PJFP.com

Pursuit of Joy, Fulfillment, and Purpose

Tag: Jevons paradox

  • Benedict Evans on Why AI Is Stuck in 1997: The Task vs the Job, Commodity Models, and Why the Jobs Apocalypse Is Overhyped

    Benedict Evans, the former Andreessen Horowitz partner and independent analyst behind the annual “AI Eating the World” presentation, sat down with Lenny’s Podcast for what the host calls the most rational take on AI you will hear this year. Instead of either doom or hype, Evans argues that AI is as big a deal as the internet or mobile, and only as big a deal as the internet or mobile, which means we are living through something closer to 1997 than to the singularity. The conversation moves through the jobs question, the difference between a task and a job, whether the model labs have any pricing power, the anti-AI backlash, and what people should actually do. You can watch the full conversation on YouTube here.

    TLDW

    Evans frames AI as a platform shift on the scale of the internet or mobile, with the crucial twist that almost nothing has been built yet, so we are in the 1997 moment where confident predictions about winners are usually wrong. He introduces his central tool, the distinction between the task and the job, to explain why “X percent of this profession is exposed to AI” studies are misleading, why the AI labs are paradoxically hiring forward deployed engineers and buying consultancies, and why accountants kept multiplying through every wave of automation (the lump of labour fallacy and Jevons paradox at work). On value capture he makes a deterministic bet that foundation models have no network effects, behave like a commodity, and will look more like cloud than like Windows, with the value moving up the stack to applications, much as it did in telecom, where a trillion-dollar industry grew data traffic thousands of times over while its stocks went nowhere. He covers distribution as the real moat, Apple Intelligence as the most compelling unshipped vision, the fuzzy anti-AI backlash (including the largely fake water panic and the very real harms of deepfakes), raising kids under radical uncertainty, and closes with the disarming admission that his own synthesis-heavy job is exactly the kind AI is currently worst at. His advice: presume radical uncertainty, dive in rather than sneer, and assume it will probably be okay.

    Thoughts

    The most useful thing in this conversation is a single question Evans keeps returning to: what is the task, and what is the job? A spreadsheet automated the arithmetic an accountant does, and the number of accountants went up for the next forty years. Claude Code can write the code, but deciding what to build, for whom, and why is the part nobody has automated. The reason the “this profession is X percent exposed to AI” studies feel hollow is that they assume a job is a neat stack of separable tasks. Evans argues, by analogy to the old expert-systems failure, that you simply cannot decompose a senior lawyer’s work that way. The 75-slide deck is the task. Walking your company, reading its politics, talking to your customers, and telling you the uncomfortable truth is the job, and that is what you actually paid McKinsey for.

    The boldest and most falsifiable claim is that the foundation-model companies look more like cloud than like Windows. No network effects means no winner-take-all, which means durable competition, which means commodity pricing and compressed margins, with the real value accruing up the stack in applications that nobody at the labs is going to build. His telecom analogy is the one to sit with. A trillion-dollar industry grew mobile data traffic by 1,500 to 2,000 times in fifteen years, and the stocks went nowhere for a quarter century, because it was a low-margin utility while all the interesting value moved to Apple and the people building apps on top. If he is right, the current token-burn economics, the person reportedly spending 1.5 million dollars a month on tokens, are the 2010 equivalent of a 50,000 dollar roaming bill, not the steady state. Evans flags openly that he could be completely wrong, which is the intellectually honest part and the part most forecasters skip.

    “It depends” and “it will probably be okay” sound like evasions, and Evans leans into that. But the 1997 framing is doing real work. The point is not that AI is small, it is that the things that will end up mattering have not been built, and that anyone confidently naming the winners today is repeating the 1997 mistake of betting on Excite over a search company with a weird logo. The discipline he is selling is to presume radical uncertainty and act anyway, because the alternative, declaring the whole thing slop and shouting about it online, buys a great feeling of moral superiority and nothing else. His repeated insistence that you can see the job that goes away but never the new job, because it does not exist yet, is the load-bearing idea under his optimism.

    The most disarming moment is the closing AI-corner answer, where the person whose entire brand is explaining AI admits he struggles to use it. His work is synthesis and precise information retrieval, and precise retrieval happens to be exactly what today’s models are worst at. He is, in his own words, the lawyer looking at VisiCalc: it is obviously transformative, and he just does not happen to make spreadsheets all day. That admission is worth more than any benchmark, because it locates the real variable. How much AI changes your life depends less on how good the model gets and more on whether your daily work sits on the part of the jagged frontier where it already works. That is a far more practical lens than arguing about whether AGI arrives in three years or thirty.

