TechFlow Logo
Login/ Sign up
ETH Gas
Gwei
Fear
gas
Valued at $41 billion, Bezos builds 'OpenAI of the physical world'

Valued at $41 billion, Bezos builds 'OpenAI of the physical world'

2026.11.20
Share
Bezos doubles down on the industrial large model Prometheus, launching a revolution in physical manufacturing from bits to atoms.

Author: Yuhangyuan, GeekPark

AI has devoured software; now, it’s coming for industry.

Five years after stepping down as Amazon CEO, Jeff Bezos couldn’t resist, taking up another job for himself.

On the local date of November 17, 2026, at Semafor's "Silicon Valley and the World" summit in Menlo Park, California, Bezos took the stage alongside physicist Vik Bajaj—a rare appearance. His newest title is Co-founder and Co-CEO of the AI startup Prometheus. This marks his first return to an active CEO role since handing over the reins of Amazon in 2021.

Bezos, who invested during the seed round back at the end of 2024, showed no restraint in his excitement on stage, bluntly stating that the project was advancing too fast for him to simply "sit on the sidelines watching."

What brought Bezos back to the battlefield was a massive bet.

Visiting Prometheus’s official website homepage reveals only a minimalist brand logo against a warm-hued halo background, with virtually no public product details visible. But beneath the surface, this mysterious company of just around 250 employees has seen its valuation skyrocket to $41 billion, having raised approximately $14.5 billion in cumulative funding.

Figure: The highly discreet Prometheus website homepage currently displays only the brand logo, with its toolkit yet to be released to the public.

Over a 32-minute conversation, Bezos pulled back the curtain for the first time. Designs for analog chips, radio equipment, or valves that used to require engineering teams of 10 to 20 people and take 6 to 9 months can now be completed by Prometheus’s system within 12 hours. His commercial aerospace company, Blue Origin, is already privately testing this toolkit on two projects.

While Silicon Valley remains mired in intense competition over chatbots and coding agents, Bezos has simply switched tables.

Prometheus’s ambition is to become the OpenAI of the physical world, using “atomic AI” to reconstruct trillions of dollars worth of the real manufacturing sector.

Newton Couldn't Build a Toaster

With Prometheus's $14.5 billion funding scale, many people's immediate reaction was that Bezos was stepping into the humanoid robot race as well.

He shut down that speculation directly at the summit: Prometheus builds neither conversational language models nor bipedal robots. Its ultimate goal is to create an "Artificial General Engineer" (AGE).

Why can today's omnipotent pure-text large language models not handle the work of hardcore engineers?

Bezos presented a vivid analogy on stage: "You can't learn engineering just by reading. Even if I read a thousand books on gymnastics, I’d still be a terrible gymnast; read a thousand books on dance, and I still wouldn’t know how to dance."

Bezos said that if you locked six top-tier physicists like Newton, Einstein, and Maxwell in a room and tasked them with building a toaster, they would most likely have no idea where to start. Understanding Maxwell's equations doesn't help build a toaster, because theoretical physics and engineering manufacturing operate on entirely different scales of granularity.

Co-CEO Bajaj, former core executive at Google X and professor at Stanford University School of Medicine, added another reality. Even in the pure language model domain, specialized division of labor is now inevitable; Anthropic's lead in programming capabilities stems precisely from this high degree of focus. In the physical world, however, even an inconspicuous radio module in a smartphone requires decades of human engineering accumulation, plus simultaneous mastery of thermodynamics, mechanics, and electromagnetic properties for packaging. In traditional engineering, merely setting up a single physical simulation often takes days to weeks.

To bridge this gap, Prometheus bypassed the pure-text route. At its foundation, it directly built neural proxy models for numerical simulation and unified multi-physics field models, forming a new generation of AI-driven CAD, physical simulation, and engineering agents.

This interdisciplinary capability demonstrated astonishing intuition during internal testing. A team member who previously worked on the Mars helicopter development for NASA's Jet Propulsion Laboratory (JPL) tried giving the system just one prompt to redesign the Mars helicopter.

