Five years after stepping down as Amazon chief executive, Jeff Bezos has found another job. On 17 November 2026 at the Semafor Silicon Valley and the World summit in Menlo Park, he appeared on stage with physicist Vik Bajaj. His new title is co-founder and co-chief executive of AI start-up Prometheus. It is the first time the billionaire has returned to an operating CEO role since handing over Amazon in 2021.
Bezos invested at the seed stage in late 2024. On stage he made no attempt to hide his excitement, saying the project was moving so fast that he could not sit on the sidelines.




Prometheus is a large bet. Its website shows almost nothing beyond a logo and a warm glow. Under the surface, a company of roughly 250 people is valued at USD 41 billion and has raised about USD 14.5 billion.
An AI engineer, not a chatbot or a humanoid
In a 32-minute conversation, Bezos lifted the cover. Work that once took a team of 10 to 20 engineers six to nine months, from analogue chip design to radio equipment and valves, can now be done by Prometheus in 12 hours. Blue Origin, his space company, is already testing the tools privately on two projects.
While Silicon Valley competes over chatbots and coding agents, Bezos has moved to another table. Prometheus wants to be the OpenAI of the physical world, using AI over atoms to reshape a manufacturing base worth tens of trillions of dollars.
Many assumed the USD 14.5 billion raise meant another humanoid robot company. Bezos ruled that out on stage. Prometheus will not build a language model for conversation, nor a bipedal robot. Its stated goal is an Artificial General Engineer, or AGE.
Why can a text model that seems capable of anything not do real engineering? Bezos offered a plain analogy. You cannot learn engineering by reading. Read a thousand gymnastics books and you are still a poor gymnast. Read a thousand dance books and you still cannot dance.
He imagined six leading physicists, including Newton, Einstein and Maxwell, locked in a room and asked to build a small oven. They would probably have no idea where to start. Understanding Maxwell’s equations does not help, because theoretical physics and engineering operate at completely different grains of detail.
Bajaj, a former Google X executive and Stanford medical school professor, added a second constraint. Even inside language models, specialisation is unavoidable. Anthropic leads on coding because it concentrates. In the physical world, even an unremarkable radio module in a phone depends on decades of human engineering and demands simultaneous command of heat, mechanics and electromagnetics to package it. A single physics simulation in traditional engineering can take days or weeks.
To cross that gap, Prometheus skipped the pure-text route. It built neural surrogate models for numerical simulation and a unified multi-physics model underneath AI-driven computer-aided design, simulation and engineering agents.
The mix produced unexpected intuition in internal tests. An engineer who worked on NASA’s Jet Propulsion Laboratory Mars helicopter asked the system, in one prompt, to redesign the aircraft. Faced with the thin Martian atmosphere, the model compared quadrotors, coaxial counter-rotating helicopters and fixed wings, then converged on a configuration close to the real Mars helicopter. It calculated blade chord and other parameters, generated requirement metrics and asked the engineer questions when the intent was unclear.
A deliberately mixed team of 250
Algorithm talent alone cannot carry this. Prometheus assembled an unusual team. Researchers came from OpenAI, Google DeepMind, Meta and Microsoft. Kyle Kosic, a xAI co-founder who led the Colossus supercomputer build, joined in early 2026 to run compute infrastructure, with a large share of funding locked into compute purchases. In November 2025 the company quietly acquired General Agents, an agentic AI lab co-founded by a former DeepMind and Tesla researcher, adding real-time computer control.
The engineering side is filled with veterans of aerospace, chemistry, materials, ocean and energy work, plus people from Epic Games, Waymo and Red Bull’s advanced technology arm. Advisers include the US Army’s chief scientist for ground systems and a former Anduril lawyer, a clear signal of interest in the defence industrial base and critical supply chains.
Alongside offices in San Francisco, London and Zurich, the company rented about 9,300 square metres of the former American Steel industrial site in West Oakland in August 2026, chosen for its high-amperage industrial power.
