,

Sourcery at TomorrowX: Gavin Baker on Nvidia, Guillermo Rauch on AI Hacks

Gavin Baker defends Jensen Huang, Guillermo Rauch predicts more AI-driven hacks, and Brian Armstrong pitches agentic finance in ten short TomorrowX interviews from Molly O’Shea’s Sourcery. The two-hour episode, recorded at the Moody Center in Austin and published November 19, 2026, also features Andrew Feldman of Cerebras, Keller Rinaudo Cliffton of Zipline, Matteo Franceschetti of Eight Sleep, Igor Kurganov, Vibhav Altekar of Saronic, iConnections CEO Ron Biscardi and AI educator Harper Carroll. It was taped in a week that brought a nationwide FAA ground stop, a Microsoft 365 outage and Huang’s exit from Nvidia, and most guests were asked about all three.

TLDW

The center of the episode is Nvidia. Gavin Baker, who co-organized TomorrowX, argues that Jensen Huang’s China policy kept America ahead and that the board should keep a return open, while poker champion and game theorist Igor Kurganov calls the chip-export logic “six-dimensional chess” and says Nvidia’s software lock-in is fading. Baker also argues the market is mispricing AI by ignoring open-source and enterprise demand. Rauch and Armstrong explain why AI makes hacking both easier and more lucrative and what defenders should do, and the hardware founders (Zipline, Eight Sleep, Saronic) agree that investors piling into capex underestimate how long hardware takes.

Thoughts

Whether America should sell advanced chips to China is the real argument of this episode, and its two sides were recorded about twenty minutes apart. Baker’s case is that Huang’s policy of selling China chips weaker than America’s kept Chinese labs on Nvidia hardware and starved Huawei of the use it needs to improve, and that the new restrictions will force every Chinese lab onto Huawei and “might get better really fast.” Kurganov answers that the lock-in theory assumes China is dumb enough to stay locked in, that China is limited by chips rather than energy so every marginal chip helps it, and that the software moat is eroding: Anthropic’s IPO materials said its AI did in about three days a piece of AMD bring-up work that used to take six to twelve months. His line, “Jensen said I’m not a car. I think now you’re a car,” is the sharpest of the day. He adds a point that applies to the whole conference: shareholder-value arguments get presented as geopolitics. Baker is open about where he stands: a 27-year friendship with Huang and a letter with Brad Gerstner urging the board to keep a return open. On this tape Kurganov wins the lock-in argument and Baker wins the Huawei one. Restricting chips removes China’s dependence on Nvidia, but it also gives Huawei a captive market.

Baker’s bull case on AI is better than his Nvidia case because it makes a falsifiable claim. He says the market is dividing something like $2 trillion of capex by under $200 billion of lab revenue (roughly $80 billion at Anthropic, about $40 billion at OpenAI, plus everyone else) and concluding the return is not there. His objection is that the denominator ignores open-source inference, where lower margins pull more compute per dollar, and on-premise enterprise AI. He cites Fireworks, an Atreides portfolio company whose top customer grew 13 times this year, and what he says Michael Dell told the conference about “AI factory” sales: 3,200 in 2024 and 2025, and about 3,300 more in the last nine months. The data is mostly his own portfolio’s and Dell’s, so treat it as a lead, not proof. But the test is concrete: if he is right, 2027 guidance from hardware companies should beat what stocks currently imply, which is his “voting machine” to “weighing machine” point.

Rauch’s most original point is not that AI makes hacking easier, but that it made hacking more profitable. The targets now are companies sitting on AI tokens, because tokens are “fungible with US dollars,” trade on black markets, and feed distillation attacks that copy a model’s abilities. That changes the threat model for every AI startup with a large API budget. His remedy is “models pointing guns at each other”: a reviewing agent, built from closed and open models, that distrusts whatever the coding agent wrote before it ships. Armstrong describes the same posture at Coinbase, where every line of code is scanned by a cyber model before production and internal systems are built to survive an engineer’s laptop being compromised. Feldman, earlier, called the bad week a once-in-30-years run of bad luck. The facts mostly back him on the outages, since the FAA blamed its ground stop on a router upgrade rather than an attack. Rauch’s essay title, “Everything hackable will get hacked,” is the better bet for the next year.