Confronted with Mars's extremely harsh environment, the model independently weighed options within a vast design space. It compared quadcopter, coaxial counter-rotating helicopter, and fixed-wing concepts, ultimately converging on a coaxial counter-rotating configuration highly similar to the actual Mars helicopter. It also calculated parameters like blade chord length tailored to Mars's thin atmosphere, automatically generated requirement metrics, and even proactively questioned engineers when prompts were ambiguous.

Figure: Concept art of NASA's Ingenuity Mars Helicopter. An engineer who participated in the project used a single prompt to get Prometheus to derive a highly similar coaxial counter-rotating configuration.

To tackle such hard problems, relying solely on algorithmic geniuses who write code is far from enough. Prometheus has assembled an "all-star team" of 250 people with an extremely eclectic mix of expertise.

For AI personnel, they recruited a large number of top researchers from OpenAI, Google DeepMind, Meta, and Microsoft. Kyle Kosic, co-founder of xAI and former lead of the construction of Musk's Colossus supercomputer, joined in early 2026 to personally oversee computing infrastructure, locking a significant portion of the company's funding directly into compute procurement. In November 2025, they quietly acquired General Agents, an agentic AI lab co-founded by former DeepMind and Tesla researcher Sherjil Ozair, securing real-time computer control capabilities.

On the engineering side, the roster is filled with veterans from aerospace, chemistry, materials, ocean, and energy engineering, along with crossover experts from Epic Games' game engine division, Waymo's autonomous driving team, and Red Bull F1 Advanced Technology.

The advisory board features the Chief Scientist for U.S. Army Ground Systems and Chase Johnson, formerly a lawyer at defense tech company Anduril, making the intention to break into the defense industrial base and critical supply chains abundantly clear.

To support their R&D, besides establishing offices in San Francisco, London, and Zurich, the company leased roughly 93,000 square meters of facility space at the former American Steel Industrial Park in West Oakland, California, in August 2026, specifically drawn to its high-amperage industrial power grid.

Untangling the Industrial Data Knot

The technical architecture is in place, but every company attempting industrial AI hits a wall: where does the data come from?

OpenAI can crawl the entire web to feed GPT, but process parameters for manufacturing aircraft engines, precision valves, and analog chips are never publicly available on the internet.

Bezos discussed precision casting processes that have continued for thousands of years at the summit. To this day, the industry heavily relies on veteran craftsmen manually inspecting industrial X-rays or tomographic scans for cracks, then applying insulation layers to specific mold locations based on experience, failing repeatedly until success. Many traditional factories lack even basic digital monitoring equipment, making data collection impossible. Even when data exists, it is a manufacturer's lifeblood. Bezos stated plainly on stage that if companies truly possess proprietary data, they "must absolutely not share it and must strictly prevent leakage."

If ready-made data can't be obtained, how does Prometheus feed its large models?

The answer lies within the laws of physics themselves. Bajaj explained that they generate high-quality synthetic data using a self-developed multi-physics simulation system, drastically reducing reliance on sensitive corporate data and enabling base model training once core proprietary processes are stripped away. Meanwhile, Bezos has turned his own aerospace ecosystem into the harshest proving ground, with two Blue Origin projects already running internal tests.

Figure: Blue Origin's New Glenn rocket launch. The company is currently privately testing Prometheus tools on two internal projects.

But what truly sent shockwaves through Wall Street is their deep layout in capital and business models.

As early as early 2026, foreign media extensively reported that Bezos and Bajaj were planning to establish a manufacturing transformation holding company, traveling to Singapore and the Middle East to court sovereign wealth funds. They plan to raise rumored capital up to $100 billion to directly acquire traditional engineering, architectural, and manufacturing firms impacted by AI, earning the nickname "Industrial Berkshire."

By the November summit, Bezos gave clearer commercial direction to this strategy. He ran the numbers: the global physical manufacturing market is a colossal sector worth tens of trillions of dollars. "A walled garden approach is impossible," he noted. Prometheus must open up and partner with thousands of external companies.

Releasing open tools is just the first step. Bezos revealed that beyond providing tools and services, Prometheus will also directly invest capital into select traditional enterprises empowered by their toolkit, positioning itself as a stakeholder to share in the growth dividends following their efficiency leaps.

