TL;DR

Jeff Bezos came out of semi-retirement to co-run Project Prometheus, a physical AI startup that raised $6.2 billion at a $30 billion valuation. The plan: buy old manufacturing and aerospace companies, plug AI into their operations, and flip the entire industrial sector into something that looks more like software. He’s also raising a separate $100 billion fund to do this at scale. Bezos is betting AI’s biggest returns will come from factories.

Bezos Is Running This One Himself

When Jeff Bezos stepped down as Amazon’s CEO in 2021, most people assumed he’d spend his time on Blue Origin and philanthropy. That lasted about four years.

In November 2025, Bezos co-founded Project Prometheus with Vikram “Vik” Bajaj, a physicist and chemist who previously ran experimental programs at Google X and co-founded the AI incubator Foresite Labs. Both hold the co-CEO title. This is the first time Bezos has taken a hands-on operational role since leaving Amazon.

But the scale here is what caught my attention.

The initial funding round pulled in $6.2 billion, partly from Bezos himself, at a $30 billion valuation. For context, that’s larger than xAI’s first round. It’s smaller than OpenAI’s $122 billion raise, but Prometheus is targeting a completely different market. All before shipping a product.

What “Physical AI” Actually Means

Every major AI company right now is fighting over the same territory: language models, coding assistants, image generators, chatbots. Prometheus is going somewhere else entirely.

The company builds AI systems designed to understand and optimize physical processes: jet engine design, semiconductor fabrication, automotive manufacturing. You can’t solve these by training on internet text. They require data from actual machines, real production lines, and years of domain-specific engineering knowledge.

Bajaj described it in a Financial Times interview as building AI that “reasons about the physical world the way LLMs reason about language.” That’s a big claim. But the hiring pattern suggests they’re serious about backing it up.

The Talent Raid

As of early 2026, Prometheus had already hired over 120 employees, poaching researchers from Meta, OpenAI, DeepMind, and Google. But the April 7 hire is the one that turned heads.

Kyle Kosic, one of xAI’s first 11 employees and a co-founder, left OpenAI (where he’d landed after xAI) to join Prometheus. Kosic helped build Colossus, the supercomputer that powers Grok. At Prometheus, he’ll lead AI infrastructure.

Losing a co-founder to a competitor is embarrassing for any company. Losing one to a five-month-old startup says something about what Bezos is offering. Multiple reports say Prometheus is paying top-of-market compensation and giving senior hires unusual equity stakes, the kind of packages that make even well-paid AI researchers reconsider.

The company now has offices in San Francisco, London, and Zurich, and it keeps growing.

The Berkshire Hathaway Play

Most AI startups build products and sell them. Prometheus wants to buy entire companies.

The long-term plan, reported by the Wall Street Journal and TechCrunch, involves raising up to $100 billion for a fund structured like a “Berkshire Hathaway-style holding company.” The strategy: acquire stakes in manufacturing firms (chipmakers, defense contractors, aerospace suppliers) and use their proprietary data to train specialized AI models. Then deploy those models back into the acquired companies’ operations.

It’s a closed loop. Buy the company, get the data, build the AI, improve the company, use the profits to buy more companies. If it works, Prometheus becomes an industrial conglomerate powered by AI.

The target sectors aren’t random picks. Semiconductor manufacturing, defense, and aerospace are all capital-intensive with decades of proprietary process data locked in internal systems. Margins are thin. Automation could widen them significantly.

A Different Kind of AI Company

The AI startup space in 2026 is crowded. OpenAI, Anthropic, Google DeepMind, xAI, Mistral, Cohere are all competing for roughly the same market: making LLMs better and selling API access or subscriptions.

Prometheus wants the data locked inside factories, on production floors, in engineering simulations that have never been digitized.

The easy wins in AI (text, code, images) are getting commoditized fast. Models keep getting cheaper, margins keep shrinking. Physical AI is harder to build but also harder to replicate. If you need data from a jet engine manufacturer’s 30-year production archive to train your model, you can’t just scrape the web for it. Owning the data source — and the company that generates it — is the actual competitive advantage.

Open Questions

I’m cautiously impressed, but let me flag the obvious problems.