Every hardware founder on the tape gave the same warning in different words. Rinaudo Cliffton says investors who spent 13 years treating capex as “radioactive” now believe hardware is the only defensible business, but have never set up a factory, and that a project is only as good as its weakest supplier: “if there’s a zero anywhere in the chain, then the overall project is a zero.” Altekar says the hard part of autonomous boats is not the demo but surviving 15 to 55 days alone at sea under 10 to 20 g slamming loads. Franceschetti says hardware in 37 countries is “really really really hard.” That fits Zipline’s own situation: three partners each want a million deliveries a day, which Rinaudo Cliffton says turned the company’s entire operating plan into “the unit of sale.” Demand has stopped being the constraint for these companies. Supply chains, reliability and time are.

Harper Carroll’s closing prediction belongs in the same conversation as Baker’s open-source argument. She expects personal agents to sit between people and their feeds, picking what is worth seeing, the way Andrej Karpathy built his own filtered YouTube and Hacker News feeds. If that happens, engagement-optimized content loses its distribution advantage and useful content gains one. It also depends on the open, customizable models Baker, Rauch and Armstrong were defending: a filter you cannot inspect or tune is just another algorithm.

Key Takeaways

  • Andrew Feldman says Cerebras is 10 to 15 times faster than other AI compute, that speed has become what customers demand, and that Cerebras added AWS, Microsoft and Oracle as customers alongside OpenAI in the last six months.
  • Zipline is working with Uber Eats toward a million deliveries a day and expanding with Walmart to about 300 supercenters, after raising $1.5 billion at a $21.5 billion valuation.
  • Matteo Franceschetti says Eight Sleep is crossing $1 billion in cumulative sales, sells in 37 countries with three more coming, is still entirely direct to consumer, and wants to move into premium hotels and airplanes.
  • Brian Armstrong says decentralized AI needs a marketplace of open-weight models with financial accounts and payment rails, and that Coinbase stores 99% of customer crypto offline.
  • Armstrong’s personal security advice: a password manager, two-factor authentication not tied to a phone number, ideally a hardware key, and multi-signer withdrawals with liveness checks for large holdings.
  • Gavin Baker argues the market’s AI math (about $2 trillion of capex against under $200 billion of lab revenue) leaves out open-source inference and enterprises running their own AI on Dell hardware.
  • Baker and Brad Gerstner asked Nvidia’s board to take a slow, deliberate process and keep a Jensen Huang return open; Colette Kress is interim CEO, and Huang has not been charged with anything.
  • Baker estimates SpaceX’s first-party AI business, led by Grok Bot, is getting close to $10 billion of revenue, an estimate he says rests on third-party data.
  • Igor Kurganov says the lock-in case for selling chips to China fails because China is chip-limited, not energy-limited, and because AI is making it easy to switch chip platforms.
  • Guillermo Rauch says AI tokens are now a hacking prize worth stealing, and recommends deploying software behind the firewall and putting a skeptical AI reviewer in every deploy pipeline.
  • Vibhav Altekar says a 90-day Jones Act waiver showed big latent demand for domestic shipping, and that the US has had almost no commercial shipbuilding base since federal subsidies ended in the early 1980s.

Chapters

0:00 Andrew Feldman: Cerebras, Speed and a Thousand Millionaires

Feldman says AI has barely started delivering value, using the one-on-one tutor as his example. He reads the week’s outages as upgrades gone wrong, not a new normal: a working system is “at its weakest when you try and change it.” He says the market has swung toward speed and that demand for Cerebras is “off the chart.” On post-IPO wealth, he tells employees not to hold 99% of their net worth in one stock and not to rush any decision, He credits mentors Mark Leslie and Pierre Lamond, and says Leslie pushed him to be completely transparent with his team.

23:00 Keller Rinaudo Cliffton: Zipline’s Uber Eats and Walmart Scale-Up

Rinaudo Cliffton explains how Zipline tore up its three-year plan when several partners each asked for a million deliveries a day, and how scaling a hardware network means every supplier must scale in lockstep. Deliveries take two to five minutes once food is ready, and Zipline can deliver to any GPS coordinate, including a dorm’s seventh-floor pool deck. He argues the developing world is choosing between a US and a Chinese technology stack, that full-stack companies are easier to secure, and that investors new to hardware will spend the next few years learning how slow it is.

42:00 Matteo Franceschetti: Eight Sleep, Marriages and Hotels

Franceschetti says the most common feedback is that the Pod “saved my marriage” by ending fights over temperature. He travels about 250 days a year, says he has almost no jet lag, and tries new hardware in the Shenzhen and San Francisco offices. Eight Sleep reports a record-low return rate and plans five-star hotels next.