    Key Takeaways

    • Evans’s headline opinion is that AI is as big a deal as the internet or mobile, and only as big a deal as the internet or mobile. Both halves of that sentence matter.
    • If you make the internet comparison honestly, we are roughly in 1997: very exciting, most of it does not work yet, most of what people will build has not been built, and it is unclear how any of it will end up working.
    • Adoption is spread across a very wide distribution. Even among teenagers, only something like 15 to 20 percent are daily active users and another 20 percent weekly, with the majority saying they do not use it at all.
    • That spread maps onto the “jagged frontier” question of where AI works, where it does not, whether you can predict where it will work in advance, and whether you can even tell after the fact.
    • Software developers are the accountants seeing VisiCalc: for them everything has already changed. Most other professions are watching, intrigued but unsure what to do with it.
    • The AI labs are investing heavily in forward deployed engineers, consultancies, and professional services. Evans jokes that a forward deployed engineer is an Accenture outsourced developer who lives in San Francisco.
    • Companies do not have spare people sitting around to reimagine every internal workflow, so reinventing a business around AI is itself a project that needs consultants, which is why the most cutting-edge labs are funding exactly the firms everyone assumed AI would kill.
    • The central framework: separate the task from the job. Sometimes the task is the job (the elevator operator pressing a lever), and automating the task ends the job. Far more often, the task is only part of the job.
    • Amazon gets you the SKU once you know which SKU you want. Knowing which one to buy is a different job. Claude Code writes the code, but knowing what code and what features to build is the job.
    • A McKinsey or Bain engagement is not really about the deck. The deck is the task. The job is walking your enterprise, understanding the politics, talking to your customers, and telling you the truth.
    • The Jevons paradox is just price elasticity applied to labour. Make something cheaper to produce and you usually do far more of it, not the same amount with fewer people.
    • Excel did not give investment bankers shorter hours. iPhone SDKs did not shrink the number of engineers even though Apple writes 90 percent of the code for you. The number of accountants rose through every wave of automation.
    • The lump of labour fallacy: since 1800, each technology automates jobs and unlocks new ones. You can always see the job that disappears and never the new job, because it does not exist yet.
    • Evans is wary of argument from authority on jobs. He wants Dario Amodei’s view on where models go in the next 6 to 12 months, not necessarily his theory of labour markets and comparative advantage.
    • The doomer scenario of every company buying ChatGPT and firing everyone in two weeks misunderstands how enterprises work. Enterprise sales cycles run 18 months or more. Nobody is ripping out SAP overnight. The full transformation takes 3 to 10 years, sector by sector.
    • AGI and superintelligence are being quietly redefined to mean whatever works now. Larry Tesler’s theorem: AI is whatever machines cannot do yet, because once they can, people call it just software.
    • We have no theory of human intelligence, no theory of why these models work, and no theory of how much better they will get, so everyone is vibes-forecasting. Even if progress stopped tomorrow, what exists is already transformative and will roll out for a decade.
    • On value capture, Evans argues models show no network effects, so no single one runs away with the market. Persistent competition plus little real product differentiation means little pricing power.
    • Sam Altman’s pitch of selling intelligence on a meter like electricity ignores the brutal margin structure of utilities. Your TV maker does not pay the power company a cut of your bill.
    • The telecom analogy: a roughly trillion-dollar mobile industry spends 15 to 20 percent of revenue on capex, grew data consumption 1,500 to 2,000 times since 2010, and its stocks went nowhere for 25 years because it is a low-margin commodity utility.
    • The elemental question: does the model do the whole thing, or does it need thousands of different apps built by different people? If it needs apps, the labs cannot build them all, just as Microsoft did not, so it looks more like AWS than like Windows.
    • If the product is a commodity, distribution becomes the moat. Google pushes Gemini through its surfaces, Meta sprayed AI across its apps and quietly ranked between ChatGPT and Gemini in usage, and incumbents with distribution have a structural edge.
    • Browsers are the warning: Microsoft used distribution to win the browser war, then it turned out winning browsers did not matter because the value was further up the stack.
    • Apple Intelligence, as shown at WWDC 2024, was the most compelling vision of a personal AI assistant Evans has seen. Apple could not ship it, but neither could anyone else, because tool-using on-device agents with no hallucinations across thousands of apps is genuinely hard.
    • The model is “the dumb thing underneath” that powers a feature. The same commodity model can sit beneath both Gemini on Android and Apple Intelligence on iOS while the products and distribution differ entirely.
    • The anti-AI backlash is a big fuzzy mess. Some is real (local electricity bills, deepfakes, real job anxiety), some is sort of true, and some is simply false.
    • The data-center water panic is largely fake. A Livermore lab study put US data-center water consumption at about 0.017 percent of US water use. Local well conflicts are planning problems, not data-center problems.
    • We have shockingly little hard data. The model labs do not publish meaningful usage numbers. There is no public daily active user figure for ChatGPT, so economists are reverse-engineering effects from government surveys.
    • Real new harms do appear with each wave. A teenager could not use Photoshop to make explicit fakes of every classmate and send them to the whole school in an afternoon. Now they can, and turn them into video.
    • The UK Post Office Horizon scandal (buggy Fujitsu software wrongly showing cash shortfalls, leading to prosecutions, bankruptcies, and suicides) is a reminder that every technology brings new ways to ruin lives, by malice or by accident.
    • You cannot reliably predict what gets exposed. In 1997 people thought taxis were safe from the internet and newspapers would be fine. The opposite happened. Today, “AI-proof” jobs like personal trainer may not be as safe as they look.
    • Uber and Airbnb show that similar-sounding companies can have very different market impact. Uber demolished and then grew the taxi market, while Airbnb’s effect on hotels was fairly marginal because business travel still wants a hotel.
    • Every new technology first lets you do the old thing but more, then unlocks things that were not possible before. Recorded music revenue is U-shaped: first “what if I do not pay 15 dollars for a CD,” then “what if 15 dollars a month gives me all the music there is.” Spotify is not an online music store, it is something else.
    • Coding was supposed to be one of the last things automated, and instead it is the most transformed role of all, which is itself a lesson in how badly we predict exposure.
    • Practical advice: do not stick your head in the sand. Dive in, submerge yourself, and come out understanding what you can do with it. Going into a shrinking job market announcing you will never use AI is not the right posture.
    • Evans’s honest coda: he struggles to find AI use cases because his job is synthesis and precise retrieval, the things models are worst at. He uses it for proofreading, images, redecorating his apartment, and dictation. He is the lawyer looking at VisiCalc.

