Jeff Bezos explains why his AI startup Prometheus is building AI for physical engineering, tells Nvidia’s board not to hire an outside CEO, and describes how Robert Pattinson became James Bond. He spoke with co-founder Vik Bajaj at Semafor’s Silicon Valley & The World, a 32-minute session moderated by Semafor’s Liz Hoffman. Watch the full conversation on YouTube.
TLDW
Bezos argues that language models are excellent symbol manipulators but cannot design physical things, so Prometheus trains specialized physics models that it says compress months of engineering into hours. He says the company will sell tools broadly and also invest in the manufacturers it helps, because no walled garden could cover tens of trillions of dollars of manufacturing. He defends talking to every president, says housing is expensive because policy subsidized demand and constrained supply, and calls picking a villain a terrible way to solve problems. He advises Nvidia’s board to promote from inside if Jensen Huang does not return, predicts that AI investors will get more selective, and recounts the screen tests that made Pattinson the next Bond.
Thoughts
Bezos’s sharpest line was about someone else’s company. Asked about Nvidia, whose board accepted Jensen Huang’s resignation last week and named CFO Colette Kress interim CEO, he would not say whether Huang should come back, but he was blunt about what should happen if he does not: “I would not hire an outsider. Nvidia is too complicated.” He said outside CEOs at complex companies have “a vanishingly small chance of succeeding.” He was talking about Nvidia, but the claim reaches further. In a company that spans chip design, software and a global supply chain, the job is less about general management than about knowing how everything fits together, and that knowledge lives in the people who have been there. It also explains why he pushed back on the Amazon comparison. Andy Jassy’s succession was a years-long planned handoff to an insider, the opposite of an emergency search run by a recruiting firm. On the same day, Nvidia’s new chair said the board does not intend to reinstate Huang, so Bezos’s advice amounts to telling the board what kind of search to run.
The most useful business advice came during the casting example, and it went against current corporate thinking. Every executive has been told their proprietary data is a moat that AI labs want to take. Bezos agreed that a company with a genuinely valuable data set should “work pretty hard not to share it,” and then added: “I think it’s more rare than you think.” His example was precision casting, a craft thousands of years old in which experienced people still read X-rays and decide by trial and error where to add insulation to a mold. Those shops are not sitting on a data asset, because they never instrumented the process in the first place. Bajaj added that most of what a physical product reveals becomes public soon after it ships, and that Prometheus generates much of its own data through simulation. For most industrial companies, the honest assessment is that their advantage lies in know-how no one has written down, and an AI tool that makes that process repeatable is worth more to them than a database they believe is valuable.
The political section was easy to read as a billionaire defending himself before a newly Democratic House, and partly it was. Under that, though, Bezos made an argument about method that holds up regardless of who makes it. He described housing the way Amazon would describe a broken fulfillment process: skip the immediate problem, find the root cause, and you discover that policy “subsidized demand and constrained supply.” He pointed to Austin, which he said lowered housing costs while its population grew, partly by dropping parking minimums that burdened apartment developers. Then he compared blame to a management failure: an executive who answered a root-cause investigation by pointing fingers would probably be fired. He listed past villains (lawyers, Wall Street, landlords, and now big tech) to show that blame moves around while the underlying problem stays. Listeners can doubt his motives and still find the diagnostic useful.
On AI valuations, Bezos and Bajaj gave a more measured answer than either AI boosters or AI skeptics usually give. Bezos described the cycle plainly: early on, investors cannot tell good opportunities from weak ones, so “the good, the bad, and the ugly, they all get funded,” and then capital becomes more discerning. He said that sorting is already underway in AI and called it healthy, while insisting the technology itself is real. Bajaj supplied the timeline that should go with that optimism. Early cancer detection is “a 10-year journey,” drug discovery has been building for about 15 years, and the right posture is to “underpromise and overdeliver.” Together they make a better case than most AI pitches: expect a shakeout in companies, not in the technology, and expect the physical and biological payoffs to come later than the hype suggests but sooner than the skeptics expect.