Where does the data come from?
Every industrial AI company hits the same wall. OpenAI can crawl the web for training text, but the process parameters for aircraft engines, precision valves and analogue chips are not on the internet.
Bezos described precision casting, a craft refined over thousands of years. Even now, the industry relies on experienced workers reading industrial X-rays or scans for cracks, then adding insulation at chosen points in the mould through trial and error. Many factories lack even basic digital monitoring. Where data exists, it is the company’s crown jewel. Bezos said openly that firms with proprietary data should never share it.
Prometheus generates high-quality synthetic data from its own multi-physics simulation, stripping out proprietary process detail before training foundation models. Blue Origin serves as a demanding internal proving ground.
The capital structure is equally unusual. Early in 2026, reports described a plan to create a manufacturing transformation holding company, with approaches to sovereign wealth funds in Singapore and the Middle East for as much as USD 100 billion, to acquire traditional engineering, construction and manufacturing businesses disrupted by AI. At the November summit Bezos framed the commercial logic more clearly. Global physical manufacturing is a market worth tens of trillions, he said, and cannot become a walled garden. Prometheus must work with thousands of outside companies. Beyond tools and services, it will invest directly in traditional firms using those tools and share in their efficiency gains. On whether he would merge Amazon, Blue Origin and Prometheus into one empire, he laughed that the idea was amusing and there was nothing to announce.
The atom fork in AI’s second half
Prometheus was incubated at Foresite Labs in 2024 under the working name FL2024-008 INC. It disclosed a USD 6.2 billion first round in November 2025, then a further USD 12 billion round in June 2026 that pushed valuation to USD 41 billion. Bezos put cumulative funding at about USD 14.5 billion in November. Shareholders include JPMorgan, Goldman Sachs, BlackRock, DST Global and ARCH Venture Partners.
The logic for top capital is straightforward. In 2026 AI is splitting from competition over bits towards a moat built from atoms. In the digital world, OpenAI and Anthropic dominate text reasoning and code generation, but text corpora are plateauing and software is converging. Amazon, where Bezos remains executive chairman, added as much as USD 25 billion to Anthropic in April 2026 in exchange for a USD 100 billion cloud spending commitment. He is securing the digital cloud base while personally moving into the harder physical world.
New entrants are dividing the territory by scale. Periodic Labs, founded by a former OpenAI research vice-president and a former DeepMind materials AI lead, raised USD 300 million in seed funding for AI scientists plus autonomous robotic labs aimed at new materials, with Bezos personally participating. Neural Concept and P-1 AI work on engineering simulation. Prometheus is positioning at the heavy end, in macro engineering design and manufacturing where commercial value appears fastest.
But moving AI from the screen to the factory floor takes tolerance to zero. A chatbot hallucination produces a bad marketing line or a few extra bug fixes. An industrial model hallucination means a valve bursting under pressure, a scrapped batch of analogue chips or a rocket breaking up on the pad. Physical verification cost and FAA or FDA grade compliance cannot be bypassed with more compute.
The deeper challenge is the word general. An early researcher told the media that the breadth of engineering is disheartening and that there is no such thing as a general engineer. Making one underlying model move freely between mechanics, thermodynamics, electromagnetics and chemical engineering is far harder than language modelling. Prometheus still has no public product release date.
There is also a social tension. In June 2026 Bezos argued that if invention costs fall to a tenth, more than ten times as many opportunities appear, eventually producing labour scarcity, one-earner households and an end to overtime. Yet over the past year Amazon, with more than 1.5 million employees, has cut tens of thousands of jobs under an automation push led by chief executive Andy Jassy.
Prometheus stole fire for the physical world. Whether that fire ends overtime or first reshapes the careers of engineers is now counting down in a heavy industrial hall in West Oakland.
Editor’s note: This is an adapted translation of the original 36Kr report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://36kr.com/p/4077584744600709.