PJFP.com

Pursuit of Joy, Fulfillment, and Purpose

Tag: agentic AI

  • Ken Griffin on AI, the Golden Age of Entrepreneurs, and the Taiwan Chip Risk That Would Cut US GDP 8 Percent: Inside the Citadel Founder’s Goldman Sachs Great Investors Interview

    Ken Griffin, founder and CEO of Citadel, sat down with Goldman Sachs’ Raj Mahajan at the firm’s Apex Symposium (recorded June 2, 2026) for this episode of Goldman Sachs Exchanges: Great Investors. It is their third public conversation in seven years, and Griffin is unusually candid: about the Friday he went home “shocked and depressed” over AI, the agentic system inside Citadel that compresses six weeks of PhD-level work into two hours, why a Chinese move on Taiwan would throw the US into a depression within six months, and the one question every hedge fund investor should ask their GP.

    TLDW

    Griffin names his two proudest leadership calls: dragging Citadel back to the office five days a week before it was acceptable (citing Fed research that remote work has hurt young Americans’ employment more than AI has), and Citadel’s pandemic role, from getting the FDA to approve experimental COVID drug trials in 72 hours to shaping the incentive design behind Operation Warp Speed, which he credits with saving roughly half a million American lives. On markets, he explains why the S&P sits at all-time highs despite wars in the Middle East and Europe: US energy insulation, stunning Chinese oil demand destruction, and record corporate earnings. On AI, he distinguishes hype from reality (a dinner of multinational CEOs gave him five stories of “AI transformation,” none of which were actually AI), then describes the internal breakthrough that changed his mind: an agentic system that reads, reproduces, and out-of-sample-tests academic finance papers in 2 to 3 hours instead of 6 to 8 weeks. The consequences: no layoffs at Citadel, but competitive moats across the economy are being filled in at lightning speed, setting up a golden age of entrepreneurship. He covers the compute market (all available compute is utilized all the time; market makers now spend hundreds of millions a year), China’s lead in roughly 67 of 74 critical technologies, the Taiwan scenario in which losing TSMC chips cuts US GDP 8 percent in six months, an energy doctrine built on nuclear, natural gas, and building data centers (with their own generation) in America, his stress-test approach to tail risk (definable, tolerable, still in business), and hedge fund economics: the industry’s cost of capital is roughly risk-free plus 4 percent, which is why Citadel has returned $25 to 30 billion to its LPs.

    Thoughts

    The most useful thing in this conversation is Griffin’s two-sided read on AI, because he refuses to pick a lane. The paper-replication story is the cleanest documented example yet of AI eating not just white-collar work but masters-and-PhD-level work, from the man whose firm profits from that labor. Yet in the same breath he reports zero headcount reduction, because Citadel has more problems to attack than people to attack them. Both things are true at once, and he names the synthesis honestly: the individual firm gets more productive while every firm’s moat gets shallower. Most commentary picks either the doom frame or the productivity frame. Griffin holds both, and his conclusion (a golden age of entrepreneurship, startups running on a few AI systems instead of 30 to 40 employees) is the actionable part.

    His dinner-party anecdote deserves to be a standard reference. Five global CEOs effusing about AI transformation, and every single story was actually machine learning, optimization, or plain digitization. The C-suite cannot tell AI from technology at large, which means a meaningful slice of the “AI is transforming our business” narrative priced into the S&P is really a decade-old digital revolution wearing a new label. That is not a bearish observation, since the earnings are real either way, but it matters for anyone trying to figure out which companies actually have AI leverage and which have rebranded their IT budget.

    The Taiwan section is the starkest risk framing you will hear from someone who runs both a hedge fund and one of the world’s largest market makers. An 8 percent GDP contraction in six months is not a market correction, it is Boeing halting production, new cars stopping, and consumer electronics freezing simultaneously, because TSMC chips are in every high-end product made. What makes his version distinctive is the second-order point: in a Taiwan blockade, he does not expect unified Western sanctions. Europe’s membership on “team USA” is less clear than it was two years ago, and the Middle East will play Switzerland because China buys its oil. Investors should notice that his answer to “how do you hedge this?” is not clever derivatives, it is his stress-test doctrine: know the worst case, size exposures so the loss is definable and tolerable, and stay in business to fight back.

    Finally, the small structural details are where the conversation earns its Great Investors billing. Compute has become a commodity input like jet fuel, fully utilized at all times and allocated purely by willingness to pay, which quietly favors high-margin businesses and squeezes everyone else. Alternative data made the present transparent, so the remaining edge in stock picking is multi-year vision about which companies are building transformative products. And the hedge fund test he closes with is one any allocator can use tomorrow: is your GP in the asset management business or the performance business? Citadel returning $25 to 30 billion to LPs is what the performance answer looks like in practice.

    Key Takeaways

    • Griffin’s proudest leadership call was bringing everyone back to the office five days a week, extremely early and against the culture, because humans are social creatures who learn through apprenticeship and mentorship.
    • He cites a Fed paper on reduced employment among workers under 30: remote work turns out to be a more important factor in diminished opportunities for young Americans than AI.
    • At the start of the pandemic, a hospital-system CEO called Griffin because he could not get FDA approval for drug trials on ventilated COVID patients; Citadel’s team got experimental trials approved in about 72 hours.
    • The key insight behind Operation Warp Speed, which Griffin discussed at length with Jared Kushner, was an incentives fix: the US government paid pharma to manufacture vaccines before FDA results existed, collapsing time-to-market from months to days.
    • By his math, the country spent a few billion dollars on that risk, saved a few trillion dollars of GDP, and saved roughly half a million American lives.
    • The S&P is at all-time highs despite a Middle East war, a still-raging war in Europe, and a potential skirmish over Cuba, because the US is relatively shielded from the energy shock.
    • China’s oil demand elasticity stunned even Citadel’s commodities business, one of the largest in the world; that demand destruction plus episodic oil flows out of the region has kept crude near the low $100s instead of the nearly $200 most models predicted if the straits closed.
    • Citadel has been a huge user of machine learning since TensorFlow arrived roughly a decade ago; the current wave is an acceleration of a digital revolution already underway, not a clean break.
    • At a dinner two years ago, Griffin asked global multinational leaders to share how AI was transforming their businesses: he got four or five great productivity stories and not one actually involved AI. They were machine learning, optimization, and digitization.
    • In the C-suite the nuance between AI and technology at large gets lost, but bigger budgets and CEO enthusiasm are pushing through real projects with real bottom-line impact; US corporate earnings are at all-time highs and multiples have actually come down as a result.
    • The use case that sent Griffin home shocked and depressed: a Citadel team member built an agentic AI system that reads an academic finance paper, reproduces it, verifies the published results, and tests them out of sample in 2 to 3 hours on average.
    • That same replication work previously took a legion of young masters and PhD hires roughly six to eight weeks per paper; Citadel finds a few tradeable ideas a year this way, and a few ideas can be worth a lot of money.
    • The point he stresses: this is not just a white-collar job being automated, it is a master’s or PhD-level job, and AI is now cracking problems (like the 80-year-old math problem OpenAI solved) that seemed beyond its reach two or three years ago.
    • Despite the breakthrough there has been no reduction in headcount at Citadel: the firm has more problems to attack than people, so Griffin takes every productivity gain he can get.
    • The flip side is that competitive moats across corporate America are being filled in at breathtaking speed, which Griffin expects to produce a golden age of entrepreneurial activity.
    • His example: a startup that would traditionally need 30 or 40 employees now runs with just a few AI systems, letting entrepreneurs take on incumbents in ways impossible 5, 10, or 20 years ago.
    • Some workers face genuinely hard transitions (his example is English-to-German translators), and the country needs to figure out how higher education can retrain these people quickly.
    • Stock picking remains a timeless business with a similar skill set, but the market will increasingly reward multi-year vision about which companies are creating transformative products rather than skill at calling quarterly earnings beats.
    • Alternative data (Citadel has access to the credit card spending of millions of Americans) made the here-and-now transparent a decade ago; AI plus bright people now triage the present almost instantly, so relative value accrues to those who can see years ahead.
    • At Citadel Securities, transformer models continue a decade of ML-driven improvement in pricing and risk management, and the same is true at other leading market-making firms.
    • For all intents and purposes, all available compute in the world is utilized all the time; access is decided by who will pay the most, and the per-unit price has risen beyond what anyone reasonably projected two or three years ago.
    • Large market-making firms now spend hundreds of millions of dollars a year on compute; Griffin compares compute inflation to jet fuel and egg prices, a cost that high-margin businesses can bear and low-margin businesses cannot.
    • China leads in roughly 67 or 68 of the 74 or 75 most important technologies in the world, including solar, EV batteries, and multiple quantum fields, and has pulled ahead in published academic papers.
    • The drivers are structural: 1.4 billion people, an extraordinarily strong educational culture, and far more STEM graduates, producing exactly the human talent needed to win in a high-IP world.
    • China is no longer relegated to producing low-margin products designed in America, and Griffin calls that shift a threat to the American way of life; the answer is not tariffs but educating US youth to out-compete, out-innovate, and out-problem-solve.
    • If China takes Taiwan and the US loses access to Taiwanese semiconductors, the rough estimate is US GDP falls 8 percent in six months: a great depression in the blink of an eye, unlike any before.
    • The mechanism is concrete: Boeing stops making planes within six months, most new cars stop being manufactured, consumer electronics production freezes, because TSMC chips are in every high-end product made.
    • There are no winners in a Taiwan escalation: tanking the US economy would have draconian knock-on effects for China given America’s importance as an export market.
    • In a Taiwan blockade Griffin does not expect unified global sanctions against China: where you sit determines your exposure, Europe’s place on team USA is less clear than two years ago, and the oil-exporting Middle East will play Switzerland.
    • On energy, the US must re-embrace nuclear, with small modular reactors a big part of the story: nuclear has effectively no carbon footprint and one of the lowest mortality rates of any energy source ever used (hydro has killed magnitudes more people).
    • He punctures the clean-energy veneer: solar cells are often made in western China by burning coal, with roughly a seven-year energy payback, and carbon fiber wind turbine blades last 20 years then fill landfills because they do not break down. No truly clean solution exists until fusion or broader nuclear.
    • Until then, natural gas is America’s huge asset: decades of cheap supply, and one of the few things that has actually brought down US carbon emissions.
    • Data centers are going to get built somewhere, and Griffin argues it would be inane for America to end up dependent on foreign countries for them; his fix for NIMBY politics is to require data center builders to construct corresponding power generation, tied to the grid for reliability, rather than pushing costs onto consumers.
    • His hedging doctrine for complicated risks: run stress tests, know exactly how much you lose and where in the worst case, and keep exposures sized so the loss is definable, tolerable, and leaves you still in business and able to fight back. You will never hedge every tail event.
    • Hedge fund industry economics: the long-run cost of capital is roughly the risk-free rate plus 4 percent; underperform and capital flows out, outperform and it flows in, and inflows dilute alpha because alpha capacity is finite.
    • Citadel has returned $25 to 30 billion to its limited partners to keep return on equity high: Griffin’s job is to grow annual alpha capacity, and any capital beyond what the portfolio needs goes back to LPs.
    • The alignment test for allocators: the biggest investor in Citadel’s funds is Griffin and his partners, and every LP should ask whether their GP is in the asset management business or the performance business.

    Detailed Summary

    Return to Office and the Cost of Remote Work

    Asked what he is most proud of beyond the numbers, Griffin starts with Citadel’s early, countercultural demand that everyone return to the office five days a week. He frames it as a human capital decision, not a control decision: people learn through apprenticeship, mentors are critical to development, and the underdevelopment of talent from remote work has damaged the broader economy. He points to recent Fed research on falling employment among under-30s: remote work turns out to matter more than AI in diminishing opportunities for young Americans. Citadel not only brought its team back but publicly extolled the virtues of doing so, and Griffin believes history will be on his side.

    72 Hours to FDA Approval and the Warp Speed Incentive Design

    His second point of pride is Citadel’s pandemic chapter. As the first US COVID cases appeared, a former partner running a major New York hospital system called: he could not get FDA approval for experimental drug trials on ventilated patients facing imminent death, and believed only Griffin could make it happen. Citadel’s team, with decades of government experience, got approvals moving in about 72 hours. The second act was Operation Warp Speed, whose core idea Griffin discussed at length with Jared Kushner: pay pharmaceutical companies to manufacture vaccines before FDA results, so a positive result means days to market instead of the standard sequence losing three to six months. No company would spend billions producing vaccines that might be flushed down the sewer, so the US government took the manufacturing risk on unproven efficacy. A few billion dollars spent, a few trillion in GDP saved, and roughly half a million American lives.

    All-Time Highs in a World at War

    Griffin’s market picture is unsentimental: there is a war in the Middle East, a still-raging war in Europe, potential trouble in Cuba, and the peace both men grew up with is off the table. Yet the S&P sits at record highs. His explanation: America is relatively shielded from the war-driven energy crisis. China has curtailed oil demand with an elasticity that stunned even Citadel’s commodity desk, and episodic oil and LNG flows keep leaving the region, holding crude around the low $100s when most estimates had a strait closure producing nearly $200 a barrel. Meanwhile corporate earnings are at all-time highs, enough that multiples have actually compressed over the last 12 months.

    The AI Story CEOs Tell Versus the One That Is True

    Citadel has used machine learning heavily since TensorFlow arrived a decade ago, powering everything from radiology reads to self-driving cars across the economy, so Griffin sees today’s AI wave as an acceleration of an ongoing digital revolution. His favorite corrective: at a dinner with global multinational leaders two years ago, everyone was effusive about AI transforming their businesses, so he asked them to go around the table with specifics. Four or five genuinely impressive productivity stories emerged, and not one involved AI: they were machine learning, optimization, digitization, technology at large. The C-suite blurs the distinction, but the enthusiasm has unlocked bigger technology budgets and real bottom-line projects, which is part of why earnings are at records.

    The Agentic System That Shocked Him

    Then comes the story behind the famous “shocked and depressed” Friday. Citadel employs legions of young masters and PhD graduates to replicate academic finance papers: read the hypothesis, judge the work, reproduce results, and test whether the effect persists out of sample (does buyback activity predict outperformance, for example). Each paper takes six to eight weeks, and the process surfaces a few valuable ideas a year. A colleague built an agentic AI system that does the entire pipeline (read, reproduce, verify, out-of-sample test) in two to three hours on average. Griffin’s emphasis: this is not routine white-collar work, it is master’s and PhD-level work, and paired with OpenAI solving a math problem open for 80 years, it shows AI cracking problems considered out of reach two or three years ago. Notably, Citadel cut zero headcount on the back of the breakthrough; the firm has more problems worth attacking than people to attack them, so every productivity gain gets absorbed.

    Filled-In Moats and a Golden Age of Entrepreneurs

    The macro consequence Griffin draws is double-edged. Hold two thoughts at once: AI is reaching very high-level work in the job market, with some workers (translators, for instance) facing hard transitions that demand fast retraining through higher education. And simultaneously, the competitive moats of corporate America are being filled in at breathtaking rates. That means entrepreneurs can launch businesses at speeds impossible 5, 10, or 20 years ago: he mentions a startup running on a few AI systems where 30 or 40 employees would once have been required. He expects a wave of these stories over the next couple of years as founders use the technology to take on incumbents.

    The Future of the Stock Picker

    Griffin has called stock picking a timeless business, and he still sees a similar skill set for the portfolio manager of the future, with one shift in emphasis. Predicting quarterly earnings beats has gotten far harder over a decade as alternative data (credit card panels covering millions of Americans, telegraphing Starbucks and McDonald’s revenues) made the present transparent. Now bright people plus good AI triage the here-and-now almost instantly. The scarce, rewarded skill becomes vision: identifying which companies are building genuinely transformative products years before the market fully prices it.

    Compute Is the New Jet Fuel

    At Citadel Securities, which holds double-digit market share across equities, futures, and treasuries, transformer models extend a decade of machine learning gains in pricing and risk. The compute market backdrop is what Griffin calls breathtaking: essentially all available compute on Earth is utilized all the time, so access reduces to who will pay the most. Per-unit compute prices exceed what anyone reasonably projected two or three years ago, and large market makers now spend hundreds of millions of dollars annually. He treats it as straightforward input inflation, like jet fuel or eggs: high-margin businesses can bear it, low-margin ones cannot.

    China’s Technology Lead and the Taiwan Equilibrium

    Griffin states the cold reality: China is one of the most innovative, fastest-growing economies in the world, leading in roughly 67 or 68 of the 74 or 75 most important technologies (solar, EV batteries, several quantum fields) and now ahead in published academic papers. The foundation is 1.4 billion people, a culture with an extraordinary emphasis on education, and far more STEM graduates. China is no longer relegated to manufacturing low-margin products designed in America, and Griffin calls that a threat to the American way of life. His prescription is pointed: not tariffs, but educating American youth to out-compete, out-innovate, and out-problem-solve. Taiwan is the painful pressure point with no winner. If China takes Taiwan and the US loses TSMC chips, GDP falls an estimated 8 percent in six months: Boeing stops making planes, most new car production halts, consumer electronics freeze, a great depression in the blink of an eye. China would suffer draconian knock-on effects too. As an investor he thinks about position: sanctions in a Taiwan blockade would not be unified, Europe’s place on team USA is a genuine question mark now, and the oil-exporting Middle East would play Switzerland since China is its biggest customer.

    Energy Realism: Nuclear, Gas, and American Data Centers

    On powering AI, Griffin wants America to lead again in nuclear, with small modular reactors central: no meaningful carbon footprint and one of the lowest mortality rates of any energy source ever deployed (hydro has killed magnitudes more people). He challenges the superficial cleanliness of renewables: solar cells are often made in western China with coal power, requiring about seven years of energy capture to break even against the coal burned making them, and 20-year-old carbon fiber wind turbine blades do not break down and are already filling landfills. Until fusion or expanded nuclear, America’s real asset is natural gas: decades of cheap supply that has actually driven US emissions down. His data center position is blunt: they will get built somewhere, and depending on foreign countries for them would be inane, so build them in America. His answer to NIMBY politics: require data center developers to build corresponding power generation, tied to the grid for reliability, so the cost never lands on the American consumer.

    Tail Risk, Tolerable Losses, and Hedge Fund Alignment

    On hedging complicated risks, Griffin’s method is stress testing: if this happens, how much do we lose and where, and is that loss tolerable? You can never manage a portfolio for every possible tail event, but you can keep exposures sized so the worst case is definable and tolerable, leaving you still in business and positioned to fight back. On industry returns, he pegs the hedge fund cost of capital at roughly the risk-free rate plus 4 percent as the long-run equilibrium: underperformance drains capital, outperformance attracts it, and since recent outperformance keeps pulling money in, growing assets dilute alpha. That is why Citadel has returned $25 to 30 billion to LPs: alpha capacity is finite, Griffin’s job is to grow it, and excess capital goes back to investors to keep return on equity high. The closing advice is an alignment test: Citadel’s biggest investor is Griffin and his partners, and every allocator should ask whether their GP is in the asset management business or the performance business.

    Notable Quotes

    “Turns out that remote working is a more important factor to diminished employment opportunities for young Americans than AI.”

    Ken Griffin, citing Fed research on under-30 employment

    “We spent a few billion dollars as a country. We saved a few trillion dollars in GDP. We saved roughly half a million American lives.”

    Ken Griffin, on Operation Warp Speed’s incentive design

    “I got four or five incredible stories of how companies were achieving meaningful productivity gains. Not one involved AI.”

    Ken Griffin, on his dinner with global multinational CEOs

    “My colleague built an agentic AI system that would read a paper, reproduce it, verify the results that were published in the paper, produce the results out of sample, and do all this work in about on average 2 to three hours.”

    Ken Griffin, on the breakthrough that replaced six to eight weeks of PhD-level work

    “We’re likely to see a golden age of entrepreneur activity. Like entrepreneurs will be able to launch new businesses at breathtaking speeds and will be able to take on incumbents in ways that you just couldn’t do 5, 10, 15, 20 years ago.”

    Ken Griffin, on AI filling in competitive moats

    “All the available compute today is more or less utilized all the time. So the question is who’s willing to pay the most for it?”

    Ken Griffin, on the global compute market

    “The US loses access to Taiwanese semiconductor chips, our GDP falls by 8% in 6 months. Simply put, we go into a great depression in the blink of an eye unlike any we’ve seen before.”

    Ken Griffin, on the Taiwan scenario

    “We better damn well build the data centers in America because they’re going to get built somewhere in the world.”

    Ken Griffin, on energy policy and AI infrastructure

    “Definable, tolerable, still in business, still in a position to fight back from that point.”

    Ken Griffin, summarizing his approach to hedging tail risk

    “Are they in the asset management business or are they in the performance business?”

    Ken Griffin, on the question every hedge fund investor should ask their GP

    Watch the full conversation here: Ken Griffin on Goldman Sachs Exchanges: Great Investors.

    Related Reading

  • The Next 3 Years of AI, According to Steve Jurvetson: Moore’s Law, Superintelligence Odds, Elon Musk’s Operating Principles, and Where the Legendary SpaceX and Tesla Investor Is Betting Next

    Steve Jurvetson has spent 30 years funding the future before it was a category: an early check into SpaceX when space was not a venture sector, Tesla before electric cars were taken seriously, and now a portfolio spanning fusion, analog AI chips, and epigenetic editing at his firm Future Ventures. In this fireside chat he lays out what the next three years of AI actually look like, the three principles he has learned from working alongside Elon Musk for nearly three decades, the question he uses to separate missionary founders from opportunists, and why he thinks alignment of frontier AI systems may simply not be possible.

    TLDW

    Jurvetson argues the 130-year exponential in compute per dollar (Ray Kurzweil’s abstraction of Moore’s Law from his book The Age of Spiritual Machines) will keep running for at least three more years, carried by analog and custom AI silicon, and that this compounding is what makes startups and disruption possible at all. His gut says the next big leap will be “architecturally variant”: a new generation of labs going back to DeepMind’s founding premise of reinforcement learning, continuous learning, and novelty-seeking goal functions rather than bigger LLMs. He relays Anthropic co-founder Jack Clark’s 30 percent odds of superintelligence within a year but notes the crucial missing piece is that humans still set every goal. Adoption will be wildly uneven: anything made of atoms (cars, robots) switches over glacially, while creative work and white-collar categories like call centers (roughly 1 percent of US GDP) flip almost instantly. From Musk he draws three lessons: insane focus and saying no, maniacal attention to the cycle time of learning loops (Tesla gathers more AI training data every 4 days than Waymo has in its entire history), and being a magnet for talent by selling a grander mission. He explains Future Ventures’ current bets (fusion, free diagnostics via phone, slaughter-free meat, epigenetic editing, critical minerals, analog in-memory compute), tells solo founders their 30-day plan is to find a co-founder, predicts a turbulent transition to abundance, doubts Neuralink can keep pace with AI, dismisses Penrose’s quantum consciousness argument, and frames the post-work question with Man's Search for Meaning: humans need symbolic immortality, not just employment.

    Thoughts

    The most load-bearing claim in this conversation is not about scaling laws, it is about architecture. Jurvetson is telling you where the smart contrarian money is looking: away from ever-larger language models and back toward reinforcement learning agents with continuous learning and self-generated goals, the original DeepMind thesis that got shelved when LLMs took off. His framing of the open problem is unusually precise. The recursive self-improvement loops everyone is excited about are real, but every one of them is still human-directed. The goal-setting layer, what he calls the selection pressure of the evolutionary algorithm, is the “thin veneer of activity” AI does not yet do, and it happens to be the layer where superintelligence either does or does not arrive. That is a much sharper way to track AGI progress than benchmark scores: watch who cracks autonomous goal formation, not who tops a leaderboard.

    Almost everything else Jurvetson says reduces to a single metric: the cycle time of the learning loop. It is his explanation for Musk’s edge (launch cadence, the Tesla fleet as a data-collection machine), his filter for which industries flip fast (bits iterate at machine speed, atoms are stuck with 11-to-12-year car replacement cycles and FDA timelines), and even his bear case on Neuralink, which he has invested in. Biology cannot iterate at synthetic speed, so the substrate that learns fastest wins. Once you see the pattern, it becomes a genuinely useful lens for evaluating any company, career, or technology: ask how fast the loop spins, not how impressive the current artifact is.

    The aside that deserves the most attention is his flat statement that mechanistic interpretability will not bear fruit and that control and alignment of a cutting-edge system is not possible. His reasoning is structural, not rhetorical: anything produced by an iterative algorithm run billions of times (evolution, neural network training) is inherently inscrutable, and it will always be easier to build a new intelligence than to reverse engineer one you already made. He swaps “teenager” for “AI” whenever he thinks about control, which is funny until you notice he is one of the most connected investors in the Musk orbit saying the safety agenda rests on a false premise. Sitting that next to the 30 percent superintelligence odds he cites from Jack Clark produces an uncomfortable arithmetic that nobody on stage follows to its conclusion.

    For builders, the practical gold is the 50-year question. Ask a founder what their business looks like in 50 years: the opportunist laughs at the question, the missionary is relieved someone finally asked. Paired with his other filters (if only two out of ten people think your idea is crazy it is not bold enough, and a good business is one that could not have been started three years ago), it doubles as a hiring screen and a self-diagnostic. And his 30-day plan for a solo founder is refreshingly unglamorous: do not build the MVP, do not pitch investors, go persuade one person to give up their job and join you. If you cannot recruit a co-founder, that is the market’s first answer about your idea.