    Detailed Summary

    AI is as big as the internet, and we are living in 1997

    Evans opens with the opinion he calls his most controversial: AI is as big a deal as the internet or mobile, and only as big a deal as the internet or mobile. To some in tech that sounds dismissive, as if he is underrating a once-in-history event. His reply is that smartphones and the internet were themselves enormous, and we are talking over the internet right now. The deeper point is the comparison’s timing. If this is like the internet, then it is like the internet in 1997: thrilling, but most of it does not work yet, most of what will be built has not been built, and nobody knows how the pieces will fit. His latest 80-slide presentation, he jokes, is essentially 80 ways of saying “we do not know,” which is partly facetious and partly the entire point.

    The jagged frontier and the wide spread of adoption

    Adoption is not uniform, it is a wide distribution. Some people in tech have bought clusters of Mac minis and stopped using Google, while most people outside tech who use AI at all touch it once every week or two. Even among 13 to 18 year olds, daily active use sits around 15 to 20 percent, weekly use adds another 20 percent, and roughly 60 percent say they do not use it. That spread maps onto what Evans calls the jagged frontier: whether a given task works, whether you can predict in advance that it will work, whether it is intuitive, and whether you can even tell after the fact. Software developers are the accountants who just saw VisiCalc, living in a clear before-and-after. Everyone else is somewhere on the curve, picking it up to varying degrees and a little puzzled about what it is for.

    Why the AI labs are buying consultancies

    One of the most counterintuitive trends is that the leading labs are pouring money into forward deployed engineers and professional services, the very category many assumed AI would erase. Evans’s explanation is grounded in how companies actually operate. Firms do not keep spare people sitting around to redesign stores, hunt down churn, or rebuild a tech stack, which is exactly why they hire Bain, BCG, McKinsey, Accenture, or Infosys when a big project appears. Reimagining every internal workflow around AI, then actually plugging vertical and horizontal systems together and retraining people, is itself a multi-month project requiring people you do not have. So the work gets outsourced, and the most advanced labs are funding the firms that do it. His joke lands the point: a forward deployed engineer is a statistician, or an Accenture developer, who happens to work in San Francisco.

