Semafor interview with Jeff Bezos and Prometheus executive Vik Bajaj, November 17, 2026
A physical-AI model trained for engineering, not language
In a Semafor interview, Jeff Bezos described Prometheus as an "AI for atoms" company built on the premise that large language models, however capable, cannot do engineering. Frontier models are trained on the internet's corpus of human knowledge, he said, which makes them strong at math and law but weak at tasks that cannot be learned by reading. "I could read a thousand books on how to be a great gymnast and I would still be a terrible gymnast," Bezos said. He offered a thought experiment: put the six greatest physicists in a room and ask them to build a toaster oven, and "they would have no idea how to do it," because "understanding Maxwell's equations doesn't help you build a toaster oven at all." His conclusion was that knowledge operates at different levels of detail and that engineering requires specialized models.
Bezos said Prometheus now has 250 employees and has raised approximately $14.5 billion, much of which has already been committed to compute. The company builds neural surrogates for numerical simulation and runs a single multiphysics model. He said the team is already designing analog chips, radios and other components that would take a team of 10 to 20 people six to nine months, and is doing so in 12 hours. He also described an internal test in which a colleague who had worked on NASA's Mars helicopter at JPL asked the tools to redesign it from a single prompt. According to Bezos, the system explored quadcopter, counter-rotating and fixed-wing configurations, converged on a design that "looks very much like the Mars helicopter," calculated blade chord for the Martian atmosphere, generated its own requirements and asked clarifying questions when it needed to.
Bajaj added that designing even a cell-phone radio draws on decades of design tradition and a command of physics, thermal properties and mechanical packaging. He noted that specialization is already visible in language models, citing Anthropic's coding models, which he said were for a time the best because they were focused on coding. "Engineering is very diverse and very heterogeneous," he said.
Not a walled garden, and not only a toolmaker
Pressed on whether Prometheus intends to become a manufacturer of toasters and jet engines, Bezos said the opportunity is too large for that. He put the manufactured world at "tens of trillions of dollars" of market and said no single approach or business model will cover it. "I see no scenario where any kind of walled garden approach could succeed here," he said, arguing that any product sits atop a vast supply chain of different manufacturing processes that can only be addressed in cooperation with thousands of companies.
The company's approach will include making tools available to customers and also investing in some of the companies it helps, so that it benefits "as a principal instead of exclusively as a tool maker," Bezos said. Asked whether Prometheus, Blue Origin and Amazon might be knitted more tightly together, as Elon Musk is attempting with his businesses, Bezos laughed and said he had "nothing to announce."
On timing, Bezos said Prometheus is in private alpha with several companies, including Blue Origin, which is using the tools on two internal projects. He declined to give details of those projects or a date for broader commercial availability.
Addressing the data-trust problem
The interviewer raised a growing concern among large companies working with frontier labs: that embedded engineers function as reconnaissance, studying a business before building a competing product. Bajaj said Prometheus collaborates with customers and respects the proprietary nature of engineering. He said the company can train on the parts of a data set that are not proprietary, noting that products become known to the public quickly after release. Prometheus also generates much of its own data and is less reliant on the kind of data that triggers these concerns, he said. He argued that shortening product development cycles, sometimes by an order of magnitude, is a large benefit to customers.
Bezos used metal casting to illustrate where the value lies. Casting is thousands of years old, he said, and precision aerospace parts are often made this way, yet it remains "very trial and error." Metal cools at different rates in thin and thick sections of a mold, producing voids and hairline fractures. Today, experienced technicians X-ray a part, identify defects, add insulation to the mold and try again. "It's actually not an easy thing to simulate," Bezos said, and such companies often have not instrumented their processes, so they have little proprietary data to protect. Where a company does hold a valuable unique data set, he said, "my advice would be don't share it," though he thinks such data sets are rarer than companies believe.
Bajaj said the casting example shows the physical economy cannot be virtualized: a foundry will always be needed, but faster output lets it take on more customers, and groups waiting 9 to 12 months for a complex cast part might get it in a month. He called that a compounding advantage for the entire supply chain....