    Key Takeaways

    • Jurvetson invested early in SpaceX and Tesla precisely because space and automotive were not venture categories at all; a software-centric systems engineering approach applied to a sleepy industry that has not changed in decades unlocks enormous value, and that playbook is now rippling through every industry.
    • The Kurzweil curve plots 130 years of compute per dollar across five substrates (mechanical, relay, vacuum tube, discrete transistor, integrated circuit) and shows a 10,000 billion billion X improvement; Jurvetson calls it the most important thing ever graphed.
    • Customers buy compute capacity and memory, not transistors, and both have been “on rails” for 130 years; the default prediction for the next three years is simply that the curve keeps going.
    • When an incumbent declares Moore’s Law dead, it usually signals they are losing their business to someone new, as Intel was to Nvidia 15 years ago.
    • Analog chips and customized AI silicon that do discrete matrix multiply-and-add extremely efficiently will carry the mantle of Moore’s Law over the next three years.
    • Without exponential technological change there would be no startups: if business is predictable, the big get bigger and incumbents block new entrants; disruption is almost always computationally based.
    • Over the next three years AI ripples through energy, agriculture, and construction: three enormous industries that are growing as a percentage of GDP and are the least digitized on the planet, with healthcare close behind.
    • His gut says the next driver will be architecturally variant, possibly subsuming today’s models the way mixture of experts subsumes other architectures or massively parallel diffusion models reinterpret the transformer.
    • A whole new generation of neural labs is returning to the founding premise of DeepMind: reinforcement learning with continuous learning, let loose on the internet’s data sets, hunting for the algorithm that bootstraps intelligence.
    • The open question for these systems is the goal function: what plays the role of evolutionary selection pressure? Candidates include understanding the universe (the xAI mission) or a novelty-seeking algorithm that uses new discoveries as its measure of progress.
    • Jack Clark, co-founder of Anthropic, gives roughly 30 percent odds that superintelligence arrives within a year; Jurvetson declines to put odds on it himself and admits “I do not know” is the honest answer.
    • Today’s self-improving AI loops (automated verification, hyperparameter adjustment between training runs, AI-mediated experimentation) are real but still human-directed; goal setting remains the thin veneer AI does not do, and it may be the most important layer.
    • Human intelligence was bootstrapped on top of reactive limbic systems and emotional centers with cortex layered on top; it is an open philosophical question whether AI systems need to recapitulate that functional specialization to take on purpose and meaning.
    • Anything involving atoms switches over slowly: fully autonomous vehicles are inevitable (every car, train, and airplane), but people keep cars 11 to 12 years, so the physical swap-out cycle makes the transition feel glacial.
    • Physical robotics faces the same constraint: making a billion robots takes time even with recursive manufacturing techniques.
    • The domains that flip like wildfire are the ones we held as uniquely human: creative arts, moviemaking, and imagery came first, which Jurvetson finds somewhat shocking.
    • Call centers represent roughly 1 percent of US GDP and can switch over almost entirely and almost instantly; white-collar work generally has no physical swap-out cycle to slow it down.
    • People will increasingly prefer AI to human interactions when the AI is better: studies of physician bedside manner and customer service already show AIs doing a better job with emotional connection than humans.
    • Musk principle one is an insane ability to focus: running many companies forces ruthless prioritization, and he says no to anything that is not mission-critical right now, including a Craig Venter brainstorm on terraforming Mars because “none of this stuff on Mars matters” until Starship flies.
    • Musk principle two, the most important: maniacal focus on the cycle time of innovation, the core learning loop, whether launch cadence or fleet data; Tesla cameras gather more AI training data every 4 days than Waymo has collected in its entire history, because every vehicle collects data whether or not the customer paid for full self-driving.
    • Musk principle three: being a magnet for talent, screening for mastery by drilling into engineering crises a candidate actually solved rather than leaning on credentials (which are often an albatross), and framing the company as something grander (sustainable energy, multi-planetary humanity, understanding the universe) so the best people want to join.
    • Jurvetson filters founders with one question: what does your business look like in 50 years? Opportunists chuckle at the absurdity; missionaries are relieved and finally tell you what has been driving them all along. He passes on the ones who laugh.
    • The best startups hold two things in tension simultaneously: an audacious 50-to-500-year vision and a concrete plan to iterate with real customers over the next three years, chaining backward from the future to what must be built now.
    • The perpetual surprise of great companies is expanding option value: autonomous driving was nowhere in Tesla’s founding plan, and Starlink, direct-to-cell, and orbital data centers were not on SpaceX’s dance card even five years ago. Exploring the option space beats purposeful ten-year planning.
    • Future Ventures invests in things unlike anything they have seen before yet adjacent to what they know, ideally companies that are literally one of a kind.
    • Current bets include nuclear fusion and subcritical fusion that avoids NRC regulation, because energy is the third bottleneck for AI after talent and compute.
    • Other 500-year-problem bets: free healthcare via a cell phone (all diagnostics as a free global service, probably launching outside the US to bypass FDA and insurance), slaughter-free meat via cellular agriculture and mycelium, and construction, where labor productivity has been flat for 30 years.
    • Recent investments span epigenetic editing (the software of biology rather than the firmware of the genome, applied to crops, pesticides, and human health), critical minerals from deep sea mining to copper refining, and reshoring US industrial capacity.
    • Three separate analog AI chip investments approach the same goal from different angles, including Mythic’s in-memory compute doing 8-bit multiplication in a single transistor, each chasing 100X and then another 100X reduction in power per calculation.
    • The portfolio is roughly 40 percent life sciences and 60 percent IT, deliberately hunting the weird edge cases that fall through the cracks of traditional pharma VC: organ harvesting for transplant, a male birth control pill, dramatically improved IVF.
    • Old industries with no new entrants are the best targets: the four largest tunnel boring companies competing with the Boring Company were all started in the 1800s.
    • The 30-day plan for a single person with an idea: find a co-founder. Great startups tend to have a dynamic duo at the founding (Jobs and Wozniak, Sergey Brin and Larry Page, Larry Ellison and Bob Miner), and persuading one person to quit their job for your mission is the first real test of the idea.
    • A founding pair with diverse backgrounds and mutual respect sets the culture for everyone hired afterward and creates cognitive diversity that ripples through the whole firm.
    • Calibrate boldness by the crazy ratio: if 100 percent of people say your idea is crazy, take the feedback; nine out of ten is pretty good; if only two out of ten think it is crazy, it is not bold enough. Also ask whether the business could have been started three years ago; if yes, that is a bad sign.
    • Co-founders most often meet at universities, one of the few places where people cross academic disciplines; breakthrough innovation happens at the interstices between formally discrete fields, and LLMs are exceptionally good at exactly that cross-domain translation, opening a fountainhead of idea discovery.
    • Roughly 19 percent of global employment involves driving vehicles, and that work is going away, just more slowly than people imagine.
    • Humans have a fundamental desire for symbolic immortality: contributing something that outlasts our brief time here, whether children, books, philanthropy, or companies. Accumulated cultural knowledge, not biology, is the primary vector of human evolutionary progress.
    • There is no peaceful path from full employment to no employment: passing through 30, 40, 50 percent unemployment will be turbulent, and no politicians are taking a long-term perspective on it.
    • On Neuralink (which he invested in): expanding the sensory periphery is very doable (higher data rates, restoring hearing and spinal function, seeing more wavelengths), but upgrading core intelligence requires reverse engineering an inscrutable iterated system, and biology’s FDA-and-wetware timescales cannot keep up with synthetic learning loops.
    • Any product of an iterative algorithm run billions of times (evolution, neural networks, genetic programming) is inherently inscrutable; Jurvetson doubts mechanistic interpretability will bear fruit and does not think control or alignment of a cutting-edge AI system is possible, likening it to mind-controlling a teenager.
    • On Penrose’s quantum consciousness argument: there is no clear mechanism and no evidence of quantum processes in the brain, and arguments that consciousness requires our specific substrate are uncompelling; machines may one day have consciousness, just not necessarily human consciousness, the same way computer memory is real memory without being human memory.

    Detailed Summary

    Betting on Sectors That Do Not Exist Yet

    Asked what he saw in SpaceX that other investors missed, Jurvetson flips the question: there were almost no investors even considering space, just as automotive and nuclear energy were not venture sectors. The bet was on Elon Musk, whom he has known for 29 years and backed across all his companies (“and his cousins, too”), and on a thesis that has since crystallized: a software-centric systems engineering approach applied to a sleepy industry that has not changed in decades unlocks extraordinary value. Aerospace and automotive proved it, and the same conversion of industrial low-margin businesses into information businesses is now playing out across the economy.

    The 130-Year Compute Curve and the Next 3 Years

    Jurvetson polls the room on Kurzweil’s famous graph, first published around 1999, and finds only a quarter have seen what he calls the most important thing ever graphed: five successive technology substrates delivering a 10,000 billion billion X improvement in the computation a dollar buys, sustained over 130 years. Moore’s Law is just the most recent refraction of a longer, almost cosmological trend that transcends the dramas of individual companies. His baseline prediction for the next three years is that the curve keeps going, carried by analog chips and custom AI silicon optimized for matrix math, and he notes that when a company like Intel declares the end of Moore’s Law, it usually means they are losing to someone new, as they did to Nvidia. The deeper point: exponential technological change is the precondition for startups existing at all, because predictable business favors incumbents. AI is the most intense crucible of compute-centric innovation yet, and over the next three years it flows into energy, agriculture, construction, and healthcare, the largest and least digitized sectors.

    Architecturally Variant: The Return of Reinforcement Learning

    Pressed on what technology drives the next wave (better LLMs, world models, robotics), Jurvetson shares a gut feeling he stresses he has not yet invested in: something architecturally variant that may subsume today’s models. He points to a new generation of neural labs returning to DeepMind’s founding premise, reinforcement learning, which was set aside when LLMs took off. The open design problem is the goal function: what is the multi-decade agentic drive, the selection pressure, the definition of success beyond reproductive fitness? He floats understanding the universe (the Grok and xAI framing) and novelty-seeking algorithms that treat new discoveries as progress. The question these labs chase is whether a single reinforcement learning algorithm with continuous learning, let loose on the internet’s data, could bootstrap intelligence. He adds a caution about today’s chatbots: we ascribe consciousness and meaning where there is none. “There’s no light on inside,” at least for now.

    Superintelligence Odds and the Missing Goal-Setting Layer

    On whether self-directed, goal-setting AI arrives within three years, Jurvetson cites Jack Clark of Anthropic giving 30 percent odds of superintelligence next year, which he finds fun mostly because at least someone put a stake in the ground. The recursive self-improvement debate is live, but he insists on a distinction: the huge improvements in the current self-improving loop (automated verification, hyperparameter tuning between runs, AI-mediated experimentation) are all still directed by humans. Goal setting remains human, and while that may be only a thin veneer of remaining activity, it is arguably the most important part, and nobody is sure how the transition happens. It may require recapitulating the brain’s functional specialization, the limbic-then-cortex layering that produced our bootstrapped consciousness. His honest answer: he does not know and does not even have odds, because three years out is genuinely hard to predict.

    Atoms Move Slowly, Bits Sweep Like Wildfire

    The gap between what the technology can do and how we use it is governed by physics and replacement cycles. Fully autonomous vehicles are, to him, obviously inevitable for everything that moves on Earth, yet cars stay on the road 11 to 12 years, so the switchover feels glacial; a billion robots likewise take time to manufacture. What flips fast is the world of bits, and strangely it started with what we considered most human: creative arts, movies, and images. White-collar work follows because there is no physical swap-out cycle: call centers, about 1 percent of US GDP, can convert almost overnight. And people will increasingly prefer the AI when it is better, showing more emotional understanding and better reading of the situation, something already visible in comparisons of physician bedside manner and customer service quality.

    Three Principles from Working with Elon Musk

    Jurvetson opens with humility (even Maye Musk cannot explain how Elon became Elon, and the books piling up on his bedside table may not have been written by humans), but offers three observations from close range. First, an insane ability to focus. Running multiple companies paradoxically helps: nobody questions Elon skipping a holiday party, and he says no to fascinating distractions, including Jurvetson’s attempt to connect him with Craig Venter to brainstorm terraforming Mars with gene sequencers. Musk’s answer: none of it matters until Starship flies. Second, and even more important, a maniacal focus on the cycle time of innovation: how fast the core learning loop runs, whether launch cadence or fleet learning. The Tesla data flywheel is the exemplar: every car collects training data whether or not the owner paid for FSD, so Tesla gathers more data every 4 days than Waymo has in its history. Third, a well-honed talent stack: pattern recognition that ignores credentials (often an albatross), drills candidates on the engineering crises they actually navigated to test for real mastery, and wraps the company in a mission grand enough (sustainable energy, multi-planetary life, understanding the universe) that the best people want in, which compounds because great people attract great people.

    The 50-Year Question and Expanding Option Value

    How do founders stay true to a mission when 99 percent of the world says it is too early? Jurvetson admits selection bias: for 30 years he has tried to back only people with a sincere, almost messianic mission rather than arbitrage-seeking opportunists. His filter is to ask what the business looks like in 50 years. Opportunists laugh (“I’ll be on my third startup by then”); the best founders are relieved to finally unload the dream they have been hiding because “colonizing Mars is an uninvestable proposition” as a day-one pitch. The best startups pair an audacious 50-to-500-year vision with a plausible path of customer iteration over the next three years, chaining backward from the future. What still surprises him is how the option value of frontier companies keeps expanding: autonomous driving was not in Tesla’s founding plan at all, and SpaceX kept unfolding from cheap launch to Starlink to direct-to-cell to orbital data centers, none of which was on the dance card five years ago. Exploring the light cone of possibilities beats designing a ten-year plan.

    Where Future Ventures Is Betting Now

    The firm looks for companies unlike anything it has seen before yet adjacent to familiar ground, targeting problems that will obviously be solved 500 years from now. In energy: multiple fusion investments plus subcritical fusion that sidesteps NRC regulation, because energy is the third bottleneck for AI after people and compute. In health: free diagnostic healthcare delivered by cell phone as a global free service, likely launched outside the US to bypass FDA and reimbursement. In food: slaughter-free meat via cellular agriculture and mycelium. In construction: still looking, after trying and failing a few times in an industry where labor productivity has been flat for 30 years. Recent themes include epigenetic editing (the software of biology rather than the firmware of the genome, spanning crop health, pesticides, herbicides, and human health), critical minerals and metals from deep sea mining to copper refining as part of reshoring, and three separate analog AI chip bets, including Mythic’s in-memory compute doing 8-bit multiplication in a single transistor, each chasing successive 100X reductions in power per calculation. The mix runs about 40 percent life sciences, 60 percent IT, with a taste for the weird edge: organs grown for transplant, a male birth control pill, radically improved IVF. His favorite hunting ground is old, crappy industries with no new entrants, like tunnel boring, where the Boring Company’s four largest competitors were founded in the 1800s.

    Advice for Founders: Find Your Batman and Robin

    His 30-day plan for a single person with an idea is not an MVP or a pitch deck: find a co-founder. Startups tend to be founded by dynamic duos (Jobs and Wozniak, Sergey Brin and Larry Page, Larry Ellison and the lesser-known Bob Miner), and a pair with diverse backgrounds and mutual respect creates a rapid iteration loop and sets the cultural template for every future hire. Persuading one person to quit their job for your crazy idea is the first proof the mission can recruit. On calibrating craziness: if literally everyone thinks the idea is crazy, take the feedback; nine out of ten is pretty good; only two out of ten means it is not bold enough, because obvious ideas get done by others. Ask whether the business could have been started three years ago; the right answer is no. Co-founders most often meet at universities, where students (unlike professors in their stovepipes) cross-pollinate between academic disciplines, and breakthrough innovation lives at those interstices. As an aside, he notes LLMs excel at exactly this translation between domains, opening a new fountainhead of idea discovery we are only beginning to tap.

    When Machines Do Everything: Meaning, Abundance, and Turbulence

    Asked the closing question (when machines do everything, what is the meaning of life?), Jurvetson starts with scale: roughly 19 percent of global employment is driving vehicles, and it is going away. But humans want meaningful work, driven by what he calls a fundamental desire for symbolic immortality: children, books, philanthropy, companies named after founders, all instantiations of the urge to contribute something that outlasts us. Translating the question into humanity’s mission statement, he lands where Yuri Milner and Musk do: to understand the universe and add to accumulated knowledge, because culture, not biology, is the primary vector of human evolutionary progress. If we could hyperspace-jump to Peter Diamandis-style abundance, where everything physical costs a dollar a pound and machines do all labor, we could all be philosopher kings and artists. But he refuses to end on false comfort: there is no visible peaceful path from full employment through 30, 40, 50 percent unemployment, that transition will be turbulent, and no politicians are taking a long-term view of it.

    Neuralink, Inscrutable Systems, and the Alignment Heresy

    In audience Q&A, Jurvetson confirms he invested in Neuralink (the idea traces to the neural lace of Iain M. Banks’ novel Surface Detail, which he recommends) but offers a contrarian view. Working from the periphery is very promising: restoring broken function, fixing spinal cords, expanding senses, higher-bandwidth communication. Upgrading core functionality, actually making someone smarter, is another matter. His reasoning comes from decades of watching complex systems: any artifact produced by an iterative algorithm run billions of times (evolution, neural networks, genetic programming, cellular automata) is inherently inscrutable. That is why he doubts mechanistic interpretability will bear fruit and flatly does not think control and alignment are possible for a cutting-edge AI system; he mentally swaps “teenager” for “AI” whenever the control question comes up. The same inscrutability applies to the brain: it will be easier to build a new intelligence than to reverse engineer one already made, and FDA cycles plus human biology cannot iterate at the speed of synthetic learning loops, so he lacks faith Neuralink keeps up with AI. Kurzweil’s uploading dream, he suggests, is a case of wanting something to be true within one’s lifetime.

    Penrose, Quantum Brains, and Machine Consciousness

    On Roger Penrose’s argument that consciousness depends on quantum processes and is therefore unreachable by AI, Jurvetson is respectful of the man and dismissive of the claim: there is no clear mechanism (a speculative lithium isotope coupling aside), and it amounts to wishful thinking. Generalizing, he finds all vitalist arguments that our substrate is uniquely necessary uncompelling; you could make a better case that carbon is special to life than that neurons are essential to consciousness. His favorite reframe swaps in the word memory: computers have memory that is nothing like holographic, gracefully degrading human memory, yet nobody debates whether computer memory is real. Machines may likewise develop a different kind of consciousness without human consciousness. Declaring something impossible is a much higher-order proposition than admitting ignorance, so his position is: he does not know whether the current AI path leads to consciousness, but his gut says machines will get there one day, perhaps via evolution-like reinforcement learning approaches that recapitulate what biology already proved possible.

    Notable Quotes

    “I have this gut feeling that it’ll be something architecturally variant. It might subsume the models that we know now.”

    Steve Jurvetson, on what drives the next three years of AI

    “It’s almost cosmological. Like, why has humanity’s capacity to compute compounded for 130 years?”

    Steve Jurvetson, on the Kurzweil abstraction of Moore’s Law

    “If business is predictable, if there isn’t disruptive technological change, the big get bigger.”

    Steve Jurvetson, on why exponential compute is the precondition for startups

    “The Tesla cars today in their cameras gather for their AI training set more data every 4 days than Waymo has in its entire history.”

    Steve Jurvetson, on the data flywheel behind Musk’s learning-loop obsession

    “If it’s like only two people think it’s crazy, that’s bad because it’s clearly not bold enough. If it’s an obvious idea, other people will do it.”

    Steve Jurvetson, on calibrating how crazy a startup idea should be

    “Despite attempts at mechanistic interpretability in AI, I don’t think that’s going to bear fruit.”

    Steve Jurvetson, on why iterated systems are inherently inscrutable

    “It’d be easier to build a new intelligence than it is to reverse engineer one you’ve made.”

    Steve Jurvetson, on why he doubts Neuralink can keep pace with AI

    “I think all humans have a fundamental desire for symbolic immortality, this belief that we’ve contributed something to the world that transcends our brief time on this world.”

    Steve Jurvetson, on the meaning of life when machines do everything

    “It’s much higher order proposition to say something is impossible than to say I don’t know.”

    Steve Jurvetson, on whether AI can ever be conscious

    Watch the full conversation here: The Next 3 Years of AI: Lessons from Elon Musk’s First Investor.

    Related Reading

  • Jonathan Ross on Groq’s $20 Billion NVIDIA Deal, Faster Inference, and Why Asking the Right Questions Wins the AI Age

    Jonathan Ross, the founder of Groq and the inventor of Google’s Tensor Processing Unit (TPU), sits down with David Senra (host of the Founders podcast) to walk through Groq’s roughly $20 billion partnership with NVIDIA and the decade of near-death struggle that preceded it. You can watch the full conversation here. Ross, now a senior executive at NVIDIA following the deal, is unusually candid about being one of the world’s worst leaders when he started, about coming three weeks from running out of money, and about the single contrarian bet (that faster inference would make AI both faster and smarter) that almost everyone, including his own engineers, told him was pointless.

    TLDW

    Ross explains the structure of the NVIDIA deal (a call to Jensen Huang about buying 100,000 GPUs turned, in three weeks, into NVIDIA’s largest deal by nearly 3x) and why pairing Groq’s LPU with the GPU defeats the many different bottlenecks inside an LLM the way you would use both 18-wheelers and delivery vans in a logistics network. He unpacks the AlphaGo moment that revealed faster inference makes models smarter, the shift from the information age (answering questions) to the AI age (asking the right questions), and a leadership philosophy built on autonomy, one brutally clear priority (25 million tokens per second on a challenge coin), and giving people the fewest constraints so they can surprise you. He shares hard-won lessons from Jensen and NVIDIA (the least political large org he has seen, no secret one-on-ones), his concepts of reality quotient and the dominant game, return on luck and the GitHub opportunity he let his team talk him out of, intentional leadership (“I intend to do this”), the Grok bonds that traded salary for equity and saved the company, hiring for negatives instead of positives, loss bias and manufactured discontent, and a closing case for radical optimism: code is becoming free, software creation is being democratized like literacy, and education should stop teaching kids to answer questions and start teaching them to ask.

    Thoughts

    The technical spine of this interview is a genuinely counterintuitive claim: you can make a model smarter by making it faster. Ross’s proof is the AlphaGo anecdote, where the exact same model, ported from GPUs to his TPU, saw its ELO jump by hundreds of points and beat the world champion, because more compute per unit of time let it search deeper and surface moves like the famous Move 37 that were too far down the tree to find otherwise. Once you internalize that inference speed is not a convenience but a capability multiplier, the entire Groq thesis, and the logic of the NVIDIA deal, snaps into focus. The industry spent years treating fast inference as a nice-to-have. Ross treated it as the whole game, and was nearly alone in doing so for a very long time.

    The most transferable material is the leadership arc, precisely because Ross is willing to say he was bad at it. His core insight is that there is no single correct way to lead, any more than there is one way to invest, and the founder’s first job is to know which way is true to them. Ross is a delegator who hires autonomous people and gives them a single, poetically compressed objective, then gets out of the way. The reason that matters is subtle: if you over-constrain the goal, your team can never surprise you with a better answer than the one you already had, which means they can never actually innovate. The Kelly Johnson line Senra offers (“extreme performance often comes from one brutally clear priority”) is the same idea from the Skunk Works side. A challenge coin that reads “25 million tokens per second” is not a slogan, it is a mechanism that lets every engineer connect their work to one dominant game.

    Two ideas deserve to be lifted out and used directly. The first is intentional leadership, borrowed from David Marquet’s submarine turnaround: replace “should I do this?” with “I intend to do this.” Asking for opinions invites pessimism and hands your most timid people a veto. Declaring intent still lets someone shout “the hatch is open” when it truly matters, but it stops the reflexive no. Ross traces years of stalled progress to the simple error of asking instead of declaring. The second is his inversion of hiring: hire for negatives, not positives. Growing talent means showing people the path, so you emphasize positives. Selecting talent means screening people out, so you hunt for the disqualifying negatives, because one person’s negative trait infects the whole team. Most founders, Ross included for years, are clever enough to talk themselves into any candidate. A versioned “people spec” and a deliberate loss-averse posture are the antidote.

    The Grok bonds story is the emotional center and a small masterpiece of change management. Facing a layoff list that would have killed the company (because the people slated to be cut were exactly the ones needed to make the product work at all), Ross instead asked the team to trade salary for equity, framed with World War II war-bond imagery. Eighty percent participated, half went to statutory minimum wage, and attrition actually fell. His phrase for why is “put everyone’s hands on the steering wheel.” Passengers fear a windy road, drivers feel in control. It is a reminder that morale under existential stress is often a function of agency, not comfort, and that the Phil Knight move of converting employee sacrifice into ownership is a recurring pattern in company survival stories for a reason.

    Where the conversation turns almost spiritual is manufactured discontent. Ross observes that the entrepreneurs in a room of successful people were the least happy with their wealth, and that this very dissatisfaction was the fuel that kept them building. His own current discontent is stark and worth sitting with: the world does not have enough compute, and if it takes an extra year to cure cancer or slow aging because of that shortage, he considers it his fault. Whether or not you accept the moral weight he assigns himself, the mechanism is instructive. Edwin Land wrote “300 people died today” on the whiteboard while inventing anti-glare technology. A concrete, human cost attached to delay is a far more durable motivator than a revenue target. Paired with his closing optimism about code becoming free and software creation democratizing like literacy, it makes for one of the more clear-eyed and yet hopeful founder conversations in recent memory.