    The task versus the job

    This is the spine of the conversation. Ask what the hard part of a job really is. Sometimes the task is the job: the elevator attendant’s whole job was driving the car, the task got automated, the job ended. Much more often the visible task is only a slice. Amazon gets you the SKU once you know which SKU you want, but knowing what to buy is a separate job. Claude Code writes the code, but deciding what to build, for whom, and how to take it to market is the job. A consulting deck is the task, while the reason you pay Bain is for them to walk your company, understand its politics, talk to your customers, and tell you the truth. Evans notes you can already generate a bad McKinsey deck with AI, and the LinkedIn grifters who do are missing that the deck was never the thing you were buying.

    Jevons paradox and the lump of labour fallacy

    The Jevons paradox is just price elasticity applied to labour: make something cheaper to do and you usually do much more of it. Excel did not hand junior bankers their Friday afternoons off, it expanded the work. iPhone developers write a fraction of the raw code because Apple wrote the drivers and file system, and there are not a tenth as many engineers, there are far more. The count of accountants climbed through adding machines, punch cards, mainframes, databases, ERP, spreadsheets, and cloud. The lump of labour fallacy is the broader version: since 1800 every technology has removed jobs and unlocked new ones, the removed jobs usually look bad in hindsight, the new ones tend to be better, and GDP keeps rising. You can always see the job that disappears and never the one that does not exist yet.

    The jobs question, Dario, and the enterprise sales cycle

    On the coming jobs apocalypse, Evans is cautious about argument from authority. Running an AI lab makes Dario Amodei worth listening to on where models go in the next 6 to 12 months, not necessarily on labour economics and comparative advantage. The doomer image of companies buying ChatGPT and firing everyone within weeks misreads reality: enterprise sales cycles run 18 months or longer, nobody is tearing out SAP overnight, and the full transformation will take 3 to 10 years, sector by sector, as people slowly work out what to do. He points to the lag in software itself. Many SaaS companies founded the day before ChatGPT launched could have been built a decade earlier, and were not, because the delay was someone realizing a problem existed and that this was the way to solve it.

    Redefining AGI and superintelligence

    Evans is skeptical of the moving terminology. He cites Larry Tesler’s line that AI is whatever machines cannot do yet, because the moment they can, people call it just software. Machine learning, image recognition, and sentiment analysis all got reclassified as not really AI once they worked, the same way jet airliners were once high technology and are now just planes. AGI is now often quietly redefined as doing some percentage of economically valuable work, which a 1975 mainframe also did, rather than anything about consciousness or a soul. Whether we reach human-level intelligence is, in his view, genuinely unknowable right now. The reassuring point is that you do not need to resolve it. Even if models hit a brick wall tomorrow, what already exists is transformative and will take a decade to deploy.

    Where the value accrues: commodity models and the telecom analogy

    Here Evans makes his most deterministic argument. Foundation models appear to lack network effects, so no single model runs away from the pack, competition persists, and product differentiation as users experience it is thin. Without differentiation or lock-in, where does pricing power come from? He skewers Sam Altman’s image of selling intelligence on a meter like electricity by pointing out that utilities have terrible margins and nobody pays the power company a cut of their TV. His telecom career supplies the analogy: mobile is a roughly trillion-dollar industry that spends 15 to 20 percent of revenue on capex, grew data traffic 1,500 to 2,000 times since 2010, and whose stocks went nowhere for 25 years because it is a low-margin commodity utility while the value sits up the stack with Apple and the app makers. If models are commodities and the real product is thousands of apps the labs will not build, the outcome looks like cloud, not like Windows.

    Distribution as the moat

    If the product is a commodity, distribution decides the winners. The web browser is the cautionary tale: the browser product is a thin wrapper around a rendering engine, tab browsing was the last real innovation 20-plus years ago, Microsoft used distribution to win, and then winning browsers turned out not to matter because the value was elsewhere. Now Google drives Gemini through its surfaces and Meta sprayed AI across its apps and, in survey data, sat between ChatGPT and Gemini in usage despite tech writing it off. An adequate product with great distribution and brand becomes a big deal, which is why OpenAI spent last year trying everything to build a flywheel before the giants defaulted everyone onto their own offering. The power of the default and sheer inertia do a lot of work.

