Jeff Bezos's $6.2B 'Physical AI' Startup Poaches the Engineer Who Built xAI's Supercomputer
Project Prometheus, the physical AI venture co-founded by Jeff Bezos, hired Kyle Kosic — one of xAI's founding engineers and the architect of Grok's Colossus supercomputer — to lead large-scale infrastructure.
Project Prometheus hired Kyle Kosic on April 7, 2026. That hire matters more than a typical executive poach. Kosic was one of xAI’s original 11 employees, came to xAI from OpenAI, and was the primary engineer behind Colossus — the supercomputer in Memphis, Tennessee that powers Grok. If you want to build AI at physical scale, he’s exactly who you recruit.
What Project Prometheus actually is
Jeff Bezos co-founded Project Prometheus with physicist Vik Bajaj, a former Google X researcher. Bezos is actively involved — his first return to an executive operating role since stepping down as Amazon CEO in 2021. The company has raised $6.2 billion and currently employs more than 120 people across San Francisco, London, and Zurich.
The pitch is “physical AI”: systems trained on scientific and physical-world experimental data rather than text. The target verticals are manufacturing, aerospace, automotive, and architecture — industries where AI has to reason about real physics, materials, and spatial constraints, not just language.
This positions Prometheus squarely against the text-heavy incumbents. The company is explicitly not trying to build another ChatGPT competitor. It’s betting that the most defensible AI will be trained on proprietary domain-specific data tied to physical processes.
Talent is pouring in from OpenAI, DeepMind, Meta, and Google in addition to xAI. Kosic’s hire is the highest-profile addition yet, partly because of his specific background: building the infrastructure that makes large-scale AI training possible at all.
The xAI exodus context
xAI’s cofounder departures have been well-documented. All 11 original cofounders have now left Elon Musk’s AI company. Kosic’s move adds a pointed narrative — the people who built xAI’s most impressive technical infrastructure are now at Bezos’s company.
That’s meaningful for a startup that needs to build massive compute infrastructure quickly. Kosic doesn’t bring vision documents. He brings the actual experience of standing up tens of thousands of GPUs and making them train reliably at scale.
Why Bezos is serious this time
Bezos has invested in AI before — he’s backed Anthropic personally and through Amazon. Project Prometheus is different. He’s not writing checks; he’s co-leading the company day-to-day alongside Bajaj. That’s a significant commitment of attention from someone who doesn’t need to work.
Reports from April also indicate Bezos is in early conversations to raise a separate $100 billion fund to acquire and automate manufacturing companies — essentially an industrial AI private equity vehicle running alongside Prometheus. If that fund materializes, Prometheus has a natural distribution channel into the physical industries it’s targeting.
The combination of Kosic’s infrastructure expertise, Bajaj’s scientific domain background, Bezos’s capital and operator experience, and a $6.2 billion war chest makes this one of the few AI startups with the credibility to challenge the incumbents in a genuinely different direction.
Physical AI is a harder problem than language AI. It requires proprietary data pipelines, custom training infrastructure, and domain expertise that can’t be bought on the open market. That’s the moat, if it holds.