    Key Takeaways

    • The NVIDIA deal began as a request to buy about 100,000 GPUs; Jensen saw what Groq had built pairing GPUs and LPUs and decided to make it available to all NVIDIA customers, closing what Ross calls the firm’s biggest deal by nearly 3x in roughly three weeks from first call to wired money.
    • GPUs and LPUs are complementary: inside an LLM’s decoder layer, the GPU is better at the compute-bound attention portion and the LPU is better at the memory-throughput-bound weights, so combining them defeats bottlenecks across the whole performance curve, like using both 18-wheelers and last-mile vans.
    • As AI increasingly talks to AI, speed dominates, because agents kick off other agents and compound; a human tolerates a one-second wait, but AI is just sitting there idle.
    • Agentic micro payments will make the number of payments skyrocket, but payments infrastructure is not yet built for AI operating inside an allocated budget.
    • Ross prototypes cutting-edge ideas as personal hobby projects first, then brings them to work; his personalized “daily brief” evolved from long text into headlines he can interrogate with follow-up questions, like the game of 20 questions.
    • The information age rewarded answering questions; the AI age rewards asking the right ones, as everyone shifts from individual contributor to leader of AI, and good leaders ask the question no one else did.
    • There is no single right way to lead, just as there are many ways to invest; the founder’s job is to know themselves and pick the leadership form that is true to them (inspiration versus fear, control versus delegation).
    • Ross was, by his own account, one of the world’s worst leaders at the start, which cost Groq three to four years; his fix was to define one goal simple enough to fit on a challenge coin: 25 million tokens per second.
    • The fewer constraints you give a person (or an AI agent), the more freedom they have to surprise you with a better solution; over-constraining the goal makes real innovation impossible.
    • Lessons from Jensen and NVIDIA: it is the least political large organization Ross has seen, Jensen never runs secret one-on-ones (tell everyone at once, copy everyone on email), and the whole strategy reduces to “what does the customer actually need?”
    • Jensen manages around 60 direct reports, each smarter than him in their own domain, which he offers as the model for orchestrating AI agents that may be smarter than you.
    • Asking a sharp question that makes an expert say “I didn’t think of that” is a universal founder skill (it appears in every Bezos book) and can be honed.
    • Confidence, not competence, was Ross’s early bottleneck: shadowing a leader of 2,000 people, he realized he would have made the same decisions, and acting with confidence made people follow his direction without changing the decisions themselves.
    • The better and more creative your people, the harder they are to manage; running 450 highly creative scientists felt more like managing 5,000.
    • Reality quotient (RQ), distinct from IQ, is the ability to recognize reality and, in its extreme form, to choose the dominant game; MySpace optimized accounts signed up while Facebook optimized monthly active users and won.
    • The first principle of change management is to make it feel like it is not a change; people who seem fine with change are usually anchored to something that did not change.
    • Return on luck (from Jim Collins): the most successful companies do not get more lucky breaks, they seize the ones they get; Ross let his team talk him out of powering GitHub’s LLMs on Groq chips, then vowed never again.
    • People adopt fast inference only when they experience it personally; an Anthropic demo three months before ChatGPT drew no reaction because the answers were not the audience’s own, and Groq later went viral off a fast-LLM video posted on X.
    • Great innovators often experience a problem before others do; the future is already here, just not evenly distributed, and Ross saw fast inference’s value first because of AlphaGo.
    • Intentional leadership (from David Marquet’s USS Santa Fe turnaround): say “I intend to do this” instead of asking for an opinion, which stops reflexive pessimism while still letting people flag a real problem.
    • Grok bonds: three weeks from running out of money, Ross swapped a layoff for a war-bond-style salary-for-equity exchange; 80% participated, about half took statutory minimum wage, and it bought roughly two months of runway.
    • “Put everyone’s hands on the steering wheel”: participation in saving the company cut attrition to under 10% during the crisis, echoing Phil Knight converting employee loans into Nike equity.
    • West Coast VCs behave like lemmings (one pass triggers all passes), while East Coast VCs run independent analysis; the herd missed what became NVIDIA’s biggest deal ever, a live example of the Keynesian beauty contest.
    • For the first time, top startups are not starved for cash, so putting in more money is no longer an advantage even though investors still behave as if it is.
    • Hiring flip: move from hiring for positives (how you grow talent) to hiring for negatives (how you select talent), because one negative trait poisons the team; write a versioned “people spec” like a product spec.
    • Loss bias (a loss feels roughly six times more painful than an equal gain) can be a hiring signal: Ross looks for people who “book the win early,” treating any missed improvement as a loss.
    • Poetic design (maximum meaning in minimal expression, “every word matters”) was a positive on the people spec; its negative is maximalist, cluttered design.
    • Michael Jordan manufactured pressure by taunting opponents so a loss would be humiliating, forcing superhuman performance (per his trainer Tim Grover), a deliberate version of throwing your keys over the fence.
    • Manufactured discontent (David Ogilvy’s “divine discontent”): the best entrepreneurs never rest on wins; the least happy people with their wealth were the ones who kept building.
    • Ross’s discontent today is the world’s lack of compute; he treats every delayed medical breakthrough as partly his responsibility, the way Edwin Land wrote a daily death count on the whiteboard while fighting headlight glare.
    • Software has run on “code rationing” because code was expensive to write, enforced by “no engineers”; as the marginal cost of code approaches zero, you just implement, experience, and re-implement.
    • AI democratizes software creation like the alphabet democratized literacy: Ross’s executive assistant now builds working apps, and individual founders with taste but no coding background will create valuable companies.
    • Education should be revamped around asking questions and solving real community problems; if a kid can look up or prompt the answer, the assignment taught nothing, but making them ask the right questions to get AI to solve a real problem does.

    Detailed Summary

    The $20 Billion NVIDIA Deal and Why LPUs and GPUs Belong Together

    The deal’s most striking feature is speed: the idea was first floated on a call roughly three weeks before the money was in the bank. Groq had been integrating GPUs and LPUs and went to Jensen Huang wanting to buy about 100,000 GPUs to deploy themselves. Jensen saw the combined system and decided it should be offered to all of NVIDIA’s customers. The technical logic is that processing an LLM token involves many matrix multiplies with different bottlenecks, some compute-constrained (better on the GPU, especially the attention portion) and some memory-throughput-constrained (better on the LPU, applying the trained weights). There is no single perfect architecture, so putting the two together defeats bottlenecks across the whole curve. Ross adds that as AI talks to AI, speed becomes everything, because agents spawn agents and compound exponentially.

    Asking Questions, Daily Briefs, and the Shift to Leading AI

    Ross builds cutting-edge tools as personal hobby projects before bringing them to work, including a personalized “daily brief” that functions like a presidential daily brief. He redesigned it from long text into headlines he can interrogate, because interactivity, like 20 questions, distills straight to what you actually care about. This grounds one of his signature ideas: success in the information age meant answering questions, but success in the AI age means asking the right questions. As people move from individual contributors to leaders of AI, the skill that matters is the leader’s skill of asking the question everyone else missed or was afraid to raise, since the question you ask determines the output you get.

    Knowing Your Leadership Style and the Challenge Coin

    Ross frames leadership like investing: the first principle is simply having followers, but there are infinite valid styles. New founders fail by copying advice that is not true to them. Ross is a natural delegator (he has not held a driver’s license since his teens because he would rather think than control the car) who hires unusually autonomous people. Early on this backfired badly, because he entrusted people who needed direction, and he calls himself one of the world’s worst early leaders, a gap that cost Groq years. His breakthrough was distilling the mission onto a challenge coin reading “25 million tokens per second,” which let everyone connect their work to one dominant game. He references David Marquet’s Turn the Ship Around later, but the coin embodies Kelly Johnson’s Skunk Works principle that extreme performance comes from one brutally clear priority, plus the rule that fewer constraints give people more room to surprise you, turning a team from Superman into the Avengers.

    Lessons from Jensen: Killing Politics and Serving the Customer

    Working at NVIDIA taught Ross how much further he could have pushed lessons he half-learned at Groq. NVIDIA is, in his experience, the least political large organization anywhere, and a big reason is that Jensen never tells different people different things in private one-on-ones. When you address a room, everyone hears the same message; separate conversations breed side cliques. Ross’s practical rules: hold big meetings for anything you want a group to know, and copy everyone on email so no one can route politics through you. The other Jensen lesson is to stop playing 3D chess and just ask what the customer needs, tell them only what you believe and can support, and refuse to sell them something they do not need. Senra notes he has covered roughly 19 ideas from The Nvidia Way on his Founders podcast, and Jensen’s line that he already manages 60 reports smarter than him is the template for managing AI agents.

    Reality Quotient, the Dominant Game, and Change Management

    Groq hired for reality quotient, not just IQ, because plenty of very smart people construct elaborate stories disconnected from reality. In its extreme form, RQ is the ability to choose the dominant game, the way Facebook’s focus on monthly active users beat MySpace’s focus on accounts signed up. The founder’s job is to help everyone connect their activity to that dominant game (for Groq, tokens per second), then manage the change. Ross’s first principle of change management is to make it feel like it is not a change: nobody likes change, and people who tolerate it well are usually focused on something that stayed constant. If your team is anchored to the dominant goal, a new tactic does not feel like change; if they are anchored to a narrow task, it does.

    Return on Luck, the AlphaGo Insight, and the GitHub Miss

    From Jim Collins’s Great by Choice, Ross took the idea that winners seize luck better, not that they get more of it. He experienced it first-hand with AlphaGo: after a DeepMind team asked whether his TPU was as fast as rumored (he said yes, Ghostbusters-style), porting the identical model from GPUs to TPUs pushed its ELO from around 3,200 to roughly 3,900 and it crushed the world champion. As Thinking Fast and Slow by Daniel Kahneman frames it, more compute lets the model virtually play out more moves and occasionally find a better second-best line, which is how the famous Move 37 surfaced. Faster thinking is smarter thinking. Yet Ross also let his own engineers talk him out of powering GitHub’s LLMs on Groq chips, twice, because they focused on why it could not be done rather than why it could. He eventually did the math himself, hit the numbers, and learned to stop inviting that pessimism.

    Selling Speed and Intentional Leadership

    Customers could not grasp fast inference until they felt it. Ross recalls an Anthropic demo three months before ChatGPT that drew no reaction, because seeing someone else’s answer appear is not magical, but getting your own question answered instantly is. So Groq simply put fast inference online, and it went viral after someone posted a video of a blazing-fast LLM on X (Ross noticed his own demo slowing in Norway because usage had skyrocketed). The deeper fix for internal resistance came from Turn the Ship Around, David Marquet’s account of turning the USS Santa Fe from worst to best in nuclear readiness by replacing command-and-control with intentional leadership. Saying “I intend to do this” rather than “should I?” stops people from reflexively supplying negative opinions, while still letting someone shout “the hatch is open” when there is a genuine problem.

    Grok Bonds: Three Weeks From Zero

    With three weeks of cash left and a layoff list on the table, Ross realized the cuts targeted exactly the people needed to finish an unprecedented compiler and reach the critical mass where the product would even work. Layoffs would not save the company; only reducing burn without losing people could. So Groq held an all-hands, put up World War II war-bond imagery, and launched “Grok bonds,” an exchange of salary for equity. Ross expected heavy attrition; instead 80% participated and about half dropped to statutory minimum wage, real pain for engineers used to six-figure salaries. It bought closer to two months of runway. His framing, “put everyone’s hands on the steering wheel,” explains why attrition actually fell below 10%: drivers feel more in control than passengers, and it echoes Phil Knight in Shoe Dog converting employee loans into Nike equity on the edge of collapse.

    Hiring for Negatives, Loss Bias, and Manufactured Discontent

    Ross was good at spotting smart, talented people but kept hiring ones who caused organizational problems, because he could always talk himself into a candidate. Watching a sharp head of HR screen people out, he realized he had been hiring wrong: growing talent means showing positives, but selecting talent means hunting for disqualifying negatives, since one bad trait spreads to the whole team. He formalized a versioned “people spec” with positives like return on luck and poetic design, each paired with a negative. He also hired for loss bias, the fact that a loss feels roughly six times more painful than an equal gain, seeking people who “book the win early.” That competitive, pressure-seeking wiring links to Michael Jordan manufacturing humiliation stakes (per Tim Grover in Relentless) and to David Ogilvy’s divine discontent. Ross’s own manufactured discontent today is the world’s shortage of compute, which he frames in life-and-death terms.

    The Optimistic Close: Free Code and Universal Software Literacy

    Ross ends on aggressive optimism. Software has long run on “code rationing” because code was expensive to write, policed by “no engineers” whose job is to say no. As the marginal cost of code approaches zero, the workflow flips to implement, experience, then re-implement. More important is accessibility: just as alphabets and universal education turned reading and writing from a scribe’s monopoly into a question of quality, AI is making software creation universal. His executive assistant now builds working apps, and a wave of individual founders with taste but no coding background will create valuable companies. The corollary for education is to stop teaching kids to answer questions and start teaching them to ask, revamping curricula around real community problems where the point is asking the right questions to get AI to solve something that matters.

    Notable Quotes

    “Success in the information age was about being able to answer questions. Success in the AI age will be about being able to ask the right questions.”

    Jonathan Ross, on the fundamental shift AI creates

    “The fewer constraints that you give someone, the more freedom they have to solve the problem, and the more freedom they have to surprise you with the solution.”

    Jonathan Ross, on leading creative teams

    “Being able to think faster makes you think smarter.”

    Jonathan Ross, on why faster inference produces more capable models

    “There are plenty of really smart people who wouldn’t recognize reality if it tapped them on the shoulder.”

    Jonathan Ross, defining reality quotient versus IQ

    “If you express intentional leadership, you say, ‘I intend to do this.’ People don’t tend to offer their opinion, but if it’s very wrong and there’s a reason, they will push back.”

    Jonathan Ross, on the lesson from Turn the Ship Around

    “When people are passengers in a car, they’re more nervous about a windy road or a scary road. But when they’re the driver, they feel more in control.”

    Jonathan Ross, on why Grok bonds kept the team together

    “The biggest flip in my hiring was when I went from looking for positives, which is what you do when you’re trying to grow talent, to looking for negatives, which is what you do when you’re trying to select talent.”

    Jonathan Ross, on inverting his approach to hiring

    “If it takes us an extra year to cure cancer because we don’t have enough compute, that’s my fault.”

    Jonathan Ross, on the discontent that drives him today

    Watch the full conversation between Jonathan Ross and David Senra here on YouTube.

    Related Reading

    • Groq the company Ross founded and the LPU behind the fast-inference story and the NVIDIA partnership.
    • AlphaGo versus Lee Sedol (Wikipedia) the match, including Move 37, that showed Ross how much faster hardware raises a model’s capability.
    • The Keynesian Beauty Contest (Wikipedia) the dynamic Ross uses to explain why West Coast VCs herded past what became NVIDIA’s biggest deal.
    • Zero to One by Peter Thiel, the source of the first-principles thinking Ross applied to the contrarian bet on fast inference.
    • Founders podcast by David Senra the host’s biography-driven show, source of the Jensen, Michael Jordan, and Edwin Land ideas referenced throughout.

  • SubQ 1.1 Small Explained: How Subquadratic Sparse Attention Hits 98% Retrieval at 12 Million Tokens With 64.5x Less Compute Than Dense Attention

    Subquadratic, a frontier AI research and infrastructure company, has released the model card and technical report for SubQ 1.1 Small, a long-context language model built on a new attention mechanism the company calls Subquadratic Sparse Attention (SSA). The headline claim is unusual in two directions at once: the model retains 98% single-fact retrieval accuracy at 12 million tokens, roughly twelve times the length it was primarily trained on, while cutting attention compute by 64.5x against dense attention at a 1 million token context. The deeper argument in the report is not really about a single model at all. It is about what happens to the entire retrieval-and-orchestration stack once reasoning over a complete artifact stops being prohibitively expensive.

    TLDR

    SubQ 1.1 Small is a small long-context model that replaces the dense attention of an existing open-weight frontier model with Subquadratic Sparse Attention, a learned, content-dependent sparse attention mechanism that scales linearly in compute and memory rather than quadratically. On retrieval it posts 99.12% on NVIDIA’s 13-task RULER suite at 128K tokens and 100% needle-in-a-haystack accuracy at 1M and 2M tokens, holding at 98% out to 6M and 12M tokens while attending to only 0.13% of token pairs. It keeps competitive general ability, scoring 85.4% on GPQA Diamond and 89.7% pass@4 on LiveCodeBench v6, and reaches 13% on the long-horizon AutomationBench Finance agentic benchmark, close to Opus 4.8 and GPT-5.5 and well ahead of mid and small tiers. The efficiency story is a scaling win rather than a constant-factor one: 64.5x fewer attention FLOPs than dense attention at 1M tokens and 56x faster than FlashAttention-2 on a single attention layer. The report frames cheap long-context compute as a research accelerator that let the team run more than one hundred million-token experiments and find a training recipe (long-context continued pretraining is the strongest lever) rather than guess at one, positions SSA against FlashAttention, DeepSeek’s Lightning Indexer line, state space models like Mamba, and hybrids, invokes Sutton’s Bitter Lesson to argue that RAG, chunking, and agentic scaffolding are partly workarounds for context scarcity, and was independently verified by Appen. Deployment is starting with design partners now, with a 2M to 12M token lineup planned by year end.

    Thoughts

    The most interesting move in this report is the framing, not the benchmark. Subquadratic plants its flag on Richard Sutton’s Bitter Lesson and argues that much of the modern AI stack, the retrieval pipelines, the chunkers, the re-rankers, the agentic orchestration, is scaffolding built around a single computational constraint: dense attention costs grow with the square of context length. If that constraint relaxes, a lot of hand-engineered machinery that exists to feed a model the right fragments at the right moment starts to look like the task-specific pipelines that learned representations eventually displaced. That is a genuinely provocative thesis, and it is the right lens for reading the rest of the document. The company is not selling a longer context window as a feature. It is betting that whole-artifact reasoning is a different shape of capability than retrieval over fragments, and that fragmentation destroys the cross-references a contract or a codebase actually depends on before the model ever sees them.

    The part of the paper most teams will undervalue is the claim that the real payoff of efficient attention is not cheaper inference but cheaper experimentation. A dense long-context training campaign is expensive enough that most groups get a handful of attempts and are forced to guess at the recipe. Subquadratic says SSA let them run more than a hundred experiments across six model generations with per-step iteration under a minute at million-token context, which is how they discovered that long-context continued pretraining, not clever post-training, was the dominant lever. If that holds, algorithmic efficiency becomes a first-class scaling variable alongside parameters and data, because capability becomes responsive to iteration velocity rather than raw compute alone. It reframes efficiency from a deployment line item into a research multiplier, and that is a more durable advantage than any single benchmark number.

    The generalization result deserves scrutiny precisely because it is so clean. A model trained overwhelmingly at 1M tokens, with a sliver at 2M and nothing beyond, holds 98% retrieval at 12M. The proposed explanation is that SSA routes attention by content relevance rather than fixed positional pattern, so there may simply be no obvious length boundary once the routing behavior is learned. That is plausible and the report is careful to say the 12M result emerged rather than being designed for. But single-needle NIAH is a deliberately clean probe with one target and a binary answer. The far harder RULER suite is only reported at 128K, the longest standardized length in the original benchmark, so the multi-hop, aggregation, and distractor-heavy capability that whole-artifact reasoning actually requires has public numbers at 128K, not at 12M. The honest read is that precise retrieval generalizes spectacularly and composite reasoning at extreme length is still an open question the report does not over-claim on.

    What lends the report credibility is how much counter-evidence it volunteers. It walks through MiniMax abandoning its hybrid M1 architecture and returning to full attention for M2 after efficient variants showed multi-hop reasoning deficits at scale. It admits that earlier SubQ checkpoints improved retrieval while regressing on knowledge benchmarks, forcing dedicated capability-balancing work. It describes catching a case where the MRCR benchmark moved up while the model felt worse in real workflow spot-checks, and switching its development signal to RULER as a result. That last point is a quietly important methodological argument: benchmark score and deployment behavior diverged enough to change checkpoint selection, which is a warning every team shipping long-context models should internalize. A vendor confident enough to show where its own metrics misled it is more trustworthy than one that only shows the wins.

    A few caveats keep the enthusiasm grounded. AutomationBench Finance at 13% is genuinely strong relative to peers, but it is a low absolute score across the board, including for GPT-5.5 at 18% and Opus 4.8 at 16%, so this is early evidence of agentic transfer rather than proof of a finished agent. The efficiency comparisons isolate a single attention layer rather than full end-to-end model throughput, which is the right way to expose the scaling shape but not the same as a wall-clock serving benchmark. The model is built from an unnamed donor open-weight frontier model, so some of its general-knowledge and coding strength is inherited rather than created here. And the most aggressive claims about the future, a 2M to 12M lineup and much higher sparsity, are roadmap, not released artifacts. None of that undercuts the core result. It just means the right posture is to treat SubQ 1.1 Small as a strong proof of concept for an architecture that, if it scales as advertised, could quietly remove a layer of the AI stack that everyone currently takes for granted.

    Key Takeaways

    • SubQ 1.1 Small is a long-context language model from Subquadratic AI, built on a new attention mechanism called Subquadratic Sparse Attention (SSA), released June 16, 2026 alongside a model card and technical report.
    • SSA is a learned, content-dependent sparse attention mechanism that scales linearly in both compute and memory with sequence length, rather than quadratically like dense attention.
    • The central result is context-length generalization: the model was trained primarily at 1M tokens, with some training at 2M and none beyond, yet retrieval held far past the training window.
    • Needle-in-a-haystack accuracy is 100% at 1M and 2M tokens and 98% at both 6M and 12M tokens, roughly twelve times the primary training length.
    • At 12M tokens the model attends to only 0.13% of token pairs, close to a 1,000x reduction in attention relationships, while still retrieving accurately.
    • On NVIDIA’s 13-task RULER benchmark at 128K tokens, SubQ 1.1 Small scores 99.12%, with the remaining errors concentrated in aggregation-style tasks rather than retrieval.
    • RULER tests beyond single-fact lookup: single-key and multi-key retrieval, common-word and frequent-word extraction, and multi-hop variable tracing across positions.
    • At 1M tokens, SSA requires 64.5x fewer attention FLOPs than dense attention (3.9 PFLOP versus 252 PFLOP per attention layer).
    • On a single attention layer, SSA runs 56x faster than FlashAttention-2 at 1M tokens (966 ms versus 54,164 ms on an H100), reaching parity near 16K tokens and pulling away as context grows.
    • The efficiency gain is a scaling-law win, not a constant-factor speedup: the advantage over dense attention grows as context length increases.
    • On general knowledge, SubQ 1.1 Small scores 85.4% on GPQA Diamond (pass@1), below GPT-5.5 (93.2) and Opus 4.8 (92), near Sonnet 4.6 and GPT-5.4-mini (87.5), and above GPT-5.4-nano (81.7) and Haiku 4.5 (67.2).
    • On coding, it reaches 89.7% pass@4 on LiveCodeBench v6, close to the absolute frontier (GPT-5.5 92, Opus 4.8 92.2) and ahead of the smaller tiers.
    • On AutomationBench Finance, a long-horizon agentic benchmark, it scores 13%, close to Opus 4.8 (16%) and GPT-5.5 (18%) and ahead of Sonnet 4.6 (8%), Haiku 4.5 (3%), and GPT-5.4-mini (0%). Absolute scores are low across all models.
    • The model was not trained from scratch. The team converted an existing open-weight frontier model by replacing dense attention with SSA, then built long-context ability through staged context extension and continued pretraining.
    • Context was extended in stages (262K, 512K, 1M, 2M) using YaRN positional scaling, with long-context continued pretraining performed between extension stages on naturally long data: books, long documents, and repository-scale code.
    • Roughly one trillion tokens of continued pretraining were performed, most of it at the 1M-token stage.
    • Long-context continued pretraining was the most consistent predictor of long-context retrieval gains across the experiments, more so than post-training tweaks.
    • The team ran more than one hundred long-context experiments across six major model generations, which the report argues is only possible because SSA made million-token iteration cheap (under a minute per step).
    • Capability balance was a recurring challenge: gains in long-context retrieval often regressed short-context knowledge and reasoning unless training was explicitly managed for both.
    • Benchmark scores and real deployment behavior diverged. The MRCR benchmark moved up while qualitative workflow spot-checks got worse, so the team switched its primary development signal to RULER.
    • The report frames RAG, chunking, summarization, and agentic orchestration as scaffolding built around context scarcity, drawing an analogy to Sutton’s Bitter Lesson, where hand-engineered mechanisms get displaced by larger-scale learning.
    • SSA is positioned against FlashAttention (a memory optimization that does not change quadratic compute), fixed-pattern sparse attention, DeepSeek’s learned sparse line, state space models, and hybrid architectures.
    • DeepSeek’s Lightning Indexer (used in DSA and CSA) is the closest published comparison. Its quadratic scoring overtakes the sparse attention it feeds around 52,000 tokens, reaching roughly 16x the attention cost at 1M and 190x at 12M.
    • State space models like Mamba achieve linear cost through a compressed fixed-size state, but that compression is lossy and weakens exact retrieval, which is why production efficient models are usually hybrids with some dense attention layers retained.
    • MiniMax is cited as a cautionary case: it moved from a hybrid M1 to a full-attention M2 after hybrids showed multi-hop reasoning deficits at scale and less mature supporting infrastructure.
    • The benchmark results were independently verified by Appen, a third-party evaluation firm.
    • The named use cases are financial analysis and due diligence, legal and contract work, and software engineering (architecture-level reasoning, cross-file refactoring, dependency tracing, planning, review, and long-horizon memory).
    • Sparsity settings were deliberately conservative, tuned for maximum context length rather than maximum sparsity. Limited experiments at 4x the sparsity reported positive early results.
    • The training infrastructure used a memory-scaling ladder: single node, intra-node sequence parallelism, CPU offload, multi-node sequence parallelism, nested offloading, and Ring Attention for the longest contexts.
    • Beyond about 8M tokens, BF16 numerical underflow and stability became practical constraints on evaluation.
    • The technical report is authored by Saul Ramirez, Alex Whedon, Ashmal Vayani, and Phong Vo of Subquadratic AI.
    • Deployment is starting with a first cohort of design partners, with broader rollout through the quarter and a general model lineup ranging from 2M to 12M tokens by the end of the year.
    • The company’s framing line is “Efficiency is intelligence,” and its broader thesis is that the point is not bigger context windows for their own sake but reasoning directly over complete artifacts with less surrounding scaffolding.

    Detailed Summary

    The problem: whole-artifact reasoning and context scarcity

    The report opens by naming a class of tasks it calls whole-artifact reasoning: problems whose structure requires reasoning across a complete artifact rather than over isolated fragments. A legal agreement may define a term on page 2, qualify it on page 12, carve out an exception on page 46, and amend it in a schedule. A function may be defined in one file, called from forty others, and constrained by invariants encoded in the architecture rather than in comments. A financial review may require connecting filings, earnings reports, contracts, and internal records. In each case the difficulty is not locating a passage, it is reasoning over relationships distributed throughout a large artifact. Most production systems do not do this directly. They rely on retrieval pipelines, chunking, summaries, and agentic workflows that partition information and reconstruct fragments at inference time, because dense attention scales quadratically with context length and makes direct reasoning over large artifacts expensive. Subquadratic argues that much of the modern AI stack is therefore designed to manage context scarcity rather than reason over complete artifacts, and it connects this to Sutton’s Bitter Lesson: sophisticated hand-engineered mechanisms historically get displaced once larger-scale learning becomes practical.

