Valued at 41 billion US dollars, Bezos is building the "OpenAI of the physical world"
Five years after stepping down as Amazon CEO, Jeff Bezos could not resist taking up a new job for himself.
On November 17, 2026 local time, at the Semafor "Silicon Valley and the World" Summit in Menlo Park, California, Bezos made a rare joint appearance with physicist Vik Bajaj. His latest title now is co-founder and co-CEO of AI startup Prometheus. This marks the first time the business giant has returned to the front line to serve as CEO since he handed over the helm of Amazon in 2021.
Bezos, who invested in the project back in its seed stage at the end of 2024, did not hide his excitement on stage, saying bluntly that the project is advancing so fast that he "could not just sit on the sidelines and watch".
What brings Bezos back to the battlefield is a huge high-stakes bet.
When you open the official homepage of Prometheus, there is currently only a minimalist brand logo and a warm glow background, with almost no public product details visible. But under the surface, this mysterious company with only about 250 employees has seen its valuation soar to 41 billion US dollars, having raised a total of approximately 14.5 billion US dollars in capital.
The extremely low-key Prometheus official homepage currently only displays the brand logo and has not opened its toolset to the public | Image source: Prometheus official website
During the 32-minute conversation, Bezos lifted the lid for the first time. The design of analog chips, radio equipment or valves that used to take an engineering team of 10 to 20 people 6 to 9 months to complete can be finished by Prometheus's system in 12 hours. Blue Origin, his commercial aerospace company, has already conducted private internal tests of this tool on two projects.
While Silicon Valley is still fiercely competing over chatbots and code-writing agents, Bezos has directly switched to a completely different game table.
Prometheus's ambition is to become the OpenAI of the physical world, using "AI for atoms" to restructure the tens of trillions of US dollars worth of physical manufacturing industry.
01. Newton Could Not Build a Small Toaster
Prometheus has raised as much as 14.5 billion US dollars, and many people's first reaction is that Bezos is also going to make humanoid robots.
He directly denied this speculation at the summit — Prometheus neither develops chat-focused large language models nor builds bipedal robots, its ultimate goal is to create an "Artificial General Engineer", abbreviated as AGE.
Why can't the all-powerful pure-text large language models today do the work of hardcore engineers?
Bezos gave a very vivid metaphor on site: "You can't learn engineering just by reading. Even if I read a thousand gymnastics books, I will still be a terrible gymnast; if I read a thousand dance books, I still can't dance."
Bezos said that if you lock six top physicists including Newton, Einstein and Maxwell in the same room and ask them to build a small toaster, these geniuses will most likely be at a complete loss. Understanding Maxwell's equations is completely useless for building a small toaster, because theoretical physics and engineering manufacturing operate at completely different granularities.
Bajaj, co-CEO who used to be a core executive at Google X and a professor at Stanford University School of Medicine, added another reality. Even in the field of pure language models, specialized division of labor is already inevitable. For example, Anthropic's leading programming capability stems from its high level of focus. When it comes to the physical world, even a trivial radio module in a mobile phone not only requires decades of accumulated human engineering experience, but also requires proficiency in thermal, mechanical and electromagnetic properties to complete encapsulation. In traditional engineering, running a single physical simulation often takes days to weeks.
To bridge this gap, Prometheus avoids the pure-text path, and directly builds a neural agent model for numerical simulation and a unified multi-physics field model at the underlying layer, creating a new generation of AI-driven CAD, physical simulation and engineering agents.
This interdisciplinary capability has shown astonishing intuition in internal tests. One of the engineers on the team, who previously participated in the development of the Mars Helicopter at NASA's Jet Propulsion Laboratory (JPL), tried to redesign the Mars Helicopter using only one prompt to the system.
Facing the extremely harsh environment on Mars, the model made independent trade-offs in the huge design space. It compared the quadcopter, coaxial counter-rotating helicopter and fixed-wing solutions one after another, and finally converged to a coaxial counter-rotating configuration that is highly similar to the real Mars Helicopter. It also calculated parameters such as the blade chord length for the thin Martian atmosphere, automatically generated requirement indicators, and even actively asked engineers questions when the intention was unclear.