Regarding outside speculation about whether he would emulate Musk and merge Amazon, Blue Origin, and Prometheus into a single super-empire, Bezos chuckled on stage:

“That’s a funny idea. There’s nothing to announce.”

The Atomic Fork in the Second Half of AI

Bezos's vision is backed by substantial financial strength; Prometheus's fundraising speed can only be described as terrifying.

Hatched in 2024 within incubator Foresite Labs under the initial codename FL2024-008 INC, the company disclosed a $6.2 billion first closing upon its debut in November 2025. In June 2026, it announced a new $12 billion financing round (including some early tranches), pushing its valuation to $41 billion. According to Bezos's latest remarks on stage in November, the company has actually raised approximately $14.5 billion cumulatively so far. The shareholder list includes JPMorgan Chase, Goldman Sachs, BlackRock, DST Global, ARCH Venture Partners, and others.

The logic behind top-tier capital clustering here is crystal clear: in 2026, the AI industry is undergoing a drastic fork from "bit-level involution" to an "atomic moat."

In the digital realm, OpenAI and Anthropic dominate text reasoning and code generation, but anxiety over plateauing pure-text corpora and software homogenization has spread. Intriguingly, Amazon, where Bezos serves as Executive Chairman, recently added a $25 billion investment to Anthropic in April 2026 in exchange for a $100 billion cloud spending commitment. While anchoring the digital cloud backbone via Amazon, Bezos himself turns and dives headfirst into the harder-to-nibble physical world.

On this new track, players are carving out territory by granularity. Former OpenAI Research Vice President Liam Fedus and the former DeepMind Head of Materials AI founded Periodic Labs, securing a $300 million seed round. Focusing on micro-new material discovery with an "AI scientist plus autonomous robotics lab" model, it even attracted a personal follow-on investment from Bezos. Other engineering simulation players include Neural Concept and P-1 AI.

In contrast, Prometheus is positioned at the heaviest segment, which also directly generates commercial value: macro-engineering design and manufacturing.

However, it is important to note that pulling AI out of screens and into factory floors instantly drives the error tolerance rate to zero.

If a chatbot hallucinates, it might just write a flawed marketing copy or force programmers to fix extra lines of bugs. But if an industrial engineering large model hallucinates, it means valves exploding under high pressure, entire batches of fabricated analog chips scrapped, or even rockets disintegrating on the launchpad. The validation costs of the physical world, along with rigorous compliance reviews at the FAA and FDA levels, cannot be easily bypassed by merely piling up compute power.

The more fundamental challenge for Prometheus and its competitors lies in the word "general." An anonymous core early researcher at Prometheus told the media candidly that the breadth of the engineering world is staggering; "the concept of 'a general engineer' simply does not exist." Expecting the same foundational model to freely generalize across highly heterogeneous disciplines like mechanical engineering, thermodynamics, electromagnetics, and chemical processing simultaneously presents a technical difficulty far exceeding that of language models.

As of today, Prometheus has still not provided a timeline for a public product release.

Beyond technology, there is an unavoidable layer of social tension. During a June 2026 interview, Bezos strongly pushed back against AI-induced job loss fears, predicting that if the cost of invention drops to one-tenth, it would create over ten times as many new opportunities, eventually triggering a "labor shortage." In the future, dual-income families would only need one person working, and people would never have to work overtime again.

But the flip side of reality is somewhat cold. Just over the past year, Amazon, where he serves as Executive Chairman and which boasts over 1.5 million global employees, has laid off tens of thousands of workers under CEO Andy Jassy's automation-driven strategy.

In Greek mythology, Prometheus stole heavenly fire, granting humanity the majestic power to forge metals and transform nature. Today, Bezos attempts to pack "the heavenly fire of engineering design" into an algorithmic black box.

Whether this hundred-billion-dollar flame will allow humanity to completely bid farewell to overtime, or instead reshape the livelihoods of countless engineers first, the countdown to the answer may have already begun inside West Oakland's heavy industrial facilities.