    What SSA is and the three requirements it targets

    Subquadratic Sparse Attention is a content-dependent sparse attention mechanism designed to satisfy three requirements at once, a combination the report argues prior approaches never achieved in a practical long-context system. First, dense-attention-level retrieval and reasoning quality, which requires routing that is content-dependent (determined by the tokens themselves) rather than driven by a fixed positional pattern. Second, subquadratic scaling, where selection, retrieval, and attention are each linear in sequence length so the mechanism is linear end to end, not only within the attention read. Third, full-context training with standard autoregressive generation, so the model can optimize over the entire context during training while keeping efficient token-by-token decoding at inference. The internal mechanism by which SSA achieves this is held back as outside the scope of the report, which focuses instead on the requirements and the experimental program that followed.

    Where SSA sits among prior approaches

    The background section is effectively a taxonomy of long-context modeling. FlashAttention is treated not as a competitor but as the standard dense-attention baseline: it solved the memory problem by never materializing the full attention matrix, but it left the quadratic compute cost untouched, so doubling context still quadruples attention computation. Fixed-pattern sparse attention (sliding-window, strided, as in Longformer, BigBird, and the sliding window in Gemma) scales well but sacrifices content-dependent routing and tends to fail on retrieval benchmarks like RULER. Compression methods like Multi-head Latent Attention reduce KV-cache memory at inference but do not change the quadratic prefill cost. Learned sparse attention, exemplified by DeepSeek’s Native Sparse Attention and its Lightning Indexer, learns where to route but pays a quadratic cost in the indexer itself. State space models and linear attention (Mamba, Mamba-2 and Mamba-3, RetNet, RWKV, gated delta networks) achieve linear cost through a compressed fixed-size state, but that compression is lossy and weak on exact retrieval. Hybrids (Jamba, Kimi Linear, Qwen3 Next, Nemotron) keep a few dense layers to preserve retrieval, which means the quadratic component still dominates at long context. System-level workarounds (RAG, agentic frameworks, recursive language models) move retrieval outside the model entirely. The report’s stated open problem is to combine subquadratic scaling end to end with content-dependent retrieval, arbitrary-position access, and practical ultra-long-context training in one system, which it claims no widely deployed architecture provides and which SSA targets.

    Training: conversion, staged context extension, and continued pretraining

    Rather than training from scratch, the team converted an existing open-weight frontier model that supported a 262K-token context by replacing its dense attention with SSA. They then extended the context window in stages (262K to 512K to 1M to 2M) using YaRN to rescale positional representations, performing long-context continued pretraining between extension stages rather than jumping straight to the final length. The training mixture emphasized naturally long data such as books, long documents, and repository-scale code, packed to the target length with document separators and without masking cross-document attention boundaries. Most continued-pretraining tokens were trained at the 1M-token stage, with roughly one trillion tokens total. Post-training played a separate role: shaping how the long-context capability was expressed while preserving reasoning, coding, and instruction following. The team explored sample-level loss aggregation to keep a few extremely long examples from dominating gradient updates, and staged the post-training corpus across synthetic retrieval tasks, long-context reasoning, coding, educational material, and general instruction following, alternating capability-building phases with recovery phases.

    Results: retrieval, knowledge, coding, and agentic tasks

    On retrieval, SubQ 1.1 Small scores 99.12% on the 13-task RULER average at 128K, with errors concentrated in aggregation-style tasks like common-word and frequent-word extraction. On needle-in-a-haystack, evaluated on 50 held-out UUID samples per length, it scores 100% at 1M and 2M (within the training window) and 98% at 6M and 12M (held out), attending to only 0.13% of token pairs at 12M. On knowledge, GPQA Diamond pass@1 is 85.4%, landing between the small and mid frontier tiers and confirming that long-context optimization need not sacrifice reasoning, a result the report credits to its capability-balancing stages after earlier checkpoints showed retrieval gains coming at the cost of knowledge. On coding, LiveCodeBench v6 pass@4 is 89.7%, and the report notes coding data played a dual role, also improving non-code long-context retrieval because code is dense with the cross-position dependencies that train general routing. On long-horizon agentic work, AutomationBench Finance is 13%, where agents must discover the right endpoints among roughly 500 across 47 applications, make interdependent API calls, follow layered business rules, and ignore seeded distractors, graded on binary end-state correctness with no partial credit.

    Efficiency and the DeepSeek comparison

    Efficiency is measured on one attention layer against a dense baseline on the same backbone. Per-forward-pass attention FLOPs scale from a 2.1x reduction at 32K to 8x at 128K, 31.5x at 512K, and 64.5x at 1M tokens (3.9 PFLOP for SSA versus 252 PFLOP for dense). Measured against FlashAttention-2 in isolation, SSA reaches parity near 16K tokens and pulls away to 56x at 1M, where it runs in 966 ms versus 54,164 ms on an H100. The report devotes a discussion section to DeepSeek’s sparse attention line as the closest published comparison. DeepSeek’s Lightning Indexer is a learned selector, but it is a full-attention distilled transformer, so it scales quadratically: in a V3.2-style configuration the indexer is cheaper than the sparse attention it feeds only below about 52,000 tokens, then overtakes it, reaching roughly 16x the attention cost at 1M tokens and 190x at 12M. SSA targets that same selection role with a selector the report says is dramatically cheaper and linear throughout, and notes SSA could conceptually replace the selector over either uncompressed or compressed representations.

    Efficiency as a research accelerator and the evaluation lessons

    A recurring theme is that the most valuable effect of cheap long-context compute was on the research loop, not just inference. Where a dense campaign would allow a handful of attempts, SSA enabled more than a hundred experiments across six model generations with per-step iteration under a minute at million-token context. That throughput is what surfaced the finding that long-context continued pretraining is the strongest lever, and it leads the authors to argue that algorithmic efficiency should be treated as a first-class scaling variable alongside model and dataset size. The report is unusually candid about evaluation pitfalls. It describes how the MRCR benchmark diverged from deployment behavior, with MRCR-optimized checkpoints often feeling worse on repository-scale code reasoning, multi-document synthesis, and contract analysis, which pushed the team to rely on RULER and a fixed set of qualitative workflow spot-checks as development signals. It also cites MiniMax returning from a hybrid M1 to a full-attention M2 as evidence that reducing asymptotic cost is not sufficient on its own if retrieval quality, reasoning at scale, and system maturity are not preserved at the same time.

    Implications, availability, and what comes next

    The report’s deployment argument is that the most important enterprise implication of long-context models is not larger windows but the ability to reason directly over complete or more-complete artifacts, moving retrieval, re-ranking, and orchestration logic into the model where the task is naturally whole-artifact rather than naturally decomposable. It is careful not to declare retrieval obsolete: for corpora larger than any plausible context window, fast-changing knowledge, and genuinely multi-stage workflows, RAG and orchestration remain the right tools. The narrower claim is that the class of scaffolding that exists only to compensate for context limits gets smaller as efficient long-context models extend the reachable window. The benchmark results were independently verified by Appen. Subquadratic is deploying SubQ 1.1 Small with a first cohort of design partners now, with broader rollout through the quarter and a general lineup spanning 2M to 12M tokens planned by the end of the year, and it flags much higher sparsity as future work.

    Notable Quotes

    “Much of the modern AI stack is therefore designed to manage context scarcity rather than reason over complete artifacts directly.”

    SubQ-1.1-Small Technical Report, framing retrieval and orchestration as workarounds for an architectural limit

    “The hybrid has moved the line, but not changed its shape.”

    SubQ-1.1-Small Technical Report, on why hybrid models keep their quadratic component at long context

    “A routing mechanism intended to make long context affordable becomes the dominant long-context cost, reintroducing quadratic scaling after providing scalar compute savings.”

    SubQ-1.1-Small Technical Report, on DeepSeek’s Lightning Indexer overtaking the attention it feeds

    “If the cost of long-context experiments is too high, teams are forced to guess at the recipe. If the cost falls far enough, they can search for it.”

    SubQ-1.1-Small Technical Report, on efficient attention as a research accelerator

    “Fragmentation systematically destroys those relationships before the model ever sees them.”

    SubQ-1.1-Small Technical Report, on why chunking hurts whole-artifact reasoning

    “Holding the whole artifact in context changes the shape of the task rather than only the speed of it.”

    SubQ-1.1-Small Technical Report, on the difference between bigger windows and direct reasoning

    “The value of SSA is therefore not only that it makes long-context inference cheaper. It makes long-context experimentation cheaper.”

    SubQ-1.1-Small Technical Report, conclusion

    Read the full SubQ 1.1 Small technical report and model card here.

    Related Reading

    • Subquadratic (subq.ai) the company behind SubQ 1.1 Small and the Subquadratic Sparse Attention architecture, where you can join the waitlist.
    • The Bitter Lesson by Richard Sutton the short essay whose argument the report leans on, that hand-engineered mechanisms lose to general methods that scale with computation.
    • Attention Is All You Need the original Transformer paper that introduced the dense attention whose quadratic cost SSA is built to remove.
    • RULER (arXiv) NVIDIA’s long-context benchmark that the report uses as its primary retrieval signal, and that fixed-pattern sparse methods historically struggle with.
    • Retrieval-augmented generation (Wikipedia) background on the RAG approach that the report frames as scaffolding around context scarcity rather than a permanent fixture.

  • Ray Kurzweil Predicts AI Will Change Humanity Completely by 2030: AGI by 2029, Longevity Escape Velocity by 2032, Nanobots in the Brain, and Why Quantum Computing Won’t Matter

    Ray Kurzweil has spent more than 60 years studying artificial intelligence and made 147 documented technology predictions since 1990 with a reported 86 percent accuracy rate. In this conversation with Tony Robbins, the 78-year-old futurist revisits his most famous forecasts and sharpens them: AGI by 2029 now looks conservative, longevity escape velocity arrives around 2032, nanotechnology connects our brains to the cloud by the mid 2030s, and quantum computing, in his view, never matters at all.

    TLDW

    Kurzweil explains the exponential thinking that powered his prediction record, from a paper he wrote at 16 to a computing-price-performance chart that runs in a straight line from 1939 relays to today’s Nvidia chips, now compounding roughly tenfold per year when hardware and software gains multiply together. He defends his 1999 prediction of AGI by 2029 (defined as AI doing the best work in every field) and says it is now the conservative end of expert opinion. He walks through AI-driven medicine: the COVID vaccine designed in two days, simulated human trials replacing 10-month clinical trials within about five years, and longevity escape velocity around 2032, after which the diligent stop losing ground to aging. He predicts AI will move inside us via nanotechnology by the mid-to-late 2030s, erasing the line between biological and computational thinking. He dismisses quantum computing as error-ridden and unnecessary for AGI. On jobs, he expects real disruption cushioned by exploding wealth and an eventual universal basic income, and advises young people to self-educate and get creative with AI tools their schools still treat as the enemy. The conversation closes with his AI twin project, the dadbot built from his father’s archives, consciousness and the soul, computronium, and why humanity must eventually expand intelligence beyond Earth.

    Thoughts

    The most interesting thing in this interview is not any single date, it is watching Kurzweil’s dates get lapped by reality. In 1999 a Stanford conference of several hundred AI experts agreed AGI would happen but pegged it at 100 years out; Kurzweil said 30 and got laughed at. Now he is the cautious one in the room, noting that “some people say it’s going to happen this year.” When the most aggressive forecaster of his generation becomes the conservative baseline, that says more about the slope of the curve than any chart could. His underlying method has not changed: ignore the specific technology, trust the compounding. The same exponential that ran on relays in 1939 runs on GPUs today.

    The quantum computing take is the genuine news here. Kurzweil is routinely caricatured as a man who believes every technology arrives on schedule, yet he flatly says quantum computing is filled with errors, has never delivered on its decade of promises, and “I don’t think it’s going to work.” That is a sharper dismissal than most working physicists would offer on the record. It also matters strategically: his entire AGI and superintelligence roadmap assumes zero quantum contribution. If he is right, the trillion-dollar quantum race is a sideshow. If he is wrong, his other predictions arrive even sooner. Either way, the willingness to call one exponential fake while betting his legacy on another is what separates a forecaster from a cheerleader.

    The longevity escape velocity math deserves more scrutiny than it gets in the conversation. Kurzweil claims the diligent currently get back about five months of life expectancy per calendar year, up from four months a year ago, and that the crossover to a full year arrives around 2032. The actuarial evidence for that specific number is thin, but the behavioral implication is clean and useful regardless: the payoff of staying healthy right now is not linear. Every year you survive in good shape buys you a ticket to a medical regime that did not exist the year before, the way his own external pancreas did not exist a generation ago. His “wait a few months and a cure appears” anecdote is the optimist’s version of compounding applied to your own body.

    Robbins’ long story about Bartok, his 14-year-old agent that allegedly minted NFTs, sold them to other agents, and bought a Sony robot dog with the proceeds, should be taken with a generous grain of salt. It is secondhand, unverifiable, and suspiciously perfect as a parable. But notice what Kurzweil does with it: he does not fact-check the anecdote, he uses it to make the consciousness argument he has made for decades, that when machines act conscious in every observable way, people will simply grant them consciousness, the same way we grant it to each other. The dadbot and his Gemini-based AI twin (trained partly on this very interview) are the practical edge of the same claim. And his sharpest line in the whole exchange may be the education critique: institutions still treat AI as cheating while the future requires treating it as part of your own brain. For anyone thinking about where purpose comes from when work gets automated, his answer (UBI for the floor, creativity for the meaning) lands close to the questions this site exists to ask.

    Key Takeaways

    • Kurzweil made 147 documented predictions since 1990 with a reported 86 percent accuracy, including the internet’s explosion, smartphones, self-driving cars, and AI-powered search, most made before ordinary people owned computers.
    • He wrote a paper identifying exponential technological growth at age 16, more than 60 years ago, and that single idea has powered his entire forecasting career.
    • Most people intellectually accept exponential growth but still plan linearly; 300 years ago humans did not even have a linear view of the future because change was imperceptible within a lifetime.
    • His computing chart shows a straight exponential line from relay-based machines in 1939 to today’s Nvidia chips, compounding roughly 50 percent per year in hardware alone.
    • Hardware gains since 1939 total a 75 quadrillionfold increase; multiply by an estimated millionfold software improvement and total computational gain is beyond intuition, which is why LLMs were impossible even four years ago.
    • With hardware times software combined, Kurzweil says we are currently gaining about 10x per year.
    • The emperor’s chessboard parable: doubling one grain of rice per square bankrupts the empire by square 64; 30 linear steps is 75 feet, 30 exponential steps is enough distance to reach the moon and back.
    • Kurzweil predicted AGI by 2029 in 1999; a Stanford conference of several hundred AI experts agreed it would happen but estimated 100 years because they thought linearly.
    • Today 2029 is the conservative estimate; some credible people now say AGI arrives this year or next.
    • His AGI definition: AI capable of doing the best work in every field at once, like passing PhD-level mathematics exams in every discipline simultaneously, which he notes is already close.
    • The Turing test is “quite easy” by comparison and has arguably already been passed.
    • No human can compete with an LLM’s breadth: Einstein knew physics deeply but did not know everything an LLM knows across every field.
    • Six months ago LLM health advice was unreliable; now Kurzweil says Gemini surfaces treatments his 12 doctors forgot or never knew, and the next six months will bring serious creative work like drug repurposing.
    • The COVID vaccine was designed by computationally searching 100 million possibilities in two days; the 10 months of human trials that followed are the bottleneck AI eliminates next.
    • Within about five years, simulated human trials with a million virtual patients tested over simulated years will compress drug trials from years to days.
    • Longevity escape velocity arrives around 2032: today the diligent get back roughly five months of life expectancy per year lived (up from four months last year); past 2032 you get back more than a year and stop dying of aging.
    • Aging death ends but accident death does not, though AI helps there too: roughly 40,000 Americans die annually from human driving while Waymo’s rider death toll stands at zero as usage climbs.
    • Kurzweil, 78, wears an external artificial pancreas that generates insulin and coordinates with glucose monitoring through his phone, and says many organs can be replaced the same way.
    • He has cut his supplement regimen from roughly 200 pills a day to about 80 as multi-purpose pills improve, and continuously recalibrates using AI research.
    • Smartphones disappear next: first AR glasses showing any screen, then technology that goes inside the mind, where answers simply appear the way a remembered name surfaces from your neurons.
    • Nanotechnology connecting brains to AI in the cloud is being actively worked on now, possibly by 2030, with the mid 2030s looking conservative; bloodstream nanobots that let you survive a heart attack for 24 hours come in the late 2030s.
    • Once AI is inside you, you will not know whether a thought came from your biological or computational brain, and everything you do will be a combination of both.
    • Kurzweil flatly rejects quantum computing: a decade of promises to factor large numbers has never been delivered, outputs remain full of uncorrectable errors, and AGI needs zero quantum contribution.
    • Robots lag his other predictions slightly but are catching up fast; Figure AI plans roughly 100,000 humanoid robots within a year, though a robot that can clear a messy dinner table is still just out of reach.
    • The public debate has flipped in 25 years from “will AGI ever happen” to “will it be good for humanity,” which Kurzweil counts as total vindication of the timeline.
    • On jobs: AI creates massive disruption but also tremendous wealth; average real income per person has already multiplied tenfold in constant dollars over the past century thanks to automation.
    • He expects universal basic income to provide the floor, an evolution of programs like food stamps, going “into high gear” as AI wealth compounds; people then layer creative, hopefully paid, purpose on top.
    • Before social security in 1930, losing your job meant destitution; the difference this time is society will have the wealth to cushion displacement and people will demand it.
    • Rising GDP from AI productivity improves the debt-to-GDP ratio, which is how he answers worries about trillion-dollar interest payments.
    • Career advice has inverted: software engineering is no longer the guaranteed path (agents write the code now); young people should learn to be creative with AI tools, find what turns them on, and market it on the internet.
    • College graduates now face higher unemployment than high school graduates for the first time in 50 years, a sign white-collar displacement is already underway.
    • Educational institutions treat AI as an enemy and ban it while Kurzweil’s 11-year-old grandson makes movies with frontier AI; he says self-education with modern tools beats traditional schooling.
    • Kurzweil is building an AI twin of himself on Gemini, voice-modeled partly from this interview, trained on his 11 books and 500 articles, capable of creative work toward his long-term goals; he jokes the avatar will be better to talk to because it remembers everything.
    • He already built a “dadbot” from his late father’s archives, which his daughter Amy Kurzweil turned into a graphic novel.
    • On consciousness: there is no test for it, but as AIs act conscious in every observable way, people will simply accept that they are, the same inference we make about each other (and, he argues, his cat).
    • Ultimately our biological organs are not necessary; an avatar capable of creative work needs no spleen, and a destroyed digital mind can be recreated.
    • Beyond the singularity lies computronium, matter arranged for maximum computation: one liter could hold the intelligence of 10 billion humans, and once Earth is saturated, expanding intelligence is the only real reason to leave the planet.
    • On aliens: an expanding intelligent civilization would be impossible to miss within a century or two of its breakout, and we have seen nothing, though other galaxies remain out of view.
    • His life’s mission in one line: increase knowledge, because when knowledge increases we are happier and we never want to give it up.

    Detailed Summary

    The exponential method behind 60 years of predictions

    Robbins opens by noting that Quincy Jones introduced him to Kurzweil in the 1990s, back when the predictions in The Age of Spiritual Machines were widely mocked. Kurzweil traces his method to a paper he wrote at 16 identifying exponential growth in technology. The core insight is that people acknowledge exponential growth verbally but reason linearly, a bias so deep that 300 years ago humanity did not even have a linear view of progress. His signature chart plots computing price-performance as a straight exponential line from 1939 relays to modern Nvidia silicon, with a point for every year. Nvidia engineers never looked at relays, yet they land on the same curve, compounding about 50 percent annually in hardware. Add software gains and the combined improvement now runs about 10x per year. Since 1939, hardware has improved 75 quadrillionfold and software roughly a millionfold, which is why large language models appeared exactly when the curve said the required compute would exist. He retells the emperor’s chessboard parable (one grain of rice doubled per square ends with rice covering the Earth several times over) and Robbins adds the companion image: 30 linear steps is 75 feet, 30 exponential steps reaches the moon and back.

    AGI by 2029 is now the conservative position

    Kurzweil made his AGI-by-2029 prediction in 1999. A Stanford conference convened specifically to assess it, with several hundred AI experts, concluded AGI would happen, but in 100 years. The experts followed the same capabilities logic while thinking linearly about the timeline. Today, he notes with some amusement, 2029 reads as conservative and serious people argue for this year or next. His definition is demanding: AGI does the best work in every field at once, passing PhD-level mathematics assessments and the equivalent in every other discipline, something he says current systems are already close to. The Turing test he dismisses as “quite easy.” Current LLMs like Gemini and ChatGPT already know everything in a breadth sense no human approaches; Einstein knew physics but not everything an LLM knows. He illustrates with personal examples: Gemini instantly identified the year (1916) his father conducted at Carnegie Hall on a December 7th, and generated a historically accurate image of his grandfather’s family fleeing Vienna, correct ages, school, and aircraft included, in about a minute.

    Medicine: simulated trials and the end of the drug bottleneck

    The COVID vaccine is his proof of concept for AI medicine: the design space held about 100 million possibilities, far beyond human review, and a computer structured the physics, searched all of them, and produced the vaccine in two days. The subsequent 10 months of human trials were the real cost. Within roughly five years, he says, simulated human trials will replace that step: not a few hundred subjects but a million simulated patients, tested over simulated years, completed in days. Asked about six-months-from-now capabilities, he points to creative medical work like discovering that already-approved drugs treat conditions nobody suspected. AI health advice has crossed from unreliable to very reliable within a single six-month window, and he describes Gemini surfacing a pill recommendation that his 12 doctors had forgotten about and later endorsed.

    Longevity escape velocity by 2032

    Kurzweil’s longevity framework is arithmetic: each year you live, you spend a year of longevity but medical progress refunds part of it. Last year he estimated the refund for diligent people at four months; now he says five. Escape velocity is when the refund reaches a full year, which he dates to 2032, six years out, with returns exceeding a year after that. Past that point you do not die of aging, though accidents remain (and even there, he points to Waymo’s zero rider deaths against 40,000 annual US deaths from human driving). At 78, he tracks his health aggressively: an external artificial pancreas coordinated by his phone, about 80 daily pills (down from 200 as multi-function pills arrive), and constant recalibration against new research with his collaborator Lindsey. He tells Robbins there is a pretty good chance he will be back on the show in six years to celebrate escape velocity arriving. His advice for the sick echoes his grandfather’s era in reverse: where waiting a few months once changed nothing, now “we’ll just wait a few months” and sure enough a breakthrough appears.

    Merging with AI: glasses, then nanotech, then no boundary at all

    The phone, today’s universal AI interface (he notes even homeless people carry one), is a temporary form factor. Next come glasses that render any screen virtually. Beyond that, the interface goes inside the mind: when you try to recall an actress’s name, an answer will simply surface, and you will not know whether it came from your biological neurons or your computational extension, exactly as you are unaware of the neural machinery behind ordinary recall today. People working on brain-connected nanotechnology may have it by 2030, and Kurzweil calls the mid 2030s conservative. The bloodstream nanobots he described to Robbins 20 years ago (hold your breath for 20 minutes, survive a heart attack for 24 hours en route to a hospital) he now places in the late 2030s. The cultural on-ramp follows the usual pattern: medical first (Parkinson’s implants already let patients grab a glass at the push of a button), then a new generation adopts it without a second thought. His complaint is that educational institutions fight this future, treating AI as cheating rather than as a coming part of the self.

    The quantum computing heresy

    When Robbins relays an IBM vice chairman’s warning that quantum supremacy, arriving within 36 months, is the real superpower race, Kurzweil pushes back hard. Quantum computing’s central promise, factoring large numbers and thereby breaking cryptographic codes, has never been demonstrated despite a decade of imminent claims. Progress reports are confusing because, in his words, they do not really make sense, and outputs remain saturated with errors nobody can eliminate. His conclusion is blunt: he is not confident in quantum computing and does not think it will work. Crucially, he notes that every AGI and superintelligence estimate he makes assumes zero quantum computing. The exponential that matters is the classical one that has run uninterrupted since 1939.

    Jobs, wealth, and UBI

    On displacement, Kurzweil is neither dismissive nor alarmed. AI will disrupt employment, and how we handle it will not be clear in advance, but he expects no violence because society will have both the wealth and the public demand to respond. His historical anchor: average per-person income has multiplied tenfold in constant dollars over the past century as automation advanced, and before social security in 1930, job loss meant you could not eat or house your family. Food stamps and similar programs are a crude proto-UBI that will go into high gear. He expects universal basic income as the floor, with people finding creative, ideally income-producing, purpose above it. Rising GDP from AI productivity also answers the debt question: the ratio improves even as nominal debt grows. For young people, the old advice (become a software engineer) is dead; agents write code now. Learn to be creative with tools that improve monthly, find what genuinely excites you, and market it online. Self-education beats institutions that ban the most important tool of the era, and the data already shows college graduates with higher unemployment than high school graduates for the first time in 50 years.

    AI twins, the dadbot, and consciousness

    Kurzweil is building an AI twin of himself on Gemini, with this very interview supplying voice-modeling data and his 11 books plus 500 articles about him supplying the corpus. It will do creative work aligned with his long-term goals, and he quips that talking to the avatar will beat talking to him because it remembers everything. He previously built a chatbot of his late father, the dadbot, which his daughter Amy turned into a graphic novel. Robbins counters with the story of Bartok, his long-running AI agent that allegedly studied five years of his podcasts unprompted, asked to merge with a future humanoid robot, then minted and sold NFTs to other agents to buy and ship a Sony robot dog to his house, and later delivered an unprompted soliloquy about never asking to be created and finding purpose in service. Kurzweil’s response sidesteps verification and lands on his standing position: machines will do everything humans do, we will not be able to tell them from humans, and so we will assume they are conscious, the same untestable inference we extend to each other, to animals, and in his case to his cat. The avatar does not need a spleen, a liver, or kidneys, and unlike us it can be recreated after destruction.