Concept art of NASA's Ingenuity Mars Helicopter. An engineer who participated in the project used one single prompt to let Prometheus re-derive a highly similar coaxial counter-rotating configuration | Image source: NASA/JPL-Caltech
To tackle such a tough problem, algorithm talents who only know how to write code are far from enough, so Prometheus has assembled an extremely mixed 250-person "dream team".
For the AI talent pool, they have recruited a large number of top researchers from OpenAI, Google DeepMind, Meta and Microsoft. Kyle Kosic, co-founder of xAI who once led the construction of Musk's Colossus supercomputer, joined in early 2026 to take charge of computing power infrastructure, locking a large part of the company's financing directly into computing power procurement. In November 2025, they also quietly acquired General Agents, an agent AI lab co-founded by former DeepMind and Tesla researcher Sherjil Ozair, bringing real-time computer control capabilities into their portfolio.
On the other engineering side, the team roster is full of veterans in aerospace, chemistry, materials, marine and energy engineering, and even cross-domain experts from Epic Games' game engine department, Waymo's autonomous driving team and Red Bull F1's advanced technology department.
The advisory board even includes the chief scientist of the U.S. Army Ground Systems, as well as Chase Johnson, who used to be a lawyer for defense technology company Anduril, making the company's intention to enter the national defense industrial base and critical supply chains in the future obvious.
To support the R&D of this group of people, the company not only set up offices in San Francisco, London and Zurich, but also leased a plant of about 9,300 square meters in the former American Steel Industrial Park in West Oakland, California in August 2026, specifically for its high-ampere industrial power supply.
02. Unlocking the Deadlock of Industrial Data
With the technical architecture in place, every company trying to do industrial AI will hit a wall: where is the data?
OpenAI can crawl web pages across the entire Internet to feed GPT, but the process parameters for manufacturing aircraft engines, precision valves and analog chips have never been available on the Internet.
Bezos talked about the precision casting process that has been used for thousands of years at the summit. Even today, this industry still relies heavily on veteran workers manually finding cracks in industrial X-ray films or tomography scans, then adding heat insulation layers to specific positions of the molds based on experience, and making repeated trials and errors. A large number of traditional factories do not even have the most basic digital monitoring equipment, so there is no way to collect data at all. Even if data exists, it is the lifeblood of manufacturing enterprises. Bezos himself said bluntly on stage that if enterprises have unique proprietary data, "never share it, and strictly prevent it from leaking out".
Since ready-made data cannot be obtained, how does Prometheus feed its large model?
The answer lies in the laws of physics themselves. Bajaj explained that they use their self-developed multi-physics field simulation system to generate high-quality synthetic data, which greatly reduces reliance on sensitive enterprise data, and can complete the training of the base model after stripping out core proprietary processes. At the same time, Bezos has turned his own aerospace ecosystem into the harshest training ground, and two projects of Blue Origin are already running internal tests.
Blue Origin's New Glenn rocket lifts off. The company is currently privately testing Prometheus tools in two internal projects | Image source: Wikimedia Commons
But what really shocked Wall Street is their in-depth layout in capital and business models.
As early as the beginning of 2026, foreign media broke the news intensively that Bezos and Bajaj were planning to set up a manufacturing transformation holding company, and went to Singapore and the Middle East to contact sovereign wealth funds, planning to raise up to 100 billion US dollars in funds as rumored, to directly acquire traditional engineering, construction and manufacturing enterprises impacted by AI, which is known to the outside world as "the industrial version of Berkshire".
At the summit in November, Bezos gave a clearer business definition for this strategy. He did the math: the global physical manufacturing industry is a huge market worth tens of trillions of US dollars, "it is impossible to build a walled garden here", and Prometheus must cooperate openly with thousands of external companies.
Opening up the tools is only the first step. Bezos revealed that in addition to providing tools and services, Prometheus will also directly invest in some traditional enterprises empowered by its own tools, as a stakeholder to share the growth dividends after their efficiency surges.