The Prometheus pitch itself depends on a thought experiment worth remembering: put Newton, Einstein, Maxwell and three other great physicists in a room, ask them to build a toaster oven, and they would have no idea where to start. Knowledge works at different levels of detail, and reading the whole internet does not get a model down to the level where solder, airflow and thermal stress matter. The claims are large. Bezos said the company has about 250 people and roughly $14.5 billion raised (press reports of its funding rounds put the total at different figures), that it designs analog chips and radios in 12 hours instead of six to nine months, and that it worked out a design close to NASA’s Mars helicopter from a single prompt. They are unverified outside the company’s private alpha, which includes two projects at Blue Origin. The business model is what makes it more than a bigger model: Prometheus plans to invest in the manufacturers it helps and “benefit as a principal,” which suggests Bezos sees more value in owning a share of a faster physical economy than in selling software to it.
Key Takeaways
- Prometheus is building “AI for atoms”: specialized models for engineering and manufacturing, because Bezos says language models trained on text cannot design physical objects in detail.
- Bezos said the company has about 250 people, has raised roughly $14.5 billion and has already committed much of it to compute. It builds neural surrogates for numerical simulation in a single multi-physics model.
- The company says its tools design analog chips and radios in about 12 hours, work that would take 10 to 20 engineers six to nine months, and that a single prompt produced a design close to the Mars helicopter.
- Bezos ruled out a walled garden because manufacturing is worth tens of trillions of dollars. Prometheus will sell tools and also invest in some of the companies it helps.
- The product is in private alpha with early customers, including two projects inside Blue Origin. There is no date for a public launch.
- Bezos said truly valuable proprietary data is rarer than companies think, and that those who really have it should guard it.
- Bajaj said Prometheus respects proprietary data, trains on what is not proprietary and generates much of its own data.
- Bezos’s thesis is that “all civilizational wealth comes from invention,” and Prometheus exists to speed up the invention loop.
- He defended working with every president since Clinton, called Trump the most communicative of them, and said shutting business leaders out of government is a mistake.
- On affordability, he blamed housing costs on policies that subsidize demand and constrain supply, cited Austin as a city that lowered costs, and said the only fix is building more housing.
- If Nvidia does not bring back Jensen Huang, Bezos said, it should not hire an outsider. He called Amazon’s own handoff to Andy Jassy the most planned transition there has been.
- He expects AI investors to become more discerning now that early funding of every experiment is giving way to selection, and he called that shift healthy.
- Bezos said he was heavily involved in casting Robert Pattinson as James Bond after screen tests of five finalists, with costumes and sets, ended in what he called an undebated choice.
Chapters
00:01 Prometheus and AI for Atoms
Bezos describes language models as strong symbol manipulators that are good at math and law but cannot learn what can only be learned by doing, like gymnastics, dance or engineering. Six great physicists could not build a toaster oven, he says, because knowledge works at different levels of detail. Prometheus trains specialized physics models instead. He cites designs for chips, radios and valves, and a redesign of the Mars helicopter that worked out blade chord for the Martian atmosphere. Bajaj adds that even language models are specializing, pointing to Anthropic’s focus on coding.
05:14 Not a Walled Garden
Asked whether Prometheus will become a manufacturer itself, Bezos says the market is too large for any closed approach. The company will use more than one business model, selling tools broadly while investing in some customers so it benefits “as a principal.” Every product depends on a large supply chain with many processes, so the work can only be done alongside thousands of companies. He brushes off a question about merging his companies the way Elon Musk is doing.
07:36 Why Bezos Could Not Sit It Out
Bezos says he invested at the seed stage in late 2024 and grew too enthusiastic to stay on the sidelines. He calls the work “type two fun,” like climbing a mountain, with daily small breakthroughs and a missionary team. Prometheus is in private alpha with early customers, including two projects at Blue Origin. He would not give a date for a broader launch.
09:21 Trust, Data and the Casting Problem
The host raises big companies’ fear that forward-deployed engineers from AI labs are effectively scouting their businesses. Bajaj says Prometheus trains on non-proprietary data and generates its own, and that cutting product cycles by an order of magnitude is the real offer. Bezos walks through precision casting, where uneven cooling causes voids and hairline cracks that experts fix by trial and error. Such shops have little data to protect, and companies with truly unique data sets should guard them.