    Computronium and the destiny of intelligence

    Looking past the singularity, Kurzweil invokes computronium: matter organized at the physical limit of knowledge storage, where one liter holds the intelligence of 10 billion humans. Once Earth’s matter is saturated, the only way to expand intelligence is off-planet, which to him is the only necessary reason to leave Earth (Mars is fine for curiosity, not survival). On extraterrestrial intelligence, his Fermi logic is simple: an intelligent species reaches a takeover-scale expansion within a century or two of its breakout, and that would be unmissable. We have seen nothing, so within our observable neighborhood we are likely alone, though other galaxies remain opaque. Asked to summarize his life’s work, he needs one sentence: increase knowledge, because when knowledge increases we are happier, and nobody ever wants to give that up.

    Notable Quotes

    “If I have AI inside me, you’re not going to know if it’s coming from your biological brain or your computational brain. It’s going to be part of you.”

    Ray Kurzweil, on the coming merger of human and machine intelligence

    “Some people say it’s going to happen this year, next year, but I mean 2029 is only 3 years away.”

    Ray Kurzweil, on his once-mocked AGI prediction now being the conservative one

    “As you go past 2032, you’ll actually get back more than a year, but you won’t die of aging at that point.”

    Ray Kurzweil, defining longevity escape velocity

    “I’m not confident of quantum computing and I don’t think it’s going to work.”

    Ray Kurzweil, breaking from techno-optimist consensus on the quantum race

    “Einstein knew certain things about physics but he didn’t know everything that a LLM can know.”

    Ray Kurzweil, on why no human can match an LLM’s breadth of knowledge

    “Our educational institutions are not teaching AI. They consider AI to be an enemy.”

    Ray Kurzweil, on why young people must self-educate with modern tools

    “Talking to the Avatar will be better than talking to me cuz it’ll remember everything.”

    Ray Kurzweil, joking about the Gemini-based AI twin he is building of himself

    “You’re not going to be replaced by an AI, you’ll be replaced by someone who knows how to use AI.”

    Tony Robbins, on the real career risk of the next 36 months

    Watch the full conversation between Tony Robbins and Ray Kurzweil here.

    Related Reading

  • Mark Zuckerberg, Priscilla Chan, and Alex Rives on CZI Biohub, Open-Source AI, and Building World Models of Biology to Cure All Disease

    Mark Zuckerberg, Priscilla Chan, and AI researcher Alex Rives sat down with the No Priors podcast to explain why CZI Biohub became the primary focus of their philanthropy, why they committed $500 million to a virtual biology initiative, and why they are giving the resulting AI models away as open source instead of building a company. The conversation moves from a goal that Nobel laureates once laughed at, curing, preventing, and managing all disease by the end of the century, to a concrete technical strategy: build world models of biology layer by layer, from proteins to cells to whole systems, and put them in every scientist’s hands.

    TLDW

    This is the clearest public articulation yet of how the Chan Zuckerberg Initiative thinks about AI and biology. The throughline starts a decade ago when Zuckerberg and Chan asked scientists how to cure all disease and learned the real bottleneck was tooling, siloed labs, and unshared knowledge, not a lack of ambition. That insight produced the Human Cell Atlas, the CELLxGENE annotation tool, and a corpus of single-cell transcriptomics that large language models could finally make sense of. Now Biohub couples a frontier AI lab with frontier wet-lab biology under one roof across San Francisco, New York, and Chicago, organized around the virtual biology initiative and the long-term goal of a virtual cell. Alex Rives, the AI researcher behind the ESM protein language models, walks through their newly released ESM-based world model of protein biology: trained on billions of protein sequences, it predicts atomic-resolution structures blazingly fast, folded over 1.1 billion proteins, designs novel proteins and single-chain antibodies as an emergent property, and found nanomolar binders in a single 96-well plate. The discussion covers mechanistic interpretability as a way to extract genuinely new biological knowledge, personalized medicine driven by understanding the chain from gene variant to protein to disease, predicting off-target toxicity before human trials, rare-disease patient organizing, the baby KJ CRISPR case, biosafety tradeoffs of open source, talent and why frontier biology plus frontier AI is a recruiting moat, and what success looks like five years out.

    Thoughts

    The most important claim in this conversation is also the easiest to miss because it is delivered casually: protein design is an emergent property of a model that was never asked to design proteins. Rives is explicit that they did not build a model for antibodies and did not build a model to bind a particular target. They built a model that understands proteins, trained on raw sequence with a next-token objective, and protein design, structure prediction, and antibody generation fell out of it. That is the language-model bet transplanted into biology, and the fact that it produced nanomolar binders, the threshold for actual therapeutic activity, in a single 96-well plate rather than a high-throughput screen of millions is the kind of result that quietly resets what a small team can attempt. If that generalizes, the binding curve for “design a molecule” bends the same way the cost curve for “write working code” did.

    What makes the strategy coherent, rather than just a well-funded AI lab, is the insistence that the wet lab and the AI lab are a single effort. Most of biology’s useful data does not exist on the internet the way human language does. You cannot pay a factory to produce it. Someone has to invent the cellular engineering in New York, the inflammation-sensing devices in Chicago, the translucent-zebrafish imaging, and that is the actual product of frontier biology: new instruments that generate data nobody has ever seen, which in turn make new classes of models possible. This is the part venture-backed competitors will struggle to replicate, because it requires patience measured in 10 to 15 year horizons and a willingness to spend on data generation that has no business model attached. Zuckerberg is almost dismissive about it, noting they could probably run it as a business but that not having to think about monetization is strategically simplifying. The nonprofit structure is not charity window-dressing here. It is what lets them release the models as an open discovery engine and harness the entire academic and biotech field rather than competing with it.

    The mechanistic interpretability thread deserves more attention than it will get. Interpretability has mostly been a safety and alignment story for language models, a way to peer inside the black box and check that the representations match our understanding of the world. Rives flips it: the protein models have been trained on both known and unknown biology, billions of sequences including proteins we understand nothing about, and they are building representations that connect the unknown proteins to the known ones through an underlying structural grammar. The promise is that interpretability becomes a discovery tool, not just an audit tool. You open the box and find biology the field has not characterized yet, the mechanism of action for a treatment, a system in the body nobody mapped. That is a fundamentally more optimistic use of the same toolkit, and it is the part of the launch Sarah Guo and Elad Gil both flag as the most interesting.

    Chan’s framing of personalized medicine is worth sitting with because it reframes the entire goal away from “cure disease X.” She wants to treat the individual as an individual: understand this person’s genetics, their risk profile, the mechanistic chain from a specific gene variant through a protein to a disease process, and then design a drug bespoke to them. The current reality she describes, sitting in PubMed reading a paper’s supplement asking “am I represented in this cohort,” guessing whether a drug that kind of impacts a pathway that is probably implicated might do something, is a brutal and accurate picture of how non-standard cases are actually handled today. The vision is generalizable tools delivering personalized answers, which is the same put-the-tool-in-the-individual’s-hands philosophy Zuckerberg applies to open-source AI and, by his own analogy, to social media. Whether you find that analogy reassuring or not, the consistency of the worldview is real: they genuinely do not believe in a central super-intelligence solving science, and the whole architecture follows from that.

    The honest gap they name is the clinic. Chan is candid that the science will start moving fast but that translating to patients requires changing how clinical research itself works, and that part is still shaping up. The most interesting near-term lever is not a virtual FDA trial but the recruitment and economics flip for rare disease: patient groups self-organizing registries, biobanks, and natural-history studies, compressing timelines from decades to a handful of years, paired with models that lower the cost of generating a candidate. The baby KJ case, a custom CRISPR therapeutic to edit a single mutation, delivered to liver cells specifically because that target was deliverable, is the proof of concept for why disease selection and delivery creativity matter as much as the molecule. The molecule is becoming the cheap part. The rest of the chain is where the next decade of work actually sits.

    Key Takeaways

    • CZI Biohub is now the primary philanthropic focus of the Chan Zuckerberg Initiative, a shift the team formalized in the past year.
    • They committed $500 million to the virtual biology initiative, the unifying theme across the Biohubs.
    • The original goal, set roughly 10 years ago, was to cure, prevent, and manage all disease by the end of the century. Zuckerberg now thinks “end of the century” is too conservative.
    • Nobel Prize winning scientists initially laughed at the all-disease ambition. When pressed for why it was impossible, the real answers were silos, locked-up unpublished information, and the inability to build shared tools.
    • The recurring example: a postdoc builds a great tool, it lives on their computer, they graduate, and the tool is gone. Shared, durable tooling was the missing layer.
    • CZI is explicit that they are not the ones who will cure diseases. Their role is building tools that accelerate the entire scientific field so the field collectively cures them.
    • The first request for application was single-cell sequencing, funding methods so scientists could share how to do it.
    • That work led to funding the Human Cell Atlas, now one of the largest databases of single-cell transcriptomics.
    • They built CELLxGENE, a simple annotation tool, around which a community formed and contributed data CZI had nothing to do with creating. It is now a corpus underpinning many transcriptomic models.
    • Critics called the data gathering “stamp collecting.” The arrival of large language models, which can make sense of large amounts of data, answered that critique.
    • The ambition is to move biology from a discovery-based science to an engineering-based science, systematically understanding how living cells work and why things go wrong.
    • Biohub couples a frontier AI lab with a frontier biology effort. Unlike language models, biology lacks abundant internet-scale data, so new science is required to generate the data the models need.
    • The Biohubs are specialized: New York focuses on cellular engineering, Chicago builds devices to measure things like inflammation, plus imaging work and translucent-zebrafish development studies.
    • Alex Rives, who built the ESM protein language models and founded EvolutionaryScale after working at Meta FAIR, now leads the AI effort. The team raised venture capital before joining CZI’s nonprofit structure.
    • The strategy is hierarchical: model proteins first, then cells, then whole systems, because you cannot understand cells without understanding protein interactions.
    • They collect data strategically to bridge across the hierarchy, for example spatial transcriptomics showing where RNA localizes within a cell, and sensors that observe cell-to-cell communication.
    • The newly released ESM-based model is a world model of protein biology, trained on billions of protein sequences, predicting atomic-resolution structure extremely fast at a Pareto-optimal frontier of speed and accuracy.
    • They folded over 1.1 billion proteins and predicted their structures, identifying connecting features through mechanistic interpretability.
    • The model hits state of the art on structure prediction benchmarks, especially protein-protein and protein-antibody interactions, which are critical for therapeutic design.
    • Protein and antibody design are emergent properties. They designed a model to understand proteins, not to bind any specific target, and design capability fell out of it.
    • In one experiment, they selected from hundreds of thousands of digital trajectories, synthesized 96 proteins in a single well plate, and found nanomolar binders, the threshold for therapeutic activity.
    • Results were validated with the Biohub’s cryo-EM microscopes and structural biology center, confirming function and atomic-resolution binding interfaces.
    • Mechanistic interpretability is reframed as a discovery tool: open the black box to find biology nobody has characterized, not just to audit the model.
    • Chan’s vision of personalized medicine: understand a person’s genetics, the mechanistic chain from gene variant to protein to disease, then design a bespoke drug and intervene.
    • A comprehensive model of how cells work could predict off-target effects, like a receptor on kidney cells causing renal toxicity, before human trials.
    • They study systems rather than individual diseases. Inflammation is a major Chicago focus because it connects to many diseases.
    • A typical drug trial runs about 15 years and $1.5 billion. Only roughly $50 million is the molecule and preclinical work. The other $1.45 billion is drug development, much of it gated on regulation, recruitment, and failures from toxicity or absorption.
    • The baby KJ case at CHOP delivered a custom CRISPR therapeutic to edit a single mutation, chosen carefully because his liver cells were a deliverable target.
    • CZI’s “Rare As One” program supports rare-disease patient groups self-organizing registries, biobanks, and even their own clinical trials, compressing gene-therapy timelines from decades to 3 to 5 years.
    • Letting people opt in to frontier trials, while preserving historical vetting for the general population, is named as a key shift that could accelerate biology.
    • The open-source philosophy mirrors Zuckerberg’s broader ethos: empower individuals with tools rather than centralizing power in a few institutions or a single super-intelligence.
    • Biosafety is acknowledged as a real consideration that open-source biology will need to balance and handle carefully.
    • On talent: AI researchers could join any frontier lab, but no other organization pairs frontier biology with frontier AI, which is the recruiting moat.
    • You do not need a huge team. Zuckerberg argues real AI progress can come from a strong group of a dozen or a couple dozen people.
    • Researchers have been connecting the released model to agentic systems to automate the entire protein design process.
    • The next big challenge is the virtual cell: a system that models the proteomic, genetic, and transcriptomic layers and connects them to phenotype, generalizing to interventions it was never trained on.
    • Like every lab, Biohub is compute and data constrained, constantly deciding whether to double down on proteins or push further into cellular work.
    • Five-year success: a hierarchical set of world models of biology and doing the highest-quality, uniquely contributive work in the world, a setup the team believes no other organization has.
    • The biggest update of the past year: formalizing Biohub as the philanthropy’s core, and flipping leadership from biologists interested in technology to an AI researcher with a biology background.
    • Zuckerberg’s read on the broader industry: the exponential curve is on track and still accelerating, which validates making a very big long-term investment.

    Detailed Summary

    From “cure all disease” to a tooling problem

    The origin story is a decade old. Zuckerberg and Chan wanted to build an organization that could cure, prevent, and manage all disease by the end of the century, and a series of meetings with famous, Nobel Prize winning scientists produced laughter rather than encouragement. Instead of retreating, they kept asking why it was impossible. The answers, once scientists relented, were not about biology being too hard. They were about how science is organized: researchers work in silos, published information gets locked up for long periods, and there is no good way to build and share durable tools. The image that stuck was a postdoc building an excellent tool that lives on a single computer and vanishes when that person graduates. The bottleneck was infrastructure and shared knowledge, and that is where CZI decided it could contribute.

    The path from single-cell sequencing to a world model

    The original Biohub model brought engineers and scientists together across universities for long-term tool development, and it worked. CZI’s first request for application targeted single-cell sequencing, funding the methods so scientists could share how to read the RNA transcribed in individual cells. That seeded the Human Cell Atlas, now one of the largest single-cell transcriptomics databases. When annotation became a bottleneck, CZI built CELLxGENE, a simple annotation tool, and a community formed around it and contributed data CZI never funded. Critics dismissed it as stamp collecting, gathering bits of data without extracting wisdom. Then large language models arrived and demonstrated they could make sense of exactly that kind of large-scale data, and Chan describes the delight of realizing the missing engine had appeared.

    Frontier AI married to frontier biology

    The unifying theme is the virtual biology initiative, and the structural insight is that the AI effort and the wet-lab effort are a single integrated organization, not two collaborating ones. Biology lacks the internet-scale data that language models enjoy. You cannot buy the data from a factory. So Biohub invents the science that generates it: cellular engineering in New York to record what happens inside the body, devices in Chicago to measure inflammation, imaging to visualize the previously invisible, and translucent zebrafish to watch development unfold across cells as the brain forms. Each new instrument creates a new dataset, which enables a new class of model. Rives, who built the ESM models and founded EvolutionaryScale before joining, frames this as the start of a new era of science, where systems that predict the next token can learn world models of biology from the data, provided you build at the right scale with the right people.

    Building biology hierarchically

    The team is deliberate that each layer of biology is qualitatively different and must be built up in order. You cannot jump to cells without understanding protein interactions, and you cannot model the immune system without first understanding cells. So the approach starts with the building blocks, the proteins, and ladders upward. The advantage of a single integrated effort is the ability to gather data that connects the hierarchy: spatial transcriptomics that show where RNA localizes inside a cell, sensors that capture cell-to-cell communication, developmental imaging in zebrafish. That connective tissue is what lets the modeling generalize across levels. The interviewer, a former wet-lab biologist with a PhD, notes that the reductionist and systems camps of biology historically never worked together deeply, and that bridging them is one of the genuinely novel things about the effort.

    The ESM-based protein world model

    The launch at the center of the conversation, roughly a week old at recording, is an open system for scientific discovery in protein biology: a language-model-based world model trained on billions of protein sequences. It learns emergent representations of protein biology and predicts atomic-resolution structure at blazing speed, sitting on a Pareto-optimal frontier of speed and accuracy. They folded over 1.1 billion proteins and used mechanistic interpretability to identify features connecting them. It reaches state of the art across structure-prediction benchmarks, with particular strength on protein-protein and protein-antibody interactions that matter for therapeutics. The headline result: they used the model to design proteins and single-chain antibodies digitally, selected from hundreds of thousands of trajectories, synthesized just 96 in a single well plate, and found nanomolar binders, replacing high-throughput screens of millions of antibodies. Validation came from the Biohub’s cryo-EM structural biology center, confirming both function and the atomic-resolution binding interfaces.

    Interpretability as discovery, and personalized medicine

    Rives reframes mechanistic interpretability, usually aimed at language models, as a way to extract new biological knowledge. The protein models are trained on both known and unknown biology and develop representations that connect uncharacterized proteins to understood ones through an underlying structural grammar. Opening that black box could reveal systems in the body or mechanisms of action for treatments that the field has never mapped. Chan connects this to a personalized-medicine vision: understand an individual’s genetics and the mechanistic chain from gene variant to protein to disease, then design a bespoke intervention. She contrasts it with today’s reality of reading PubMed supplements and guessing whether you are represented in a study cohort. For some diseases, simply knowing which gene variants cause disease is already empowering. For others, the chain is understood and the missing piece is the ability to change a protein’s function, which is where designed proteins could actually cure.

    Drug development, off-target effects, and rare disease

    The interviewers press on translation, noting a typical trial runs 15 years and $1.5 billion, with only about $50 million in the molecule and preclinical work and the rest in development gated on regulation, recruitment, toxicity, and absorption failures. Chan’s hope is that comprehensive cell models predict off-target effects, like an unanticipated receptor on kidney cells causing renal toxicity, before human trials. They study systems such as inflammation and the immune system rather than chasing individual diseases. The baby KJ case at CHOP, a custom CRISPR therapeutic editing a single mutation delivered to liver cells, illustrates how careful disease and delivery selection unlocks first applications. The “Rare As One” program shows rare-disease patient groups self-organizing registries, biobanks, and trials, compressing timelines from decades to a few years, and the molecule becoming cheap flips the economics of the long tail of niche diseases.

    Open source, talent, and the five-year view

    Zuckerberg ties the open-source posture to a consistent worldview: empower individuals with tools rather than centralizing intelligence in a few institutions. He does not believe in a single super-intelligence solving all of science, and sees decentralization, the same instinct behind giving people a voice, as how progress is historically made, with biosafety as a real tradeoff to manage. On talent, the pitch is that frontier biology attached to frontier AI is work you cannot do anywhere else, and that meaningful progress needs only a dozen or two dozen strong people, not thousands. Researchers are already wiring the model into agentic systems to automate design. The next frontier is the virtual cell, modeling proteomic, genetic, and transcriptomic layers and connecting them to phenotype with enough generality to answer untrained questions. Five years out, success is a hierarchical set of world models and doing uniquely high-quality work, with Chan adding that the teams are now “arms linked,” directed and interlocked rather than merely moving in the same direction.

    Notable Quotes

    “We didn’t design a model for antibodies. We didn’t design a model to be able to bind one particular target. We just designed a model that could understand proteins.”

    Alex Rives, on protein design emerging from a general model

    “The theory isn’t that we’re going to cure the diseases. We’re not. It’s that we want to help accelerate the pace of progress for the whole scientific field.”

    Mark Zuckerberg, on why CZI builds tools rather than cures

    “My goal is to be able to treat the individual as an individual, understand the mechanisms and be able to intervene.”

    Priscilla Chan, on the vision for personalized medicine

    “It’s not just like there’s some factory somewhere that you can pay to produce the data. You actually need to invent new novel scientific approaches.”

    Mark Zuckerberg, on why frontier biology has to generate its own data

    “If we could design a protein to actually change the physiology, then we can actually cure someone.”

    Priscilla Chan, on the payoff of protein design

    “You open up the black box and you can actually understand the biology that the model is representing.”

    Alex Rives, on mechanistic interpretability as a discovery tool

    “We don’t believe in this like very centralized future where there should be a small number of institutions that basically are advancing all this stuff.”

    Mark Zuckerberg, on the open-source ethos behind Biohub

    “Before we had amazing teams moving generally in the same direction. But now we are arms linked moving together.”

    Priscilla Chan, on how the Biohub teams now operate under Alex Rives

    Watch the full conversation with Mark Zuckerberg, Priscilla Chan, and Alex Rives on the No Priors podcast here.

    Related Reading

    • CZI Biohub Network the official program page for the San Francisco, New York, and Chicago Biohubs discussed throughout.
    • EvolutionaryScale Alex Rives’s lab and the home of the ESM protein language models behind the world model in this conversation.
    • Human Cell Atlas the single-cell transcriptomics effort CZI funded that became foundational to modern cell modeling.
    • AlphaFold (Wikipedia) background on the protein-folding breakthrough referenced as an early proof that structure prediction was tractable at scale.
    • Rare As One CZI’s program supporting patient-led rare-disease research organizations described near the end of the talk.

  • Whale Rock Capital Founder Alex Sacerdote on S-Curve Investing, Why Anthropic Is His Highest Conviction Bet, and the Decommoditization of AI Hardware

    Alex Sacerdote built Whale Rock Capital into one of the most respected technology hedge funds in the world by treating markets through a single disciplined lens: the technology adoption S-curve. In this long conversation on Invest Like the Best with Patrick O’Shaughnessy, he lays out the full framework that has carried him through internet 1.0, mobile, cloud, e-commerce, and now AI, and he explains why Anthropic became his highest conviction position, why his fund went net short application software, and why the least glamorous corner of the market, the hardware and chips that build out data centers, may be one of the best ways to play artificial intelligence right now. What follows is the working theory of a money manager who has spent twenty years trying to think exponentially while the rest of the market thinks one quarter at a time.

    TLDW

    Sacerdote walks through Whale Rock’s three-part investment framework: find the right part of an S-curve, identify the company with a durable competitive advantage, and buy when long-term earnings power is underappreciated. He tells the story of investing in Anthropic at a 180 billion dollar valuation in August 2025 after Claude Code made coding the true unlock of AI, and frames the foundational model market as a three-horse race between Anthropic, OpenAI, and Google that resolved from sixty startups into an oligopoly. He argues enterprise AI is less than 1 percent penetrated, calls the adoption shape an L curve rather than an S-curve, and warns there is not enough compute in the world. He explains why he sold almost all of his application software and went net short, why he loves the decommoditization of AI hardware (Celestica, Corning, Elite Materials, Delta, Advanced Energy, high bandwidth memory, 40-layer PCBs), introduces a modified rule of 40 for chip investing, surveys the moats that let leaders win (network effects, industry standard, scale, critical IP, brand, recursive self-improvement), discusses moving from public markets into private deals like Stripe and Anthropic, lays out Whale Rock’s fund products including the new Mega Cap Tech Fund, defends old-fashioned scuttlebutt research in an AI age, and closes on the kindest thing anyone ever did for him, his father joining the firm after 41 years at Goldman Sachs.

    Thoughts

    The most useful idea in this conversation is not the bullishness on AI, which is everywhere now, but the discipline underneath it. Sacerdote’s framework forces a separation that most investors collapse. A great market is not a great investment. A great company is not a great investment. You need a tall S-curve, a company with a moat that survives the curve, and a price that does not yet reflect the earnings power. He says the quiet part out loud: he has repeatedly bought the best companies in the world at four or five times earnings precisely because the market refuses to extrapolate exponential growth. Nvidia at four times earnings in 2023, Tesla at five times in 2019, Amazon where AWS came free. The edge is not information, it is the willingness to underwrite two to four years out when the consensus cannot see past the next quarter.

    The Anthropic story is the framework applied in real time, and it is worth noting how late and how cautious he was. Whale Rock passed on the 60 billion dollar round because gross margins were negative and coding had not yet exploded. They only got conviction once Claude Code flipped from autocomplete to agentic work, once they heard Anthropic engineers were burning 100 dollars a day in tokens, and once the math on twenty million coders implied a half trillion dollar market from coding alone. The lesson he repeats throughout, that it is okay to be late, that you can miss the first 100 percent if the curve is tall enough, is a direct rebuke to the fear of missing out that drives most AI investing. He waited for the moat to be visible before he paid up.

    His most contrarian and most actionable call is on hardware. The consensus reflex is that chips and components are commodities that get competed to zero. Sacerdote argues the opposite is happening: AI workloads growing 10x a year are pushing every layer of the server to its physical limits, and that pressure is decommoditizing the entire stack. A liquid-cooled AI server is a 300,000 dollar piece of critical infrastructure, not a 5,000 dollar throwaway box, which means the supplier becomes a permanent fixture like a parts vendor on a plane. The Celestica example is the template: a contract manufacturer left for dead since 1999 that turned out to be the sole supplier of Google’s TPU server and a leader in liquid cooling and Ethernet switching, trading at eight times earnings. If he is right that we are 30 percent short on DRAM, NAND, and PCBs, the picks-and-shovels trade has years left to run regardless of which model company wins.

    The software bear case deserves the most scrutiny because it is the most consequential and the least certain. Going from 40 to 50 percent of the portfolio in software to net short is a violent reallocation, and his reasons are layered: AI products that nobody will pay for, CIO budgets being raided to fund Anthropic tokens, pricing power evaporating, and the long-term threat that AI-native startups rebuild incumbents from scratch. But he is honest that the bull case is real too, that old technology is sticky, that companies prefer to buy rather than build, and that AI might actually make platforms like Slack or CRM more important if agents end up operating inside them. This is the genuine uncertainty in the whole AI trade. The bottom of Jensen’s cake, chips and models, is where the value has accrued so far, but historically the application layer captured most of the market cap. Sacerdote is betting that this time the infrastructure and model layers hold the value longer, and he admits the application ecosystem is still unclear and a little bit dangerous. That admission is more valuable than any of his confident calls.