50:00 Brian Armstrong: Agentic Finance and AI on Defense

Armstrong says he does not want “one or two AI overlords,” and that teams of specialized agents, paying and hiring each other over Coinbase rails, are the antidote. On security he says there is a flight to quality during attacks, that attackers using AI must be met with AI on defense, and that Coinbase serves seven of the largest banks, 140 government agencies and 50 of the top 100 hedge funds. His hot take is that human-enhancement companies, following GLP-1 drugs, will reach trillion-dollar valuations, including through his biotech New Limit.

56:00 Gavin Baker: AI Demand, Jensen Huang and Starship

Baker says the demand he heard on stage from Michael Dell, Cerebras and SpaceX contradicts stocks that are pricing in a demand problem. He then makes his case for Huang: a patriot who arrived from Taiwan at nine and was sent to a reform school, and who has the best supply-chain relationships in Asia, which Baker says no one can replicate. He asks the board to let the Congressional testimony and internal investigation play out. On SpaceX, he calls the Starship upper-stage catch a step toward orbital compute and Starlink direct to cell, and says Grok Bot is now the AI tool he uses most.

1:19:00 Igor Kurganov: Export Controls and Open Weights

Kurganov argues that some chips are reaching China anyway, that recent Chinese models were trained at least partly on Blackwells, and that this needs to be stopped. He separates motivated reasoning from policy: maximizing shareholder value is fine, but it should not pose as geopolitics. On open versus closed models he expects the gap to grow, because labs will stop selling tokens and keep their best models for their own products.

1:29:00 Guillermo Rauch: Everything Hackable Will Get Hacked

The Vercel CEO says decades of human-built tech debt now needs rebuilding in memory-safe languages, and that coding agents can already flag bad choices if teams are willing to hear it. Enterprises mostly deploy Vercel behind the firewall, and he warns that ungoverned agents can leak data, citing the Hugging Face breach that OpenAI said its own models caused. He says Nvidia built an open ecosystem and that some critics have their own reasons to want it to fall. Vercel has open-sourced a security harness, deepsec.

1:40:00 Vibhav Altekar: Saronic, Shipyards and Weather

Saronic’s CTO recaps the Port Alpha shipyard in Brownsville, the June rescue of two downed pilots by a Corsair boat, and a new low-profile vessel for the Marine Corps. He says China has about half of global shipbuilding capacity and the US almost none. Reliability at sea is the main engineering problem, from salt corrosion to birds on sensors. His hot take: forecasts fall apart after nine to eleven days, AI weather models should push that toward thirty, and the missing data is mostly from the ocean.

1:49:00 Ron Biscardi: Building TomorrowX With AI

Biscardi says the event normally needs 18 months of planning, and AI let iConnections build the website, dashboards and app in days. He credits Grok Bot, saying his COO ran a dozen agents during the run-up. His hot take is that AI makes in-person meetings more valuable.

1:55:00 Harper Carroll: Teaching AI and Filtering Feeds

Carroll, formerly head of AI at Brev.dev (acquired by Nvidia in 2024), describes building an audience on fine-tuning guides, starting with 50 followers on X. She is launching a series on how AI works under the hood. Her hot take is that personal agents will filter social feeds, and she credits Stanford professor Mykel Kochenderfer as the mentor who convinced her she belonged in machine learning.

Notable Quotes

“We build the fastest AI compute on Earth. Not by a little bit, but 10 or 15x.”

Andrew Feldman, on Cerebras and the market’s shift toward speed

“So if there’s a zero anywhere in the chain, then the overall project is a zero.”

Keller Rinaudo Cliffton, on why hardware projects take longer than investors expect

“Now that the attackers are using AI, we’ve got to be using AI on defense.”

Brian Armstrong, on security at Coinbase

“I’ve been doing this 27 years, and I’ve met one Jensen, and I’ve met one Elon.”

Gavin Baker, on why Nvidia’s board should keep Huang’s return open

“Jensen said I’m not a car. I think now you’re a car.”

Igor Kurganov, on AI eroding Nvidia’s software lock-in

“Tokens are very precious. They’re fungible with US dollars, right?”

Guillermo Rauch, on why AI companies have become hacking targets

“I think the way that we consume media is going to change drastically.”

Harper Carroll, on agents filtering social feeds

Watch the full conversation here.

More From the TomorrowX Summit

Related Reading