    Apple Intelligence and the model as the dumb thing underneath

    Evans calls the Apple Intelligence segment of WWDC 2024 the most compelling vision of a personal AI assistant he has seen: tool-using, on-device, agentic, with no prompt injection or hallucinations across a standardized API spanning thousands of apps. Apple could not ship it, but neither could anyone else, because that is genuinely hard. The episode illustrates his framing that the model is “the dumb thing underneath” that powers a feature. The same commodity model can sit beneath Gemini intelligence on Android and Apple Intelligence on iOS, with different products, different distribution, and different decisions about what the feature should be. Apple has a billion edge-capable devices, while Google’s “coming soon to our most powerful devices” really means it will not work on most Android phones.

    The anti-AI backlash, water, and real harms

    The backlash, Evans says, is a big fuzzy mess of very different things. Some is tangible, like a higher local electricity bill in a small number of places. Some is essentially fake, like the water panic. He dug into a Livermore lab study putting US data-center water use at about 0.017 percent of national consumption. Local well conflicts are planning failures, not data-center failures. The jobs piece is genuinely unresolved, with charts pointing both ways and a youth employment slowdown that shows up regardless of degree or AI exposure. He stresses how little hard data exists, since the labs publish no meaningful usage numbers and there is no public daily active user figure for ChatGPT. He compares the moment to the social media backlash, compressed, where some fears were true, some half true, and some simply false. The real new harms are real, though: deepfakes let a teenager generate explicit fakes of an entire school in an afternoon, and the UK Post Office Horizon scandal shows how buggy software plus institutional denial can destroy lives.

    You cannot predict what gets exposed, and what to actually do

    Evans dismisses the O*NET-style exercise of scoring what percentage of each profession AI can do as deluded, the modern version of the expert-systems problem, where you try to describe a job as 700 logical steps and it never works. You cannot say a senior partner’s work is 17 percent automatable. The history of prediction is humbling: in 1997 people thought taxis were safe from the internet and newspapers would simply save on printing, and both were wrong. Coding, supposedly one of the last things to automate, became the most transformed role of all. Personal trainers might be next once your phone can watch your form. His closing advice is to presume radical uncertainty and act anyway: do not retreat into sneering moral superiority, dive in, internalize what the tools can do, and make yourself a great hire. He ends with a candid admission that his own synthesis-and-retrieval job is exactly what AI is currently worst at, so he is the lawyer looking at VisiCalc, sure it changes everything while not personally making spreadsheets all day.

    Notable Quotes

    “My most controversial opinion is that I think that AI is as big a deal as the internet or mobile, and only as big a deal as the internet or mobile.”

    Benedict Evans, stating the thesis that frames the whole conversation

    “If you’re going to make the internet comparison, it’s like we’re in 1997. It’s very exciting. Most stuff kind of doesn’t work yet. Most of the stuff that people are going to do hasn’t been built yet.”

    Benedict Evans, on why confident predictions about AI winners are usually wrong

    “You can’t look at a senior partner at a law firm and say, well, 17 percent of their work could be automated. This is horseshit.”

    Benedict Evans, on why O*NET-style job-exposure scoring fails

    “Claude Code can write you the code, but what code do you want? It can make you the features, sure, but what features do you want? Who’s your customer? What’s the right product for that customer?”

    Benedict Evans, drawing the line between the task and the job

    “There’s this quote from Sam Altman where he said we’re going to be selling AI intelligence on a meter like water or electricity, and you look at this and think, my dear sweet child, you need me to explain the margin structure of the utility industry to you.”

    Benedict Evans, on why model labs may lack pricing power

    “The model is just the dumb thing underneath that powers the feature. The model is the commodity that powers different decisions about what the feature should be.”

    Benedict Evans, on why value moves up the stack to applications

    “Every time we have a new technology it automates away a bunch of jobs, and then that automation unlocks a bunch of new jobs, and you don’t know the new job because it doesn’t exist yet.”

    Benedict Evans, on the lump of labour fallacy and 200 years of automation

    “Don’t stick your head in the sand and say I hate all of this stuff. That gives you a great feeling of moral superiority, but that’s not going to help. What helps is you diving into this and coming out understanding what you can do with it.”

    Benedict Evans, on what to actually do about AI right now

    “AI is good at stuff that computers are bad at, and bad at stuff that computers are good at.”

    Benedict Evans, quoting an observation that explains why he struggles to use AI in his own work

    This is a curated set of pulls, not a transcript. To hear the full argument in context, including the telecom and recorded-music charts and the lightning round, watch the full conversation on YouTube here.

    Related Reading

  • 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.