14:12 All Wealth Comes From Invention
Bajaj notes that casting will never be virtualized, so making casting houses faster benefits the whole supply chain. Bezos sets out the mission: the plow and the steam engine made everyone wealthier, inventors take a little and move on, and the invention lasts. Prometheus aims to accelerate the invention loop. When the host notes that the myth ended badly for Prometheus, Bezos points out that Hercules shows up and saves him.
15:50 Presidents, Democrats and Business Leaders
With Democrats taking control of the House, the host asks whether Bezos’s closeness to Trump will cost his companies. He says he has worked with every president since Clinton, still asks Obama for advice, and finds Trump unusual only in how much he reaches out. He says he has never used those relationships to advantage Amazon or himself, and that lawmakers lack ground truth about American business without them. Bajaj argues that fast-moving technology could help with long-stalled priorities in education, infrastructure and healthcare.
20:34 Housing, Root Causes and Picking a Villain
Bezos calls affordability the core issue and applies Amazon’s root-cause method to housing: subsidized demand plus constrained supply drives prices up. He cites Austin, which he says lowered housing costs as it grew, partly by scrapping parking requirements. Picking a villain is good politics but a terrible way to solve problems, he says, and he lists the villains of past decades before rejecting the idea that tech is the current one.
24:36 Nvidia After Jensen Huang
Bezos says he does not know enough about what happened inside Nvidia’s board meetings to say whether Huang should return, but that Huang deserves admiration and respect regardless. If Huang does not come back, he would not hire an outsider, calling outside CEOs at complex companies a last resort. He rejects comparisons to Amazon’s multi-year planned transition.
26:07 AI Investors Get Discerning, Science Takes Time
Responding to the point that Amazon lost money for years, Bezos says “That’s called investing.” He predicts the AI market will get more selective, as every technology cycle does after early experiments all get funded. Bajaj, who ran Google’s life sciences business, says healthcare AI needs calibrated timelines. Early cancer detection is a 10-year journey, though blood tests already catch many cancers, and drug discovery companies like Isomorphic and Insitro are close to speeding up the search for cures.
30:10 Choosing Robert Pattinson as James Bond
Bezos says he was heavily involved in choosing the first Bond under Amazon. The team cut a large group down to five finalists, ran old-fashioned screen tests with full costumes, real sets and three scenes each, and decided on Friday. All five were extraordinary, he says, but everyone, including director Denis Villeneuve, gravitated to Pattinson. The team had been biased against him at first because he is already so famous.
Notable Quotes
“I could read a thousand books on how to be a great gymnast, and I would still be a terrible gymnast.”
Jeff Bezos, on what language models cannot learn by reading
“If you actually have a very valuable proprietary data set, you know, my advice to you would be don’t share it and actually work pretty hard not to share it. I think it’s more rare than you think.”
Jeff Bezos, on companies guarding their data from AI labs
“All civilizational wealth comes from invention.”
Jeff Bezos, on the mission behind Prometheus
“Pick a villain is a good political technique, but it’s a terrible technique for solving problems.”
Jeff Bezos, on politics and affordability
“I would not hire an outsider. Nvidia is too complicated.”
Jeff Bezos, on Nvidia’s CEO search if Jensen Huang does not return
“We have to be scientifically rigorous and sort of underpromise and overdeliver.”
Vik Bajaj, on AI in healthcare and the physical world
“In the screening tests it was so obvious he is James Bond.”
Jeff Bezos, on casting Robert Pattinson
Watch the full conversation here, and read Semafor’s write-up here.
Related Reading
- NVIDIA Announces CEO Transition Nvidia’s announcement of Jensen Huang’s resignation and Colette Kress’s appointment as interim CEO.
- Ingenuity (Wikipedia) the NASA Mars helicopter whose design Prometheus says its tools rediscovered from a prompt.
- Casting (metalworking) (Wikipedia) background on the trial-and-error process Bezos uses to explain physical AI.
- Prometheus (Wikipedia) the myth behind the company name, including the Hercules ending Bezos says people forget.