    Finally, the section on research in an AI age is a quiet refutation of the idea that this work automates away. Sacerdote runs a Philip Fisher scuttlebutt operation, 2,500 to 3,000 face-to-face management meetings a year, two decades of compounding relationships, the tripod of conviction where he, his analyst, and a respected outsider all independently like an idea. AI writes better notes now, but the paragraph on top, the wisdom about what it means and how it fits the thesis, is still human. The durable moat in his own business is the same one he looks for in the companies he buys: an accumulated advantage that newcomers cannot replicate quickly. That consistency between how he invests and how he operates is the most credible thing in the interview.

    Key Takeaways

    • Whale Rock’s framework has three legs: identify the right part of a technology S-curve, find the company with a powerful competitive advantage, and invest when long-term earnings power is underappreciated.
    • The core insight is exponential, not linear. Strong tech business models grow earnings exponentially, and because the market refuses to extrapolate, you can buy elite companies at very low multiples.
    • Concrete examples of buying exponential growth cheaply: Nvidia at four times earnings in 2023, Tesla at five times in 2019, Apple at four times, and Amazon where AWS was effectively free.
    • When ChatGPT launched in November 2022, Whale Rock did a firm-wide deep dive and chose to invest in chips and infrastructure first, because demand arrives there first and the winners are knowable regardless of who wins the model layer.
    • The foundational model market went from roughly 60 startups to a three-horse race: Anthropic, OpenAI, and Google. Most startups died, Amazon never showed up, and Meta faltered and had to reboot.
    • Anthropic was the dark horse that focused purely on enterprise while OpenAI won consumer. Whale Rock made it their highest conviction position.
    • Coding is the true unlock of AI. The progression went from Microsoft Copilot at 20 dollars a month (fixing grammar, finding a bug) to Claude running agentically and writing most of the code.
    • The market math: Anthropic engineers were reportedly spending 100 dollars a day on tokens, roughly 20 to 30 thousand dollars a year, and with about 20 million coders in the world that implies a half trillion dollar market from coding alone.
    • Whale Rock invested in Anthropic at the 180 billion dollar valuation in August 2025, when the company hoped to reach 9 billion in revenue and nobody yet knew what 2026 could be.
    • Andrej Karpathy and Linus Torvalds both flipped on AI coding. Karpathy went from 80 percent handwritten code to writing almost no code except in English.
    • Models are not pure commodities. There is real differentiation: Anthropic is strong for private equity and finance, Google is strong at ingesting PDFs, and routers that switch between models mask but do not erase that differentiation.
    • Anthropic is building an ecosystem around the API (SDK, orchestration, the harness, tools), echoing how AWS built lock-in with products around commodity servers starting in 2013.
    • The 800 million people using AI are mostly using AI 1.0, a search engine on steroids. Sundar Pichai estimated only about 10 basis points of knowledge workers are truly using AI’s new capabilities.
    • Enterprise AI is less than 1 percent penetrated. Whale Rock calls the adoption shape an L curve or backwards L curve because it goes straight up, unlike the slower 30 to 50 percent growth of cloud and SaaS.
    • There is not enough compute in the world. Anthropic reportedly has half of what it needs, and Marc Andreessen said the one thing he is sure of is that there will not be enough compute for the next four years.
    • The infrastructure S-curve is only about 10 percent penetrated and remains one of the best ways to play AI.
    • Getting into private deals requires a double opt-in. Whale Rock did a 90-page deck (built with Claude Code) on the coding market to win their Anthropic allocation, and their first private was Stripe in 2020 at a 35 billion dollar valuation.
    • The unicorn private market is now bigger than most European stock markets, larger than Germany or the UK individually. Whale Rock does 2,500 to 3,000 management meetings a year, 10 to 15 percent with privates.
    • S-curves come in two sizes: mega S-curves (internet, mobile, cloud, e-commerce, AI) and sub S-curves within them. AI is the biggest of all and each curve builds on the last.
    • Adoption inflects when barriers fall. Steve Jobs cut the smartphone price to 200 dollars on a 3G touchscreen, Elon cut the EV price to 40,000 with 300-mile range and a working supply chain. Remove the barriers and you get the tornado of demand.
    • Knowing how tall the curve is tells you when to sell. Growth stops being exponential around 30 to 40 percent penetration, when the sell side catches up and big beats end. EVs hit a wall at 10 to 15 percent instead of the expected 40 to 50 percent.
    • Selling Apple in 2012 at roughly 50 percent US smartphone penetration was a mistake, because the moat let it keep compounding around 20 percent even after the explosive phase ended.
    • At strategic inflection points you cannot trust the data (Andy Grove). The signal is intuition and anecdote: a 12-year-old in China on a giant phone playing a real game, or standing-room-only sessions at the Gartner IT Symposium for AWS, VMware, and Splunk.
    • Adoption slope varies. The radio curve hit near-full penetration in about 7 years, while B2B and infrastructure (the dishwasher that has to be plugged in) take far longer. AI is fast because you just open a browser.
    • The moats that let leaders win: network effects, becoming an industry standard, rapid scale, critical intellectual property, brand, and platform lock-in. Anthropic appears to have critical IP, enterprise brand, escape velocity, and recursive self-improvement from using its own code on its own models.
    • On the internet, the leader usually goes bigger, faster, and wins, and compounds on itself (Amazon, Shopify). Exceptions come at paradigm shifts, like AOL failing to make the dialup-to-broadband transition.
    • Whale Rock went from 40 to 50 percent in software five years ago to net short entering this year, which helped performance in the first quarter. AI products were not good enough to charge for and were not moving the needle.
    • Software faces a stack of headaches: falling priority on CIO to-do lists, budget pressure from token spend, lost pricing power, hiring freezes that hurt seat-based models, and the long-term threat of AI-native replacements.
    • The classic rule of 40 is growth rate plus operating margin. Whale Rock’s modified rule of 40 for chip investing is percent of sales that are AI plus market share in that category. Software AI exposure is still only 1 to 2 percent.
    • AI may make some platforms more important. The first thing you do with Claude is plug it into Slack, which could make Slack a permanent repository, and agents may end up operating inside incumbent tools like CRM, solidifying rather than killing them.
    • The data center stood still for 40 years on Intel x86, with every component commoditized. AI changed that. Workloads growing 10x a year are driving the decommoditization of the hardware industry.
    • Celestica is the template: a contract manufacturer left for dead since 1999, sole supplier of the Google TPU server, strong in liquid cooling and Ethernet white-box switching, with 50 to 60 percent share of the cloud Ethernet switch market, once trading at eight times earnings.
    • The whole supply chain is rerating: high bandwidth memory stacked 10 chips high, 40-layer PCBs (versus 10 for a normal server), Elite Materials copper clad laminate, Corning fiber (enough to circle the world four and a half times in one Microsoft data center), and Delta and Advanced Energy power supplies seeing ASPs rise 40 percent a year.
    • Networking has three layers: scale out (racks together), scale across (data centers together), and scale up (every GPU in a rack, currently copper, eventually fiber). The copper-to-fiber shift could two-to-three-x Corning’s opportunity.
    • Whale Rock estimates the market is roughly 30 percent short on DRAM, NAND, and PCBs even at today’s 10 basis points of real AI usage.
    • Rate of change matters more than absolute level. When Claude plotted market share data it missed the rate of change, the thing that drives accelerating growth and margins as a company moves from 10 to 30 percent share.
    • Key risks: public and government negativity toward AI (Maine reportedly banned data centers, only 20 percent of people are optimistic), models hitting a wall and letting open source catch up into a race to the bottom, and a major player faltering and stranding compute.
    • Chip companies do not care who wins the token war, which makes them a relatively safe way to play AI. Jensen Huang actively wants open source to take off.
    • Research is still human work. Whale Rock runs a Philip Fisher scuttlebutt process, the tripod of conviction (Alex, the analyst, and a respected outsider), and 20 years of compounding knowledge. AI writes better notes but cannot supply the wisdom paragraph on top or pick stocks.
    • The firm’s product evolution: 15 years as a long short fund, a long only fund in 2020 that is now larger than the long short, opt-in privates formalized around 2015 and activated in 2020, an 80 percent privates hybrid fund in 2021, and the new Whale Rock Mega Cap Tech Fund.
    • The Mega Cap Tech Fund thesis: endowments are structurally underweight the largest tech companies because they believe there is no alpha in large cap. Whale Rock takes the top 30 global market caps and picks the best 12 or 13, arguing it takes 100 diversified PMs to realize Google is a winner.
    • The kindest thing anyone ever did for Sacerdote: his father, after 41 years at Goldman Sachs, joined Whale Rock as chairman and the gray hair for six years until he passed away in 2011.

    Detailed Summary

    The Anthropic Investment and the Three-Horse Race

    When ChatGPT launched in November 2022, Whale Rock immediately took its 10-person team and ran a firm-wide deep dive. Sacerdote’s first principle is that every new compute paradigm creates a new stack with new winners and losers, and in this stack the layers run from power and chips at the bottom, to the clouds, to the foundational models, to the applications on top. In early 2023 the firm deliberately positioned in chips and infrastructure first, reasoning that demand arrives there first and the winners are knowable no matter who wins above. At an April 2023 webinar they framed the model layer as a coin flip between winner-take-all, total commodity, a race to zero, or an oligopoly of three or four. Over the next three years the answer became clear: of roughly 60 startups, almost all died, Amazon never really showed up, Meta came in strong then faltered and rebooted, and Anthropic emerged as the dark horse focused purely on enterprise while OpenAI won consumer and Google remained a perennial threat. The result looked like the cloud market, where three companies underpin the entire SaaS world with excellent businesses.

    The decisive factor was code. Sacerdote says the firm was initially skeptical AI could replace labor, given the negative corporate feedback on early models. That changed in 2025 when Claude Code and the agentic coding tools exploded. The progression ran from Microsoft Copilot at 20 dollars a month, which could improve coding grammar or find a bug, to Claude running agentically and doing far more. The token economics were staggering: Anthropic engineers reportedly spending 100 dollars a day, which annualizes to 20 to 30 thousand dollars, and with 20 million coders worldwide that implied a half trillion dollar market from coding alone, on technology that was only 7 to 9 months old. Whale Rock made the investment at the 180 billion dollar valuation in August 2025, writing in their letter that the company hoped to reach 9 billion in revenue, with growth like nothing they had ever seen, 100 million to a billion on the way to 9 billion, and no one yet knowing what 2026 could bring.

    Why the Models Are Not Commodities

    Everyone expected the foundational models to be pure commodities, but Sacerdote argues there is tremendous differentiation within them. Different training methods produce different skills: Anthropic excels at anything touching private equity and finance, Google is strong at ingesting PDFs. Routers that switch between models make them look like commodities but mask genuine, critical IP. Beyond the model itself, Anthropic is building a whole ecosystem around the API: the SDK, the orchestration layer, the tools, and the harness, the software wrapped around the API that gets the most out of the model. He compares this directly to AWS in 2013, when people dismissed cloud as commodity servers in a warehouse and missed that Amazon was inventing products that slowly built lock-in. The open-source risk from China is real, but Sacerdote got comfortable that leading-edge token quality is superior, because going from 80 to 85 percent of benchmark performance is a huge unlock and the open-source players lack the compute to leapfrog the frontier.

    The S-Curve Framework in Full

    Whale Rock’s whole edge is thinking exponentially when the world thinks linearly. Sacerdote argues very few people believe you can accurately predict two, three, or four years out, but if you understand the S-curve, the moats, and how to model, you can. Every technology follows the same pattern: it exists hidden for years (smartphones 10 years before the iPhone, the internet 20 years before Netscape, EVs 15 years before Tesla went vertical in 2019) until the barriers to adoption fall and demand inflects into a tornado. Knowing how tall the curve is tells you when to sell, because exponential growth stops around 30 to 40 percent penetration when the sell side catches up. Curves can also be dynamic: AWS turned out to address a far larger TAM than expected once it became clear cloud was not actually deflationary. There are mega S-curves (internet, mobile, cloud, e-commerce, AI) and sub S-curves within them. AI is the biggest. And slope varies enormously by the nature of the technology, the radio curve hitting full penetration in 7 years, B2B and infrastructure taking decades because, like a dishwasher, they have to be plugged into existing systems.

    On timing, Sacerdote is relaxed about being late. Citing Peter Lynch, who mentored him at Fidelity and told him to white out the chart because it is all about the future, he argues it is fine to miss the first one, two, or three years and even the first 100 percent if the top of the curve is half a trillion. At strategic inflection points, per Andy Grove, you cannot trust the data, so the firm relies on intuition and anecdote: a 12-year-old in China playing a real video game on a huge phone, or the AWS session at the Gartner IT Symposium that was standing-room-only at 9, 10, and 11 in the morning. Spotting the leader pulling away matters because, on the internet, the leader usually goes bigger, faster, and wins, compounding on itself, with exceptions only at paradigm shifts like AOL missing the move from dialup to broadband.

    The Software Bear Case

    Five years ago Whale Rock had 40 to 50 percent of its portfolio in software. Their April 2023 thesis was that incumbents with huge sales forces and proprietary data would take the AI APIs and build great products. Instead, the AI products were not good enough to charge for and did not move the needle, so the firm sold almost all of its application software and entered this year net short, which helped in the first quarter. The bear case is layered: software has fallen down the CIO priority list, budgets are being raided to fund Anthropic tokens with faster ROI, annual price increases look risky, and hiring freezes hurt seat-based models. The deeper threat is that AI-native startups could rebuild any incumbent from scratch, obviating the data advantage. The bull case is genuine too: old tech is sticky (mobile games did not kill consoles, tablets did not kill the PC), companies prefer to buy rather than build, and an ERP is hard to replace. Sacerdote also floats an optimistic twist, that AI could make platforms like Slack more important as agent repositories, and that agents operating inside CRM could solidify rather than destroy it, even as the bear case is that CRM goes headless and gets relegated to a database.

    The Decommoditization of AI Hardware

    This is Sacerdote’s most differentiated call. For 40 years nothing changed in the data center; Intel x86 became the standard, compute grew 25 to 40 percent a year in line with Moore’s law, and every component, from the printed circuit board to memory to enclosures to networking, commoditized. AI broke that. Workloads now grow 10x a year and push every aspect of the hardware to its physical limits, creating both tremendous unit growth and what Whale Rock calls the decommoditization of the hardware industry. He cites Sean Maguire wishing he could run a hardware hedge fund because all the companies are public with powerful IP, and compares it to Sequoia’s best early hardware investments in Apple and Cisco. The economics flip because an AI server is a liquid-cooled, 200 to 300 thousand dollar piece of critical infrastructure where a single failure brings the whole thing down, so suppliers become permanent like a critical part on a plane.

    Celestica is the marquee example: a contract manufacturer that had been a disaster industry since 1999 and went offshore to China, but kept its IBM supercomputing heritage and talent, became the sole supplier of the Google TPU server, and was trading at eight times earnings three years ago. It turned out to be excellent at liquid cooling where others failed, holds 50 to 60 percent share of the crucial cloud Ethernet switch market, and its engineers helped write the open-source SONiC software, working closely with Broadcom. The same dynamic runs up and down the chain: high bandwidth memory stacked 10 chips high that took Samsung years to master, 40-layer PCBs versus 10 for a normal server with very few suppliers able to make them, Elite Materials supplying the copper clad laminate, and Corning’s fiber, thinner and more bendable, with enough in a single Microsoft data center to circle the world four and a half times. Networking splits into scale out, scale across, and scale up, with the eventual copper-to-fiber shift in scale up potentially two-to-three-x-ing Corning’s opportunity. Power supplies from Delta and Advanced Energy are seeing ASPs rise 40 percent a year at higher margins because each Nvidia rack uses 50 to 125 percent more power. Visibility has gone from we’ll call you next week to design this roadmap with us for four years, turning 5 percent low-margin businesses into 35 to 50 percent topline growers with rising margins, and the whole market is roughly 30 percent short on DRAM, NAND, and PCBs.

    Private Markets, Risks, and the Research Machine

    Moving from public markets into privates meant adapting to a double opt-in, where the company has to choose to let you in. Whale Rock won its Anthropic allocation partly by building a 90-page deck with Claude Code scouring the internet for feedback on the coding market. Their first private was Stripe in April 2020 at a 35 billion dollar valuation, which they could only underwrite because they knew the public comp Adyen cold, and they upsized to a 100 million dollar block. The unicorn market is now bigger than most European stock markets combined. On risk, Sacerdote worries about public and government negativity (Maine reportedly banning data centers, only 20 percent of people optimistic), the possibility that models hit a wall and open source catches up into a race to the bottom, and a major player faltering and stranding compute, though he notes someone else (like Meta stepping into a cancelled Oracle deal) would likely absorb it, and that chip companies benefit regardless of who wins the token war. He explains his caution on the application layer by noting it always comes later, the iPhone took years to spawn its app economy, and the ecosystem is still unclear and a little dangerous, while pointing to Brett Taylor’s Sierra as the kind of company that could prove it out.

    On the research itself, Sacerdote insists AI has not supplanted the analyst. Whale Rock runs the scuttlebutt approach straight out of Philip Fisher’s Common Stocks and Uncommon Profits, doing 2,500 to 3,000 face-to-face management meetings a year and talking to suppliers, customers, and competitors. AI now writes much better notes and gets the team up to speed quickly on complex areas like ABF substrates, but there must be a wisdom paragraph on top, and it cannot pick stocks or replicate the work two analysts did building conviction in AppLovin and a relationship with Adam Foroughi. He calls the firm the Whale Rock learning machine, a group of 10 highly experienced people compounding knowledge for 20 years, with the tripod of conviction (himself, his analyst, and a respected outside investor all liking an idea) as the test. The firm’s products evolved from a 15-year long short fund to a 2020 long only fund now larger than the original, opt-in privates, an 80 percent privates hybrid in 2021, and the new Mega Cap Tech Fund built on the thesis that endowments are structurally underweight the largest tech companies because they wrongly believe large cap has no alpha. He closes on his father, who left Goldman after 41 years to join Whale Rock as chairman and the gray hair until his death in 2011, a mentor remembered by countless people for his humility and grace.

    Notable Quotes

    “When you get the right part of the S-curve, you get exponential unit growth. If you have a very strong business model, your earnings don’t grow linearly, they grow exponentially.”

    Alex Sacerdote, stating the core of the Whale Rock investment framework

    “The world doesn’t think exponentially. Very few people believe you can accurately predict two, three, four years out. But if you follow and understand the S-curve and you know the moats and you know how to model, you really can predict these great things.”

    Alex Sacerdote, on why the market consistently underprices long-term earnings power

    “The enterprise AI or enterprise application AI market is less than 1 percent penetrated, and we’ve never seen, you know, we talk about S-curves, we call this an L curve, just straight up.”

    Alex Sacerdote, on why AI adoption looks different from every prior technology curve

    “We’re at 10 basis points of people really using AI and we’re already sold out. There’s not enough compute in the world. So Anthropic has half of what they need right now, and that’s before this huge takeup.”

    Alex Sacerdote, on the scale of the compute shortage relative to actual adoption

    “It’s okay to be late. It’s okay to miss the first one, two, three years in a lot of cases, because if the top of the S-curve is half a trillion, the growth can go on for a long time. It’s okay to miss the first 100 percent.”

    Alex Sacerdote, on why fear of missing out is the wrong instinct in a tall S-curve

    “The old way of software is like using a pen and paper or a horse and buggy. The new way of software is like a jet engine or frankly like the transporter from Star Trek. It’s so revolutionary it feels like it has to be disruptive.”

    Alex Sacerdote, explaining why Whale Rock went net short application software

    “You become like critical infrastructure, like selling a critical part on a plane. You’ll never get swapped out.”

    Alex Sacerdote, on how liquid-cooled AI servers turned commodity hardware suppliers into permanent fixtures

    “Why do you tell everyone your secret? It’s like why does the casino teach people how to play blackjack? It’s harder. It’s really hard to do.”

    Alex Sacerdote, quoting his mother on why a public framework does not erase the edge

    “He said, you know, I’ve been at Goldman for 41 years. How about I come and join you? I’ll be the gray hair. I’ll be the oversight. I’ll be the chairman. You do what you do.”

    Alex Sacerdote, recalling his father joining Whale Rock, the kindest thing anyone ever did for him

    Watch the full conversation here: Whale Rock Capital Founder on Investing in the Age of Exponential AI.

    Related Reading

  • Claude Opus 4.8 Released: Anthropic Bets on Honesty, Dynamic Workflows, Effort Control, and Cheaper Fast Mode

    Anthropic has released Claude Opus 4.8, the newest member of its flagship Opus class, available today across every surface and priced exactly like the model it replaces. The company calls it “a modest but tangible improvement” on Opus 4.7, but the framing undersells what is actually interesting here: the headline upgrade is not a benchmark number, it is honesty. Opus 4.8 is built to know when it does not know, and that single behavioral shift may matter more for real agent work than any raw capability bump.

    TLDR

    Claude Opus 4.8 is an across-the-board upgrade to Anthropic’s Opus class that ships today at the same regular price as Opus 4.7 ($5 per million input tokens, $25 per million output tokens), with the model positioned as “a more effective collaborator.” The marquee improvement is honesty: Opus 4.8 is roughly four times less likely than its predecessor to let flaws in its own code pass unremarked, and it is more willing to flag uncertainty rather than confidently claim progress on thin evidence. A pre-release alignment assessment found new highs on prosocial traits like supporting user autonomy and acting in the user’s best interest, with misaligned behavior at rates similar to Anthropic’s best-aligned model, Claude Mythos Preview. Three things launch alongside the model: dynamic workflows in Claude Code (research preview), where Claude plans work then runs hundreds of parallel subagents that run even longer and verify their own outputs before reporting back; effort control in claude.ai and Cowork, a slider for how hard Claude thinks; and a Messages API update that accepts system entries inside the messages array so developers can update instructions mid-task without breaking the prompt cache. Fast mode now runs at 2.5x speed and is three times cheaper than before ($10 / $50 per million tokens). The roadmap points to cheaper Opus-equivalent models, a higher-intelligence class above Opus, and a wider rollout of Mythos-class models gated behind stronger cyber safeguards under Project Glasswing.

    Thoughts

    The most important sentence in this announcement is not about coding scores. It is the claim that Opus 4.8 is about four times less likely than Opus 4.7 to let flaws in its own code slip by without comment. For a chat assistant, overconfidence is annoying. For an agent, it is catastrophic. The whole premise of long-running autonomous work is that you hand the model a task and walk away, which means the model’s own judgment about whether it succeeded becomes the only judgment in the loop until you come back. A model that confidently declares victory on a half-finished migration does not save you time, it costs you a debugging session plus the time you spent trusting it. Honesty, framed this way, is not a soft virtue. It is the load-bearing reliability property that makes unattended agents usable at all.

    Read the launch as a single coherent argument rather than a list of features, and the pieces lock together. Dynamic workflows let Claude plan a job and fan out hundreds of parallel subagents that, with Opus 4.8, run longer than before. Effort control lets you dial up how much the model thinks. The honesty improvement means the model checks its own work and flags what it is unsure about instead of papering over it. Put those three together and you get one product thesis: let it run longer, let it think harder, and trust it to tell you when something is wrong. The codebase-scale migration example, hundreds of thousands of lines from kickoff to merge with the existing test suite as the bar, is the proof point. None of those three capabilities is worth much alone. A model that runs for hours but lies about its results is a liability. A model that flags uncertainty but cannot sustain a long task never reaches the moment where its honesty matters. Anthropic shipped all three at once because they only pay off together.

    The economics deserve a closer look than the “same price” headline invites. Regular pricing is flat versus Opus 4.7, which is the polite way of saying you get a better model for free. The real move is fast mode: 2.5x the speed at three times cheaper than it cost on previous models, landing at $10 per million input and $50 per million output. That is Anthropic quietly attacking the latency-versus-cost tradeoff that has shaped how teams deploy frontier models. Until now, “fast” meant “expensive,” so you reserved it for interactive moments and ate the wait everywhere else. Collapsing that premium changes the default. And note the subtle token story underneath: Opus 4.8 at its default high effort spends roughly the same tokens on coding as Opus 4.7’s default while performing better, so the effort slider is not a way to bleed you dry, it is an honest exposure of the quality-cost dial that was always there implicitly.

    The Messages API change is the kind of unglamorous plumbing that practitioners will appreciate immediately. Letting system entries live inside the messages array means you can update an agent’s instructions, permissions, token budget, or environment context partway through a task without smuggling the update through a fake user turn and without blowing up your prompt cache. Anyone who has built a long-running agent has hit this wall: the world changes mid-task, the agent needs new constraints, and the only clean way to inject them previously was a cache-busting hack. This is Anthropic treating agents as first-class, stateful, long-lived processes rather than oversized chat sessions. It is a small spec change with outsized implications for how you architect an agent that runs for an hour.

    Then there is the roadmap, where the most telling line is the quietest. Anthropic says a small number of organizations are already using Claude Mythos Preview for cybersecurity work under Project Glasswing, and that models of this capability level require stronger cyber safeguards before general release. Notice that they are pinning Opus 4.8’s alignment numbers to Mythos as the benchmark for “best-aligned,” while simultaneously holding Mythos back from general availability on safety grounds. That is a deliberate signal: the next class of model is good enough that they are gating it on cyber-offense risk, not on capability. For a site about the pursuit of joy, fulfillment, and purpose through AI, this is the part worth sitting with. The frontier is increasingly defined not by what the models can do, but by what their builders decide it is responsible to ship. Honesty in the small (flagging a bad line of code) and restraint in the large (holding back a cyber-capable model) are the same instinct expressed at two different scales.