First, the $100 billion fund doesn’t exist yet. The $6.2 billion is real. The $100 billion is what Bezos reportedly wants to raise. Those are very different things. Even for someone worth $200+ billion, convincing LPs to commit that much to an unproven AI-manufacturing thesis is a huge ask.

Physical AI is slow. Training an LLM on internet text can happen in weeks. Getting useful training data from manufacturing processes takes months or years. Prometheus needs partners willing to share extremely senstive operational data. That’s a harder sell than most VC pitch decks acknowledge.

Then there’s the patience gap. Amazon famously operated at a loss for years while building scale. Bezos can stomach that. But can his investors? The Berkshire model works when you’re buying cash-flowing businesses. If AI optimization takes years to show results, those acquired companies still need to turn a profit in the meantime.

And AI researchers move fast. Average tenure at top labs is around 18 months. Prometheus needs these people to stick around long enough to build something that works in physical domains, which takes a lot longer than shipping a chatbot feature.

How It Stacks Up Against the Competition

Prometheus isn’t the only company thinking about AI for physical systems. NVIDIA has been pushing its Omniverse platform for digital twins in manufacturing. Google DeepMind has done work on materials science and protein folding. And Musk is building Terafab, a $25 billion chip fab alongside Tesla’s Optimus robots.

But none of them are taking the “buy the whole company” approach. NVIDIA sells tools, DeepMind publishes papers, and Tesla builds robots. Prometheus wants to own the entire value chain, from data acquisition to operational deployment.

The closest comparison might actually be Amazon itself. Bezos grew Amazon from a bookstore to a logistics operation to a cloud provider. He knows horizontal expansion. Bezos knows how to build horizontally across industries. Whether he can do it again at 62, in a completely different market, is the real question.

What Happens Next

The Kyle Kosic hire suggests Prometheus is moving into its infrastructure-buildout phase. You don’t hire the person who built xAI’s supercomputer to write business plans. They’re building something big, probably a compute cluster tailored for physical simulation and domain-specific training.

Reports indicate the $100 billion fund pitch is happening now, targeting sovereign wealth funds and large institutional investors. Even if that fundraise closes at half the target, it would be one of the largest private capital raises in history.

For the broader AI industry, Prometheus represents a real divergence. Everyone else is racing to build the best general-purpose model. Bezos is betting that the real money is in vertical AI: models trained on data you can’t find on the internet, deployed in industries where a 2% efficiency gain is worth billions.

He might be right. Manufacturing, aerospace, and defense represent trillions in global GDP. If AI moves the needle even slightly in those sectors, the returns dwarf anything you’d get from $20/month chatbot subscriptions.

FAQ

What is Project Prometheus?

Project Prometheus is an AI startup co-founded by Jeff Bezos and Vikram Bajaj in November 2025. It focuses on “physical AI” — building AI systems that optimize manufacturing, aerospace, and industrial processes rather than generating text or images.

How much funding has Project Prometheus raised?

The company raised $6.2 billion in its initial round at a $30 billion valuation. Bezos is also reportedly seeking to raise a separate $100 billion fund to acquire manufacturing companies.

Who is leading Project Prometheus?

Jeff Bezos and Vikram “Vik” Bajaj serve as co-CEOs. Bajaj is a physicist and chemist who previously led programs at Google X and co-founded Foresite Labs. Kyle Kosic, an xAI co-founder, recently joined to lead infrastructure.

How does Prometheus differ from OpenAI or Anthropic?

While OpenAI and Anthropic build general-purpose language models, Prometheus focuses on AI for physical processes — manufacturing optimization, aerospace engineering, and industrial automation. It also plans to acquire companies directly.

Is Project Prometheus publicly traded?

No. Prometheus is a private company as of April 2026, though the Berkshire Hathaway-style structure suggests it may eventually go public or operate as a holding company with publicly traded subsidiaries.

Bottom Line

The $6.2 billion is already deployed. The talent pipeline is filling up. The Berkshire-style acquisition model is at least plausible on paper. What Prometheus doesn’t have yet is proof that AI can actually optimize a jet engine production line better than the engineers who’ve been doing it for decades. That’s the gap between a $30 billion valuation and a $30 billion company. I’m watching this one closely — the gap between “plausible thesis” and “working product” is where most ambitious plays die, but Bezos has closed that gap before.