As for outside speculation that he will follow Musk's example to merge Amazon, Blue Origin and Prometheus into a super empire, Bezos laughed on the spot and said:
"That's a funny idea, there is nothing to announce".
03. The Atomic Fork in the Second Half of the AI Race
Bezos's vision is backed by real money strength, and Prometheus's financing speed can only be described as staggering.
This company, which was incubated at the Foresite Labs incubator in 2024 with the early code name FL2024-008 INC, disclosed a $6.2 billion first round of financing when it first appeared in November 2025; in June 2026, it announced a new round of $12 billion financing (including some early funding arrangements), pushing its valuation to $41 billion. According to Bezos's latest statement on stage in November, the company has raised a total of approximately $14.5 billion in capital so far. Its shareholder list includes JPMorgan Chase, Goldman Sachs, BlackRock, DST Global, ARCH Venture Partners and many other well-known institutions.
The logic behind top capital piling heavy bets is very clear: the AI industry in 2026 is going through a drastic fork from "bit involution" to "atomic moat".
In the digital world, OpenAI and Anthropic dominate text reasoning and code generation, but the anxiety of peaking pure text corpora and software homogenization has spread. It is thought-provoking that Amazon, where Bezos serves as Executive Chairman, just invested an additional 25 billion US dollars in Anthropic in April 2026, in exchange for a 1 trillion US dollar cloud spending commitment. While using Amazon to stabilize the cloud infrastructure of the digital world, Bezos himself turned around and jumped into the more challenging physical world.
On this new track, players are dividing their territories by granularity. Liam Fedus, former Vice President of Research at OpenAI, and the former head of materials AI at DeepMind founded Periodic Labs, which raised a $300 million seed round to focus on micro new material discovery with "AI scientists plus autonomous robot labs". Bezos also personally participated in the follow-up investment. There are also other engineering simulation players such as Neural Concept and P-1 AI.
In contrast, Prometheus is occupying the heaviest macro engineering design and manufacturing link that directly generates commercial value.
However, it is worth noting that when AI is brought from the screen into factory workshops, the fault tolerance rate drops to zero instantly.
If a chatbot hallucinates, at worst it writes a wrong marketing copy, or makes programmers modify a few more lines of bugs. But if an industrial engineering large model hallucinates, it means that a valve bursts under high pressure, an entire batch of manufactured analog chips is scrapped, or even a rocket disintegrates on the launch pad. The verification cost in the physical world, as well as the strict FAA and FDA-level compliance reviews, cannot be easily bypassed just by stacking computing power.
The more fundamental challenge for Prometheus and its competitors lies in the word "general". An anonymous early core researcher at Prometheus confessed to the media that the breadth of the engineering world is so desperate that "there is no such thing as a 'universal engineer' at all". To make the same underlying model generalize freely across highly heterogeneous disciplines such as mechanics, thermodynamics, electromagnetism and chemical engineering, the technical difficulty far exceeds that of language models.
To this day, Prometheus still has not given a public product release timetable.
Beyond technology, there is another unavoidable social tension. In an interview in June 2026, Bezos strongly refuted the unemployment panic caused by AI, predicting that if the cost of invention drops to one-tenth of the current level, more than 10 times new opportunities will be created, which will eventually lead to "labor scarcity". In the future, only one person in a dual-income family needs to work, and people will never have to work overtime.
But the other side of reality is a bit harsh. In the past year alone, Amazon, where he serves as Executive Chairman and has more than 1.5 million employees worldwide, has laid off tens of thousands of employees under the automation strategy promoted by CEO Andy Jassy.
In ancient Greek mythology, Prometheus stole fire to give humans the power to forge metals and transform nature. Now Bezos is trying to put the "fire of engineering design" into an algorithmic black box.
Whether this tens of billions of dollars worth of flame will let humans completely say goodbye to overtime, or reshape the jobs of countless engineers first, the answer may start counting down in the heavy factory buildings in West Oakland.
*Cover image source: Semafor / AP