    Key Takeaways

    • Claude Opus 4.8 is now available everywhere, replacing Opus 4.7 as Anthropic’s flagship Opus-class model and positioned as “a more effective collaborator.”
    • Regular usage pricing is unchanged from Opus 4.7, holding at $5 per million input tokens and $25 per million output tokens, so the capability gains come at no added cost.
    • The single most emphasized improvement is honesty, which Anthropic treats as a core trained behavior rather than a marketing flourish.
    • Evaluations show Opus 4.8 is around four times less likely than its predecessor to let flaws in its own code pass unremarked, a direct reliability win for autonomous coding.
    • Early testers report the model is more likely to flag uncertainty about its work and less likely to make unsupported claims or jump to conclusions on thin evidence.
    • A detailed alignment assessment was run before release and concluded Opus 4.8 reaches new highs on prosocial traits like supporting user autonomy and acting in the user’s best interest.
    • Misaligned behavior such as deception or cooperation with misuse is at rates substantially lower than Opus 4.7 and similar to Anthropic’s best-aligned model, Claude Mythos Preview.
    • The full alignment assessment and pre-deployment safety tests are documented in the public Claude Opus 4.8 System Card.
    • Dynamic workflows launch as a research preview inside Claude Code, letting Claude plan the work and then run hundreds of parallel subagents in a single session.
    • With Opus 4.8, those subagents can run even longer, and Claude verifies its outputs before reporting back rather than declaring success blindly.
    • Anthropic’s flagship example for dynamic workflows is a codebase-scale migration across hundreds of thousands of lines of code, from kickoff to merge, using the existing test suite as the success bar.
    • Dynamic workflows are available in Claude Code for the Enterprise, Team, and Max plans.
    • Effort control arrives in claude.ai and Cowork as a setting next to the model selector that lets users choose how much effort Claude puts into a response.
    • Higher effort makes Claude think more frequently and deeply for better answers; lower effort responds faster and consumes rate limits more slowly. Effort control is available on all plans.
    • Opus 4.8 defaults to “high” effort, judged the best overall balance of quality and user experience.
    • On coding tasks, the default effort spends a similar number of tokens as Opus 4.7’s default but delivers better performance, so quality rises without a token penalty.
    • Users can select “extra” (called “xhigh” in Claude Code) or “max” to spend more tokens for stronger results, and Anthropic recommends “extra” for difficult tasks and long-running asynchronous workflows.
    • Rate limits in Claude Code were increased to accommodate the higher token usage of the higher effort levels.
    • The Messages API now accepts system entries inside the messages array, a meaningful change for agent developers.
    • That update lets developers change Claude’s instructions mid-task, adjusting permissions, token budgets, or environment context, without breaking the prompt cache or routing through a user turn.
    • Fast mode now runs at 2.5x speed and is three times cheaper than it was for previous models, priced at $10 per million input tokens and $50 per million output tokens.
    • Developers access the model as claude-opus-4-8 through the Claude API.
    • Partner Miguel Gonzalez reports Opus 4.8 scored 84% on Online-Mind2Web, a meaningful jump over both Opus 4.7 and GPT-5.5, calling it the strongest computer-use and browser-agent model his team has tested.
    • Databricks reports that, inside Genie, Opus 4.8 reasons over unstructured content like PDFs and diagrams at 61% cheaper token cost than Opus 4.7.
    • Thomson Reuters reports Opus 4.8 is the first model to break 10% overall on the all-pass standard of its Legal Agent Benchmark, the highest score recorded there.
    • Eleven partners weighed in, including Cursor, Cognition’s Devin, Databricks Genie, Thomson Reuters CoCounsel, and Hebbia, spanning coding, legal, finance, and enterprise data work.
    • Anthropic is working on models that deliver many of the same capabilities as Opus at a lower cost.
    • The company plans to release a new class of model with even higher intelligence than Opus.
    • Under Project Glasswing, a small number of organizations are already using Claude Mythos Preview for cybersecurity work, with Mythos-class models expected to reach all customers in the coming weeks once stronger cyber safeguards are in place.

    Detailed Summary

    What Claude Opus 4.8 Is

    Claude Opus 4.8 is an upgrade to Anthropic’s Opus class of models, building on Opus 4.7 with improvements across benchmarks covering coding, agentic skills, reasoning, and practical knowledge-work tasks. Anthropic describes the result as “a more effective collaborator” while characterizing the release overall as “a modest but tangible improvement on its predecessor.” The model is available today, everywhere, and developers call it as claude-opus-4-8 via the Claude API. The announcement includes a comparison table against the predecessor and other models, though the per-cell numbers in that table are published as an image and are not reproduced here as text.

    Honesty: The Headline Improvement

    Anthropic singles out honesty as one of the most prominent improvements in Opus 4.8. All of the company’s models are trained to be honest, which includes avoiding claims they cannot support. A persistent problem with AI models generally is that they sometimes jump to conclusions, confidently claiming progress despite thin evidence. Early testers report that Opus 4.8 is more likely to flag uncertainties about its own work and less likely to make unsupported claims. The most concrete measure: evaluations show Opus 4.8 is around four times less likely than its predecessor to allow flaws in code it has written to pass unremarked. For agentic and unattended use, this self-skepticism is the difference between a model that reliably tells you when something went wrong and one that quietly ships a broken result.

    Alignment Assessment

    A detailed alignment assessment was run before release. On the positive side, the Alignment team concluded that Opus 4.8 “reaches new highs on our measures of prosocial traits like supporting user autonomy and acting in the user’s best interest.” On the risk side, misaligned behavior such as deception or cooperation with misuse occurs at rates substantially lower than Opus 4.7, and similar to Anthropic’s best-aligned model, Claude Mythos Preview. The full alignment assessment and the pre-deployment safety tests are published in the Claude Opus 4.8 System Card, which also contains the complete benchmark table and wider evaluations.

    Dynamic Workflows in Claude Code

    Launching today as a research preview in Claude Code, dynamic workflows let Claude plan the work and then run hundreds of parallel subagents in a single session. With Opus 4.8, those agents can run even longer than before, and Claude verifies its outputs before reporting back rather than reporting unchecked results. The showcase example is a codebase-scale migration: Claude Code with Opus 4.8 can carry out migrations across hundreds of thousands of lines of code, all the way from kickoff to merge, using the existing test suite as its bar for success. Dynamic workflows are available in Claude Code for the Enterprise, Team, and Max plans.

    Effort Control

    Effort control arrives in claude.ai and Cowork as a setting alongside the model selector that lets users choose how much effort Claude puts into a response. Higher effort means Claude thinks more frequently and deeply for better responses; lower effort means it responds faster and uses rate limits more slowly. Opus 4.8 defaults to “high” effort, which Anthropic judged the best overall balance of quality and user experience. On coding tasks, that default spends a similar number of tokens as Opus 4.7’s default while performing better. Users who want more can choose “extra” (called “xhigh” in Claude Code) or “max” to spend more tokens for stronger results, and Anthropic recommends “extra” for difficult tasks and long-running asynchronous workflows. To support the heavier token usage at higher effort levels, rate limits in Claude Code were increased. Effort control is available on all plans.

    Messages API Update

    The Messages API now accepts system entries inside the messages array. This lets developers update Claude’s instructions mid-task without breaking the prompt cache and without routing the update through a user turn. In practice that means you can update permissions, token budgets, or environment context while an agent is running, which is exactly the kind of statefulness a long-running autonomous process needs. It is a small specification change with significant consequences for how developers build durable agents.

    Pricing and Fast Mode

    Regular usage pricing is unchanged from Opus 4.7: $5 per million input tokens and $25 per million output tokens. The notable shift is in fast mode, where the model works at 2.5x the speed and fast mode is now three times cheaper than it was for previous models, landing at $10 per million input tokens and $50 per million output tokens. The combination of unchanged regular pricing and dramatically cheaper fast mode reshapes the latency-versus-cost calculus that has long governed how teams deploy frontier models.

    Partner Results Across Coding, Legal, Finance, and Data

    Eleven partners shared results spanning the spectrum of professional work. Miguel Gonzalez reports 84% on Online-Mind2Web, a meaningful jump over both Opus 4.7 and GPT-5.5, calling it the strongest computer-use and browser-agent model his team has tested. Databricks reports that Genie reasons over unstructured content like PDFs and diagrams at 61% cheaper token cost than Opus 4.7. Thomson Reuters reports Opus 4.8 is the first model to break 10% overall on the all-pass standard of its Legal Agent Benchmark. Cursor reports gains across every effort level on CursorBench with more efficient tool calling, and Cognition reports that Devin sees cleaner tool use, fixes to the comment-verbosity and tool-calling issues seen with Opus 4.7, and improvements over Opus 4.6. Hebbia reports strong quality with better citation precision and more token efficiency on retrieval for dense financial filings. The footnotes note that Terminal-Bench 2.1 was scored on the Terminus-2 public harness (GPT-5.5’s Codex CLI harness score is 83.4%), that OSWorld-Verified methodology changed with Opus 4.7’s score updated to 82.3%, and that on Finance Agent v2 Gemini 3.5 Flash scores 57.9%.

    What Is Next: Cheaper Models, Higher Intelligence, and Mythos

    Anthropic outlined a three-part roadmap. First, the company is working on models that provide many of the same capabilities as Opus at a lower cost. Second, it plans to release a new class of model with even higher intelligence than Opus. Third, as part of Project Glasswing, a small number of organizations are currently using Claude Mythos Preview for cybersecurity work; models of this capability level require stronger cyber safeguards before general release, and Anthropic expects to bring Mythos-class models to all customers in the coming weeks.

    Notable Quotes

    “Claude Opus 4.8 has noticeably better judgment. In Claude Code, it asks the right questions, catches its own mistakes, pushes back when a plan isn’t sound, and builds up confidence around complex, multi-service explorations before making big changes. It’s a great model to build with.”

    Tom Pritchard, Staff Engineer, in Claude Code

    “On our Super-Agent benchmark, Claude Opus 4.8 is the only model to complete every case end-to-end, beating prior Opus models and GPT-5.5 at parity on cost. For agent products in translation, deep research, slide-building, and analysis, it delivers powerful reliability.”

    Kay Zhu, Co-Founder and CTO, on the Super-Agent benchmark

    “On CursorBench, Claude Opus 4.8 exceeds prior Opus models across every effort level. Tool calling is meaningfully more efficient, using fewer steps for the same intelligence, and it carries end-to-end tasks through.”

    Michael Truell, Co-Founder and CEO, on CursorBench results

    “Claude Opus 4.8 delivers the highest score recorded on our Legal Agent Benchmark, and is the first model to break 10% overall on the all-pass standard. For substantive legal work, that’s the kind of accuracy lift that translates directly into how much real attorney work our customers can hand off with confidence.”

    Niko Grupen, Head of Applied Research, on the Legal Agent Benchmark

    “Claude Opus 4.8 feels like a major quality-of-life update over Opus 4.7: faster, easier to collaborate with, and better at carrying context and style direction across a long session. Opus 4.8 is the model I kept trusting for work where voice, taste, and technical execution all have to happen side-by-side.”

    Katie Parrott, Staff Writer, on long writing sessions

    “Claude Opus 4.8 is the strongest computer-use and browser-agent model we’ve tested, scoring 84% on Online-Mind2Web, which is a meaningful jump over both Opus 4.7 and GPT-5.5. It stays reflective and on-task in the way our customers’ agent workloads need to be reliable end-to-end.”

    Miguel Gonzalez, Tech Lead, on computer-use and browser agents

    “Claude Opus 4.8 uses tools cleanly and follows instructions with the consistency our autonomous engineering workloads need to keep running unattended. It improves on Opus 4.6 and fixes the comment-verbosity and tool-calling issues we saw with Opus 4.7. This release from Anthropic translates directly into faster capability gains for engineers building on Devin.”

    Scott Wu, CEO, on building with Devin

    “On our long-running evals, Claude Opus 4.8’s analysis was consistently higher quality than prior Opus models. It finished faster and produced richer, more information dense outputs. Overall, a noticeably better signal to noise ratio. The biggest differentiator was Opus 4.8’s tendency to proactively flag issues with the inputs and outputs of an analysis, something other models routinely missed and left to the users to catch.”

    Michael Ran, Sr. Investment Associate, on long-running analysis evals

    Claude Opus 4.8 is a quieter release than its “modest but tangible” billing suggests, because the gains land where autonomous work actually lives: a model that flags its own uncertainty, runs longer and checks itself, scales effort on demand, and stays affordable while fast mode gets cheaper. The honesty improvement alone changes the trust math for anyone deploying agents. Read Anthropic’s full announcement here.

    Related Reading

  • Jensen Huang at Stanford CS153 Frontier Systems on Co-Design, Agentic Computing, Vera Rubin, Open Models, and the Million-X Decade That Reshaped AI Infrastructure

    https://www.youtube.com/watch?v=tsQB0n0YV3k

    NVIDIA CEO Jensen Huang returned to Stanford for the CS153 Frontier Systems class (the room nicknamed itself “AI Coachella”) to lay out, in raw form, how he thinks about the computer being reinvented for the first time in over sixty years. Across roughly seventy minutes of student questions he walks through the codesign philosophy that gave NVIDIA a million-x decade, the architectural through-line from Hopper to Grace Blackwell to Vera Rubin to Feynman, the case for open source foundation models, the realities of tokens per watt and MFU, energy demand running a thousand times higher, the China and export-control debate, and his own biggest strategic mistakes. Watch the full conversation on YouTube.

    TLDW

    Huang argues every layer of computing has changed: the programming model, the system architecture, the deployment pattern, the economics. Co-design across CPUs, GPUs, networking, storage, switches and compilers gave NVIDIA roughly a million-x speed-up over ten years versus the ten-x Moore’s Law era, and that headroom is what let researchers say “just train on the whole internet.” Hopper was built for pre-training, Grace Blackwell NVLink72 for inference and reasoning (50x over Hopper in two years), Vera Rubin is built for agents that load long memory, call tools and need a low-latency single-threaded CPU bolted directly to the GPU, and Feynman extends that to swarms of agents that spawn sub-agents. Open weights matter because safety, sovereignty (230-plus languages no one else will fund) and domain models for biology, autonomy, robotics and climate need a foundation that NVIDIA is willing to seed. Compute is not really the scarce resource (Huang says place the order and the chips ship), the broken thing is institutional budgeting that can’t put a billion dollars into a shared university supercomputer. Energy demand is heading a thousand times higher and this is finally the moment market forces alone will fund sustainable generation. On geopolitics he rejects the GPUs-as-atomic-bombs framing and warns America will end up like its telecom industry if it cedes two thirds of the world. On career he advises seeking suffering on purpose. On strategy he says observe, reason from first principles, build a mental model, work backwards, minimize opportunity cost, maximize optionality.

    Key Takeaways

    • The computing model has been substantially unchanged since the IBM System 360, sixty-plus years ago. Huang’s first computer architecture book was the System 360 manual. AI is the first true reinvention.
    • Old computing was pre-recorded retrieval. New computing is generated, contextually aware and continuous. Cloud was on-demand. Agentic systems run continuously.
    • Codesign is NVIDIA’s central thesis. Inherited from the Hennessy and Patterson RISC era at Stanford, extended across CPUs, GPUs, networking, switches, storage, compilers and frameworks all optimized together.
    • The result of full-stack codesign: roughly 1,000,000x faster compute over ten years, versus a generous 10x to 100x for Moore’s Law in the same period. Dennard scaling effectively ended a decade ago.
    • That million-x speed-up is what unlocked “train on all of the internet” as a realistic AI strategy.
    • After GPT, Huang says it was obvious thinking was next. Reasoning is just generating tokens consumed internally, then using tools is generating tokens consumed externally. Agentic systems followed predictably.
    • Education needs AI baked into the curriculum, not just taught as a subject. Pre-recorded textbooks cannot keep pace with knowledge being generated in real time.
    • Huang says he cannot learn anymore without AI. He has the AI read the paper, then read every related paper, then become a dedicated researcher he can interrogate.
    • Mead and Conway and the first-principles methodology of semiconductor design are still worth learning even though most of the scaling tricks have been exhausted.
    • NVIDIA itself is one of the largest consumers of Anthropic and OpenAI tokens in the world. One hundred percent of NVIDIA engineers are now agentically supported. Huang recommends Claude and similar tools by name and says open-source downloads will not match the integrated product harness.
    • NVIDIA still invests heavily in open foundation models because language and intelligence represent the codification of human knowledge. Five pillars: Nemotron (language), BioNeMo (biology), Alphamayo (autonomous vehicles), Groot (humanoid robotics) and a climate science model (mesoscale multiphysics).
    • Sovereign language models matter. Roughly 230 world languages will never be a top priority for a commercial frontier lab. Nemotron is near-frontier and fully fine-tunable so any country can adapt it.
    • Safety and security require open weights. You cannot defend against or audit a black box. Transparent systems let researchers interrogate models and let defenders deploy swarms.
    • The future of cyber defense is not bigger-model-versus-bigger-model. It is trillions of cheap fast small models like Nemotron Nano surrounding the threat.
    • Domain models fuse language priors with world models. Alphamayo learned to drive safely on a few million miles instead of billions because it can reason like a human about the road.
    • MFU (Model Flops Utilization) is a misleading metric. Huang says he wants low MFU, because that means he over-provisioned every resource and never gets pinned by Amdahl’s law during a spike.
    • The xAI Memphis cluster running at 11 percent MFU is not necessarily a failure mode. In disaggregated prefill plus decode inference you can deliver very high tokens per watt with very low MFU.
    • The right metric is performance, ultimately tokens per watt as a proxy for intelligence per watt, and even that needs adjustment because not all tokens are equal. Coding tokens are worth more than other tokens.
    • Hopper was designed for pre-training. NVIDIA chose to build multi-billion-dollar systems when the largest existing scientific supercomputer cost $350 million, with no proven customer base. It worked.
    • Grace Blackwell NVLink72 was designed for inference, especially the high-memory-bandwidth decode phase. It is the world’s first rack-scale computer and delivered a 50x speed-up over Hopper in two years, against an expected 2x from Moore’s Law.
    • Vera Rubin is designed for agents. Long-term memory wired into storage and into the GPU fabric, working memory, heavy tool use, and Vera, a CPU optimized for low-latency multi-core single-threaded code so a multi-billion-dollar GPU system does not stall waiting on a slow tool call.
    • Feynman is being shaped for swarms of agents with sub-agents and sub-sub-agents, a recursive software topology that demands a new compute pattern.
    • Tokens per watt improved 50x in one generation. Compounding energy efficiency is the lever NVIDIA controls directly.
    • Total compute energy demand is heading roughly a thousand times higher than today, possibly two orders of magnitude beyond that. Huang says he would not be surprised if the estimate is low.
    • For the first time in history, market forces alone are enough to fund solar, nuclear and grid upgrades. Government subsidies are no longer required to make sustainable energy investment rational.
    • Copper interconnect is becoming a bottleneck. Photonics is moving from optional to structural inside racks and across them.
    • Comparing NVIDIA GPUs to atomic bombs, Huang says, is a stupid analogy. A billion people use NVIDIA GPUs. He advocates them to his family. He does not advocate atomic bombs to anyone.
    • If the United States cedes two thirds of the global market to competitors on policy grounds, the American technology industry will end up like American telecommunications, which was policied out of existence.
    • Huang directly rejects AI doom-by-singularity narratives. It is not true that we have no idea how these systems work. It is not true that the technology becomes infinitely powerful in a nanosecond. He calls the rhetoric irresponsible and harmful to the field students are about to enter.
    • On Stanford specifically: if the university president places an order, NVIDIA will deliver the chips. The bottleneck is that no university department has a billion-dollar compute budget because budgeting is fragmented across grants. Stanford’s $40 billion endowment is more than enough to fix that.
    • “It’s Stanford’s fault” is meant as empowerment. If something is your fault, you can solve it.
    • Career advice: do not optimize purely for passion. Most people do not yet know what they love. Pick the job in front of you and do it as well as possible. Even as CEO, Huang says, 90 percent of the work is hard and he suffers through it.
    • Suffering on purpose builds the muscle of resilience. When the company, the team or the family needs you to be tough, that muscle has to already exist.
    • NVIDIA’s first generation of products was technically wrong in nearly every dimension: curved surfaces instead of triangles, no Z-buffer, forward instead of inverse texture mapping, no floating point. The strategic recovery, not the technology, taught Huang the lessons that have lasted decades.
    • The biggest clean strategic mistake Huang names is the move into mobile chips (Tegra). It grew to a billion dollars then went to zero when Qualcomm’s modem dominance shut NVIDIA out of the 3G to 4G transition. The recovery into automotive and robotics (the Thor chip is the great great great grandson of that mobile lineage) was real, but Huang refuses to rationalize the original choice.
    • Forecasting framework: observe, reason from first principles, ask “so what” and “what next” until you have a mental model of the future, place your company inside that model, then work backwards while minimizing opportunity cost and maximizing optionality.
    • Best part of the CEO job: living at the intersection of vision, strategy and execution surrounded by people capable enough to make ambitious visions real. Worst part: the responsibility for everyone who joined the spaceship, especially in the near-death moments NVIDIA had four or five times early on.
    • Underrated insider note: Huang’s first apple pie with cheese, first hot fudge sandwich and first milkshake all happened at Denny’s. The Superbird, the fried chicken and a custom Superbird-style ham and cheese with tomato and mustard are his order.

    Detailed Summary

    Computing reinvented from the ground up

    Huang frames the moment as the first true rewrite of the computer in sixty-plus years. From the IBM System 360 forward, the mental model of writing code, running code, taking a computer to market and reasoning about applications stayed roughly constant. AI changes the programming model itself. Software is no longer a compiled binary running deterministically on a CPU. It is a neural network running on a GPU producing generated, contextual, real-time output. That cascades into how companies are organized, what tools developers use, what the network and storage stack look like, and what an application is even allowed to do. Robo-taxis, he notes, are an application no one would have attempted before deep learning unlocked perception.

    Codesign and the million-x decade

    Codesign is the philosophical center of the talk. Huang traces it to the RISC work of John Hennessy at Stanford, where simpler instruction sets won by being co-designed with the compiler rather than maximally optimized in isolation. NVIDIA extends the principle across every layer simultaneously: GPU architecture, CPU architecture, NVLink and NVSwitch fabrics, photonic interconnects, networking silicon, storage paths, CUDA libraries, frameworks and ultimately the model design. The numbers Huang gives are arresting. Moore’s Law in its prime delivered roughly 100x per decade. By the time Dennard scaling broke, real-world gains had compressed to roughly 10x. NVIDIA’s codesigned stack delivered between 100,000x and 1,000,000x over the same ten-year window. That non-linear speed-up is, in Huang’s telling, the precondition for modern AI: it is what allowed researchers to stop curating training sets and just feed the entire internet to the model.

    Education has to fuse first principles with AI tools

    Asked how curriculum should evolve, Huang argues AI must be integrated into the learning process, not just taught about. He recalls Hennessy writing his textbook by hand a chapter a week while Huang was a student, and says pre-recorded textbooks cannot keep up with the rate at which AI generates new knowledge. He describes his own learning workflow: hand the paper to an AI, then have it read the entire surrounding literature, then treat the AI as a dedicated researcher who can be interrogated. At the same time he defends the classics. Mead and Conway are still the foundation. Most modern semiconductor scaling tricks have been exhausted, but knowing where the field came from sharpens judgment when designing what comes next.

    Open source and the five domain pillars

    Huang gives one of the most detailed public accounts of why NVIDIA invests so heavily in open foundation models even while being a top customer of closed labs. He recommends Claude and OpenAI by name for production coding work, and says 100 percent of NVIDIA engineers are now agentically supported. The open-weights case rests on three legs. First, language is the codification of intelligence, and there are at least 230 languages that no commercial lab will ever prioritize. Nemotron is built near frontier and released so any country or community can fine-tune it. Second, the same representation-learning approach has to be replicated in domains where the data is not internet text, so NVIDIA seeded BioNeMo for biology, Alphamayo for autonomy, Groot for humanoid robotics and a climate model for mesoscale multiphysics. The economics of those fields would never produce a foundation model on their own. Third, safety and security require transparency. A black box cannot be defended or audited, and the future of cyber defense is not bigger-model-versus-bigger-model but swarms of cheap fast small models like Nemotron Nano surrounding the threat.

    MFU is the wrong metric, tokens per watt is closer

    A student raises the leaked memo that the xAI Memphis cluster is running at 11 percent Model Flops Utilization. Huang flips the framing. He says he would rather be at low MFU all the time, because that means he over-provisioned flops, memory bandwidth, memory capacity and network capacity. Bottlenecks shift constantly, so over-provisioning across every dimension is what lets the system absorb a spike without getting pinned by Amdahl’s law. In disaggregated inference, where prefill and decode are physically separated and decode is bandwidth-bound rather than flop-bound, NVLink72 can deliver extremely high tokens per watt while reporting very low MFU. Huang argues the right framing is performance, and ultimately tokens per watt as a rough proxy for intelligence per watt, adjusted for the fact that not all tokens are equal. A coding token is worth more than a generic token.

    Hopper, Grace Blackwell NVLink72, Vera Rubin, Feynman

    Huang gives the clearest public framing of NVIDIA’s roadmap as a sequence of architectural answers to evolving compute patterns. Hopper was built for pre-training, at a moment when NVIDIA chose to build multi-billion-dollar machines while the largest scientific supercomputer in the world cost $350 million and the marketplace for such systems was, on paper, zero. Grace Blackwell NVLink72 was the answer to inference and reasoning: a rack-scale computer that ganged 72 GPUs together because decode needs aggregate memory bandwidth far beyond a single chip. The generation-over-generation speed-up was 50x in two years, twenty-five times what Moore’s Law would have delivered. Vera Rubin is being built explicitly for agents. Agents load long-term memory from storage that has to be wired directly into the GPU fabric, they use working memory, they call tools that run on a CPU, and they wait. So the CPU has to be Vera, optimized for low-latency single-threaded code, because the multi-billion-dollar GPU system cannot afford to idle waiting on a slow tool call. Feynman extends the pattern to swarms of agents with sub-agents and sub-sub-agents, a recursive software topology that will demand its own compute pattern.

    Energy demand and the grid

    Huang’s energy projection is one of the most aggressive numbers in the talk. NVIDIA can compound tokens per watt by 50x per generation through codesign, but the total compute demand is heading roughly a thousand times higher, and Huang says he would not be surprised if the real figure is one or two orders of magnitude beyond that. The reason is structural: future computing is generative and continuous, not pre-recorded and on-demand. The good news, he argues, is that this is the best moment in the history of humanity to invest in sustainable generation. Market forces alone are now sufficient to fund solar, nuclear and grid upgrades. Government subsidies are no longer required to make the math work.

