Jeff Bezos is building a two part AI strategy through Project Prometheus: a $10B AI lab at $38B valuation focused on physical AI, and a separate $100B vehicle modeled on Berkshire Hathaway to acquire industrial manufacturers whose operational data will train the AI systems.
Jeff Bezos has two bets on artificial intelligence, and the larger one does not build AI models at all.
BlackRock and JPMorgan confirmed on Monday they are investors in Project Prometheus, the five-month-old lab developing AI systems that interact with the physical world — robots, drones, manufacturing equipment — rather than generating text or images. The institutional names are the confirmation that matters: two of the largest asset managers in the world are not buying a software play. They are buying a hardware play.
Separately, Bezos is raising up to $100 billion for what investor documents describe as a Berkshire Hathaway-style vehicle to acquire industrial companies — chipmakers, aerospace manufacturers, defense contractors — whose operational data would feed training runs for Prometheus, per Reuters, citing the Wall Street Journal. No capital is committed yet. Prometheus builds the AI; the acquisition vehicle secures the factories that run it. The $38 billion valuation attached to the current $10 billion raise is the price, not the architecture.
The talent makes the strategy legible. Prometheus hired Kyle Kosic this month — he built the Colossus supercomputer at xAI, the largest GPU cluster Elon Musk has assembled to date. It acquired General Agents, a startup co-founded by former DeepMind researcher Sherjil Ozair, per Tech Funding News. David Limp, the CEO of Blue Origin, sits on the board, per Reuters. Prometheus has no disclosed product, no published model, no named commercial partners.
The closest competitor running the same physical AI thesis without the valuation premium is Periodic Labs, founded by William Fedus, who led OpenAI's reasoning models before leaving. Periodic Labs raised $300 million in seed from a16z, per TechCrunch. No Berkshire-style acquisition vehicle, no $38 billion valuation. Fedus took a slower, cheaper path.
Physical AI for industrial use remains unproven outside narrow applications, analysts at AInvest have noted. The question is whether Bezos can compress a decade of industrial AI development into months by throwing capital at the training data problem — and whether the factories he needs to buy will even sell.