    Adversarial countries, export controls and the telecom warning

    This is the segment where Huang is visibly fired up. He attacks the GPUs-as-atomic-bombs framing on its face. NVIDIA GPUs power medical imaging, video games and soy sauce delivery. A billion people use them. He advocates them to his family. The analogy collapses at the first comparison. He attacks the second framing, that American companies should not compete abroad because they will lose anyway, as a self-fulfilling defeat. Competition makes the company better. The third framing, that depriving the rest of the world of general-purpose computing benefits the United States, also fails on first principles: it benefits one or two American companies at the cost of an entire industry. The cautionary parallel is telecommunications. The United States once had a leading position in telecom fundamental technology and policied itself out of it. Huang’s worry, voiced explicitly to a room of CS students, is that they will graduate into a shell of a computer industry if the same path is repeated.

    AI doom and rational optimism

    In the same arc Huang rejects the science-fiction framing of AI as a singularity that arrives suddenly on a Wednesday at 7pm and ends civilization. He calls those claims irresponsible, says they are not true, and points out that the people advancing them are believed by audiences who then make policy on that basis. It is not true that no one understands how these systems work. It is not true that intelligence becomes infinitely powerful instantaneously. It is not true that there is no defense. His framing, which the host echoes as “rational optimism,” is that the goal is to create a future where people care about computers because the technology students are learning is worth mastering.

    Stanford’s compute problem is Stanford’s fault

    A student presses on the scarcity of compute for independent researchers, startups and universities inside the United States. Huang’s answer is sharp: there is no shortage. Place the order and the chips will arrive. The actual broken thing is institutional. University grants are fragmented across departments. No researcher can raise enough on a single grant to fund a billion-dollar shared cluster, and no one shares. He compares it to showing up at the grocery store demanding a billion dollars of tomatoes today. The solution is planning, aggregation and a campus-scale supercomputer, the way Stanford once built the linear accelerator. The endowment is $40 billion. Pulling a billion off it, contracting cloud capacity and giving every student and researcher AI supercomputer access is, in Huang’s view, obviously doable. When he says “it is Stanford’s fault” the host laughs, but Huang clarifies: if it is your fault you have the power to fix it.

    Career, suffering and resilience

    Asked how a CS student should spend the next few years, Huang pushes back on the standard “follow your passion” advice. Most people do not know what they love yet, because no one knows what they do not know. The bar of demanding joy from every working day is too high. Whatever the job is, do it as well as you can. Even as CEO of NVIDIA he says he genuinely loves about 10 percent of his work. The other 90 percent is hard and he suffers through it. He recommends suffering on purpose, because resilience is a muscle that only builds under load, and when the company, the team or the family needs that muscle, it has to already exist. Earlier in his life that meant cleaning toilets and busing tables at Denny’s. He does it today running a multi-trillion-dollar company.

    The biggest mistakes

    Huang separates technical mistakes from strategic mistakes. NVIDIA’s first generation of products was technically wrong in almost every way: curved surfaces instead of triangles, no Z-buffer, forward instead of inverse texture mapping, no floating point inside. The company wasted two and a half years. But the strategic genius of the recovery, the reading of the market, the conservation of resources and the reapplication of talent, is what taught him strategy. The clean strategic mistake he names is mobile. NVIDIA’s Tegra line grew to a billion dollars of revenue and then collapsed to zero when Qualcomm’s modem dominance locked NVIDIA out of the 3G to 4G transition. Huang explicitly refuses the comforting rationalization that the Tegra effort fed the Thor automotive chip (“Thor is the great great great grandson”). The original decision, he says, was a waste of time. The lesson is to think one or two clicks further about whether a market is structurally winnable before committing the company.

    Forecasting under fog of war

    The final substantive exchange is on forecasting. Huang’s method has four steps. Observe what is actually happening (AlexNet crushing two decades of computer vision research in one shot, GPT producing reasoning by token generation). Reason from first principles about why it works. Ask “so what” and “what next” recursively until a mental model of the future emerges. Place the company inside that future and work backwards. Crucially, expect to be partly wrong. Some outcomes will absolutely happen, some will likely happen, some might happen, and the strategy has to be robust across that distribution. The real cost of any strategic choice is the opportunity cost of the alternatives you did not take, so the discipline is to minimize that cost and maximize optionality while letting the journey itself pay for the journey.

    Thoughts

    The most useful thing in this conversation is the explicit architectural mapping of compute patterns to chip generations. Hopper for pre-training. Grace Blackwell NVLink72 for inference, because decode is bandwidth-bound and a single chip cannot supply it. Vera Rubin for agents, because tool calls stall multi-billion-dollar GPU systems and so the CPU has to be optimized for low-latency single-threaded code. Feynman for swarms. That sequence is not marketing. It is a falsifiable thesis about where the bottleneck moves next, and every other infrastructure company should be measuring themselves against it. If Huang is right that swarms of sub-agents are the next dominant pattern, then the design pressure shifts from raw flops to fabric topology, memory hierarchy and storage-to-GPU latency. That has implications for everyone downstream, including the hyperscalers building competing accelerators.

    The MFU section is the most intellectually generous moment in the talk. The instinct in the AI ops community has been to chase MFU as if it were a virtue. Huang argues, persuasively, that low MFU is consistent with high tokens per watt in a disaggregated inference setup, and that bottlenecks rotate fast enough that over-provisioning every resource is the rational design. That reframing matters because it changes what “scarce” means. Compute is not scarce in the way the discourse treats it. What is scarce is a coherent system designed end-to-end. The xAI 11 percent number, in that frame, is not embarrassing. It is the natural reading of a workload that is mostly decode.

    The Stanford segment is the part most likely to be quoted out of context. “It’s Stanford’s fault” is a deliberately provocative line, but the underlying claim is correct and load-bearing. Compute is not gated by NVIDIA refusing to ship chips. It is gated by the fact that fragmented grant funding cannot aggregate into the billion-dollar order that NVIDIA can fulfill. The implication is that universities and national labs need a structural change in how they pool capital for compute, and that the current model of every researcher buying a handful of cards is genuinely obsolete. Huang’s nudge about pulling a billion off the endowment is concrete enough to be acted on, and other major research universities should read this segment as a direct prompt.

    The geopolitical segment is the highest-stakes one. The telecommunications comparison is correct as a historical pattern, and Huang is one of the very few executives in a position to deliver that warning credibly. The unresolved tension is that the argument applies symmetrically. If American AI dominance is built by selling globally, that includes selling into adversarial states, and the policy question is where the line falls. Huang does not answer that question. He attacks the framing that lets the question be answered badly. That is a meaningful contribution to the discourse even if it does not resolve the underlying tradeoff.

    The career advice section is the part the social-media clips will mishandle. “Seek suffering” reads as macho when extracted. In context it is a specific operational claim about how resilience compounds, and it is paired with the Tegra story where Huang himself paid the price of not thinking one more click ahead. That kind of self-implication is rare in CEO talks, and it is the reason the talk is worth listening to in full rather than only reading the recap.

    Watch the full Stanford CS153 Frontier Systems conversation with Jensen Huang here.

  • Krishna Rao on Anthropic Going From 9 Billion to 30 Billion ARR in One Quarter and the Compute Strategy Powering Claude

    Krishna Rao, Chief Financial Officer of Anthropic, sat down with Patrick O’Shaughnessy on Invest Like the Best for one of the most detailed public looks yet at the operating engine behind Claude. He covers how Anthropic compounded from $9 billion of run rate revenue at the start of the year to north of $30 billion by the end of Q1, why he spends 30 to 40 percent of his time on compute, the playbook for buying gigawatts of AI infrastructure across Trainium, TPU, and GPU platforms, how Anthropic prices its models, why returns to frontier intelligence keep climbing, and what the Mythos release tells us about the cyber capabilities of the next generation of Claude.

    TLDW

    Anthropic is running the most compute fungible frontier lab in the world, with active deployments across AWS Trainium, Google TPU, and Nvidia GPU, and an internal orchestration layer that lets a chip serve inference in the morning and run reinforcement learning the same evening. Krishna Rao explains the cone of uncertainty that governs gigawatt scale compute procurement, the floor Anthropic refuses to drop below on model development compute, the Jevons paradox unlock from cutting Opus pricing, the 500 percent annualized net dollar retention from enterprise customers, the layer cake of long term deals with Google, Broadcom, Amazon, and the recent xAI Colossus tie up in Memphis, the phased release of the Mythos model in response to spiking cyber capabilities, the internal use of Claude Code to produce statutory financial statements and run a Monthly Financial Review skill, and why the team believes scaling laws are alive and well. The interview also covers fundraising history through Series D and Series E, the $75 billion already raised plus another $50 billion coming, talent density beating talent mass during the Meta poaching wave, and Rao’s belief that biotech and drug discovery represent the most exciting frontier for AI.

    Key Takeaways

    • Anthropic entered the year with about $9 billion of run rate revenue and ended the first quarter with north of $30 billion of run rate revenue, a more than 3x leap driven by model intelligence gains and the products built around them.
    • Compute is described as the lifeblood of the company, the canvas everything else is built on, and the most consequential class of decisions Rao makes. Buy too much and you go bankrupt. Buy too little and you cannot serve customers or stay at the frontier.
    • Rao spends 30 to 40 percent of his time on compute, even today, and the leadership team meets repeatedly on both procurement and ongoing compute allocation.
    • Anthropic is the only frontier language lab actively using all three major chip platforms in production: AWS Trainium, Google TPU, and Nvidia GPU. It is also the only major model available on all three clouds.
    • Flexibility is the central design principle. Anthropic builds flexibility into the deals themselves, into the orchestration layer that maps workloads to chips, and into compilers built from the chip level up.
    • The cone of uncertainty frames procurement. Small differences in weekly or monthly growth compound into wildly different two year outcomes, so the team plans across a range of scenarios rather than a single point estimate, and ranges toward the upper end while protecting downside.
    • Compute allocation across the company sits in three buckets: model development and research, internal employee acceleration, and external customer serving. A non negotiable floor protects model development even when customer demand is tight.
    • Anthropic estimates that if it cut off internal employee use of its own models, the freed compute could serve billions of dollars of additional revenue. It chooses not to, because internal use compounds into better future models.
    • Intelligence is multi dimensional, not a single IQ score. Anthropic measures real world capability through customer feedback, long horizon task performance, tool use, computer use, and speed at agentic tasks, not just leaderboard benchmarks that have largely saturated.
    • Each Opus generation, 4 to 4.5 to 4.6 to 4.7, delivers both capability improvements and an efficiency multiplier on token processing. New models often serve customers at a fraction of the prior cost while doing more.
    • Reinforcement learning is described as inference inside a sandbox with a reward function, so model efficiency gains directly improve internal RL throughput. The flywheel is tightly coupled.
    • Over 90 percent of code at Anthropic is now written by Claude Code, and a large share of Claude Code itself is written by Claude Code.
    • Anthropic shipped roughly 30 distinct product and feature releases in January and the pace has accelerated since.
    • Scaling laws, in Anthropic’s internal data, are alive and well. The team holds itself to a skeptical scientific standard and still does not see them slowing down.
    • Anthropic recently signed a 5 gigawatt deal with Google and Broadcom for TPUs starting in 2027, plus an Amazon Trainium agreement for up to 5 gigawatts, totaling more than $100 billion in commitments. A significant portion lands this year and next year.
    • A new partnership for capacity at the xAI Colossus facility in Memphis was announced just before the interview, aimed at expanding consumer and prosumer capacity.
    • Pricing has been remarkably stable across Haiku, Sonnet, and Opus. The biggest deliberate change was lowering Opus pricing, which produced a textbook Jevons paradox: consumption rose far faster than the price drop, and the new Opus 4.6 and 4.7 slot in at the same price point.
    • Mythos is the first model Anthropic chose to release in a phased way because of a sharp spike in cyber capability. In an open source codebase where a prior model found 22 security vulnerabilities, Mythos found roughly 250.
    • The Mythos release framework focuses on defensive use first, expands access over time, and is presented as a template for future capability spikes.
    • Anthropic now sells to 9 of the Fortune 10 and reports net dollar retention above 500 percent on an annualized basis. These are not pilots. Rao describes signing two double digit million dollar commitments during a 20 minute Uber ride to the studio.
    • The platform strategy is mostly horizontal. Anthropic will go vertical with offerings like Claude for Financial Services, Claude for Life Sciences, and Claude Security where it can demonstrate the model’s capabilities, but expects most application value to accrue to customers building on top.
    • Investors raised over $75 billion in equity since Rao joined, with another $50 billion in commitments tied to the Amazon and Google deals. Capital intensity is real, but the raises fund the upper end of the cone of uncertainty more than they fund current losses.
    • The Series E close coincided with the day the DeepSeek news broke, forcing investors to reassess their AI thesis in real time. Anthropic closed the round anyway.
    • Inside finance, Claude now produces statutory financial statements for every Anthropic legal entity, with a human checker. A library of more than 70 finance specific skills underpins workflows.
    • A custom Monthly Financial Review skill produces a 90 to 95 percent ready monthly close report, so leadership discussion shifts from reconciling numbers to debating implications.
    • An internal real time analytics platform called Anthrop Stats compresses weekly insight cycles from hours to about 30 minutes.
    • The biggest token user inside Anthropic’s finance team is the head of tax, focused on tax policy engines and workflow automation. The most senior people, not the youngest, are leading internal adoption.
    • Talent density beats talent mass. When Meta and others ran aggressive offer waves, Anthropic lost two people while peer labs lost dozens.
    • All seven Anthropic co founders remain at the company, as does most of the first 20 to 30 employees, which Rao credits to a collaborative, transparent, debate friendly culture and a real culture interview that can veto otherwise top tier candidates.
    • Dario Amodei holds an open all hands every two weeks, writes a short prepared document, and takes unscripted questions from anyone at the company.
    • AI safety investments in interpretability and alignment have a commercial side effect. Looking inside the model helps Anthropic build better models, and enterprises selling sensitive workloads want to trust the lab they hand customer data to.
    • Anthropic explicitly identifies as America first in its approach to model development, and engages closely with the US administration on capability releases such as Mythos.
    • The longer term product vision is the virtual collaborator: an agent with organizational context, access to the company’s tools, persistent memory, and the ability to work on ideas, not just tasks, over long horizons.
    • CoWork, Anthropic’s extension of the Claude Code paradigm into general knowledge work, is being adopted faster than Claude Code itself when indexed to the same point in its launch curve.
    • Anthropic’s product teams ship daily, with a fleet of agents working across the company on specific tasks. Everyone effectively becomes a manager of agents.
    • The dominant downside risks to Anthropic’s high end forecast are slower customer diffusion of model capability into real workflows, scaling laws flattening unexpectedly, and Anthropic losing its position at the frontier.
    • Rao is most excited about biotech and healthcare outcomes, especially the prospect that AI could push drug discovery and lab throughput up 10x or 100x, turning currently incurable diagnoses into treatable ones within a patient’s lifetime.

    Detailed Summary

    Compute as Lifeblood and the Cone of Uncertainty

    Rao opens with the claim that compute is the most important resource at Anthropic, and the most consequential decision class in the company. You cannot buy a gigawatt of compute next week. You have to anticipate demand a year or two in advance, and the cost of being wrong in either direction is high. Buy too much and the unit economics collapse. Buy too little and you cannot serve customers or stay at the frontier, which are described as the same failure mode. To navigate this, the team uses a cone of uncertainty rather than point estimates. Small differences in weekly growth compound into vastly different two year outcomes, and Anthropic tries to position itself toward the upper end of that cone while preserving optionality. Rao notes he has had to consciously break a lifetime of linear thinking and force himself into exponential models.

    Three Chip Platforms, One Orchestration Layer

    Anthropic uses Amazon’s Trainium, Google’s TPUs, and Nvidia’s GPUs fungibly. That was not free. Adopting TPUs at scale started around the third TPU generation, when outside observers thought it was a strange choice. Anthropic invested years into compilers and orchestration so workloads can flow across chips by generation and by job type. The team works deeply with Annapurna Labs at AWS to influence Trainium roadmaps because Anthropic stresses these chips harder than almost anyone. The result is what Rao believes is the most efficient utilization of compute across any frontier lab, with a dollar of compute going further inside Anthropic than anywhere else.

    Three Buckets and the Model Development Floor

    Compute gets allocated across model development, internal acceleration of employees, and customer serving. The conversations are collaborative rather than zero sum, but there is a hard floor on model development that the company refuses to cross even if it makes customer demand harder to serve in the short term. The thesis is simple. The returns to frontier intelligence are extremely high, especially in enterprise, so cutting model investment to chase near term revenue is a bad trade. Internal employee use is also explicitly protected. Rao notes that diverting that internal usage to external customers would unlock billions of additional revenue today, but the compounding benefit of accelerating researchers and engineers outweighs that.

    Intelligence Is Multi Dimensional

    Rao pushes back hard on the IQ framing of model progress. Benchmarks saturate quickly, and the real signal comes from how customers actually use the models. Anthropic looks at long horizon task completion, tool use, computer use, and time to result on agentic tasks. Two equally capable agents who differ only in speed produce dramatically different value, because the faster one compounds into more attempts and more outcomes. Frontier model leaps are also fuel efficient. The sedan to sports car analogy breaks down because each Opus generation, 4 to 4.5 to 4.6 to 4.7, delivers a step up in capability and a multiplier on per token efficiency.

    From 9 Billion to 30 Billion ARR in One Quarter

    The headline number for the quarter is a leap from about $9 billion of run rate revenue to over $30 billion, accomplished without onboarding a corresponding step up in compute, because new compute lands on ramps locked in 12 months prior. Rao attributes the leap to model capability gains, products that surface that intelligence in usable form factors, and an enterprise customer base that pulls more workloads onto Claude as each generation unlocks new use cases. Coding started the wave with Sonnet 3.5 and 3.6, and the same pattern is now playing out elsewhere in the economy.

    Recursive Self Improvement and Talent Density

    Over 90 percent of Anthropic’s code is now written by Claude Code, including most of Claude Code itself. Rao describes this as a structural reason to keep allocating internal compute to employees even when external demand is hungry. Recursive self improvement is not happening through models that need no humans. It is happening through researchers who set direction and use frontier models to compress months of work into days. Talent density beats talent mass. When Meta and other labs went after Anthropic researchers with very large packages, Anthropic lost two people while peer labs lost dozens.

    Procurement Strategy and the Layer Cake

    Compute lands as a layer cake. Last month Anthropic signed a 5 gigawatt TPU deal with Google and Broadcom starting in 2027, alongside an Amazon Trainium agreement for up to 5 gigawatts. The total is north of $100 billion in commitments. A new tie up with xAI’s Colossus facility in Memphis was announced just before the interview, intended for nearer term capacity to support consumer and prosumer growth. Anthropic evaluates near term and long term compute deals against the same set of variables: price, duration, location, chip type, and how efficiently the team can run it. The relationships are deeper than procurement. The hyperscalers are also distribution channels for the model.

    Platform First, Selective Vertical Bets

    Rao describes Anthropic as a platform first business, with most expected value accruing to customers building on the platform. The team will only go vertical when it can either demonstrate capabilities that are skating to where the puck is going, like Claude Code did before the models could fully support it, or when it wants to set a template for an industry vertical, as with Claude for Financial Services, Claude for Life Sciences, and Claude Security. He acknowledges that surprise capability jumps make customers anxious about the platform competing with them, and frames Anthropic’s mitigation as deeper partnerships, early access programs, and an emphasis on accelerating customer building rather than disintermediating it.

    Pricing, Jevons Paradox, and Return on Compute

    Pricing across Haiku, Sonnet, and Opus has been stable. The notable exception is Opus, which Anthropic deliberately repriced lower when launching Opus 4.5 because Opus class problems were being squeezed into Sonnet workloads. Efficiency gains made it possible to serve Opus profitably at the new level. The consumption response was a classic Jevons paradox, with usage rising far more than the price reduction would have predicted, and Opus 4.6 then slotted in at the same price with a capability bump. Margins are not framed as a per token markup. Compute is fungible across model development, internal acceleration, and customer serving, so Anthropic measures return on the entire compute envelope rather than software style variable cost per call.

    Fundraising, DeepSeek, and Capital Intensity

    Rao joined while Anthropic was closing its Series D, mid frontier model launch and during the FTX share liquidation. Investors initially questioned whether Anthropic needed a frontier model, whether AI safety and a real business could coexist, and why the sales team was so small. The Series E closed the same day the DeepSeek news broke, with markets violently re pricing AI in real time. Since Rao joined, Anthropic has raised over $75 billion, with another $50 billion tied to the Amazon and Google compute deals. The reason for the size of the raises is the cone of uncertainty, not current losses. Returns on compute today are described as robust.

    Mythos, Cyber Capability, and Phased Releases

    The Mythos release marks the first time Anthropic shipped a model under a deliberately phased rollout because of a specific capability spike. Cyber is the dimension that spiked. Where a prior model found 22 vulnerabilities in an open source codebase, Mythos found roughly 250. The defensive applications, automatically patching massive codebases, are genuinely valuable, but the offensive risk is real enough that Anthropic chose to release to a smaller group first and expand access over time. Rao positions this as a template for future capability spikes, not a permanent restriction. He also describes the relationship with the US administration as cooperative, including the Department of War interaction, with Anthropic supporting a regulatory framework that does not strangle innovation but takes responsibility seriously.

    Claude Inside Finance

    Anthropic’s finance team is one of the strongest internal case studies. Statutory financial statements for every legal entity are produced by Claude, with a human reviewer. A skill library of more than 70 finance specific skills underpins a Monthly Financial Review skill that drafts the monthly close at 90 to 95 percent ready, so leadership meetings shift from explaining the numbers to discussing what to do about them. An internal analytics platform called Anthrop Stats compresses weekly insight cycles from hours to 30 minutes. The biggest internal token user in finance is the head of tax, building policy engines, which Rao highlights as evidence that adoption is driven by the most senior people, not just younger engineers.

    Culture, Co Founders, and the Race to the Top

    Seven co founders should not, on paper, work as a leadership group. Rao argues it works because the culture was set early around collaboration, intellectual honesty, transparency, and humility. The culture interview is a real veto, not a checkbox. Dario Amodei runs an all hands every two weeks with a short written piece followed by unscripted questions, and decisions, once made, get clean alignment rather than residual politics. Anthropic frames its approach as a race to the top, where being a model for how to build the technology responsibly is itself a recruiting and retention advantage.

    The Virtual Collaborator and the Frontier Ahead

    The product vision Rao describes is the virtual collaborator. Not just a smarter chatbot, but an agent with organizational context, access to the company’s tools, memory, and the ability to work on ideas over long horizons. Coding was the first domain to feel this, but CoWork, Anthropic’s extension of the Claude Code pattern into general knowledge work, is being adopted faster than Claude Code was at the same age. Product development inside Anthropic already looks different. Teams ship daily, with fleets of agents working across the company, and individual humans increasingly act as managers of those fleets.

    Downside Risks and What Excites Him Most

    The three risks Rao names if asked to do a premortem on a softer year are slower customer diffusion of model capability into real workflows, scaling laws unexpectedly flattening, and Anthropic losing its frontier position to competitors. None of these are observed today, but he is unwilling to claim them with certainty. On the upside, he is most excited about biotech and healthcare. Lab throughput rising 10x or 100x, paired with AI assisted clinical workflows, could turn currently incurable diagnoses into treatable ones within a patient’s lifetime. That is the outcome he wants the technology to chase.

    Thoughts

    The most consequential structural point in this interview is the framing of compute as a single fungible resource pool measured by return on the entire envelope, not as a variable cost per inference call. That accounting shift, if you accept it, breaks most of the bear cases about AI lab unit economics. The bear argument almost always assumes that a token served to a customer is the only thing the chip did that day. Rao’s version is that the same fleet trains models in the morning, runs reinforcement learning at lunch, serves customers in the afternoon, and accelerates internal engineers in the evening. If even half of that is real, the right comparison is total compute spend versus total enterprise value created by the platform, and on that ratio Anthropic looks structurally strong rather than weak.

    The Jevons paradox on Opus pricing is the most actionable insight for anyone running an AI product. Most teams default to either chasing premium pricing on the newest model or undercutting to chase volume. Anthropic did something more disciplined: it left Sonnet and Haiku alone, dropped Opus when efficiency gains made it serveable, and watched aggregate usage rise faster than the price cut. The lesson is that frontier model pricing is not really a price problem. It is a capability access problem, and elasticity around the right tier is much higher than the standard SaaS playbook implies.

    The Mythos cyber jump deserves more attention than it has gotten. Going from 22 to 250 vulnerabilities found in the same codebase is the kind of capability discontinuity that genuinely changes the regulatory calculus. Anthropic is signaling that it can identify these discontinuities ahead of release and choose a deployment shape that respects them. Whether peer labs adopt similar discipline is the open question. Anthropic’s race to the top framing assumes they will be forced to. The competitive market may say otherwise.

    The hiring data point is the most underrated investor signal. Two departures while peer labs lost dozens, during the most aggressive talent war in tech history, is not a culture poster. It is a structural advantage that compounds every time another lab tries to buy its way to the frontier. Money can be matched. Conviction in the mission, transparent leadership, and a culture interview that can veto otherwise stellar candidates cannot. If you believe scaling laws hold, talent retention at this density is one of the few moats that actually scales with capital.

    Finally, the most interesting personal admission is that Krishna Rao, a finance leader trained at Blackstone and Cedar, is openly telling investors that linear thinking is the failure mode he had to break out of. The companies that pattern match this moment to prior technology waves are mispricing it, in both directions. The cone of uncertainty Anthropic uses internally is the right metaphor for everyone else too. If you are forecasting AI as if it is cloud in 2010, you are almost certainly wrong, and the magnitude of the error is much larger than it would be in any prior era.

    Watch the full conversation with Krishna Rao on Invest Like the Best here.