Jeff Bezos's Project Prometheus Recruits xAI Co-Founder to Build Industrial AI Systems

Jeff Bezos Taps xAI Co-Founder for Secretive AI Startup

Kyle Kosic, an xAI co-founder who helped build the infrastructure behind Elon Musk's Colossus supercomputer, has joined Jeff Bezos's Project Prometheus, marking another high-profile move in the fierce competition for top AI talent1

1

. Kosic departed OpenAI to join the secretive AI startup, where he will continue working on AI infrastructure projects alongside Bezos and co-founder Vikram Bajaj, a former Google executive who also co-founded life sciences venture Verily4

4

.

The recruitment of top AI talent has intensified as labs compete with substantial salaries to poach engineers and researchers from rivals. Kosic's journey reflects this volatility: he joined OpenAI in 2021 as technical staff, left in 2023 to co-found xAI with Musk, led infrastructure development there, then returned to OpenAI in 2024 before now joining Project Prometheus3

3

. Meanwhile, all 11 of Musk's xAI co-founders have now departed, with the last two leaving at the end of March amid complaints about Musk's management.

Industrial-Focused AI Systems Target Physical World Applications

Project Prometheus aims to create AI systems that operate in the physical world, moving beyond the large language models that power chatbots like ChatGPT or coding tools. The company focuses specifically on the industrial sector, envisioning models that understand the laws of physics and are trained on domain-specific data such as jet engine design. According to sources close to the company, Project Prometheus has already assembled the largest corpus of data on engineering and how such systems work.

The startup has hired hundreds of staff across its San Francisco headquarters and offices in London and Zurich, targeting engineers, AI researchers and specialists with experience building out massive infrastructure projects2

2

. The team includes alumni from OpenAI, Google DeepMind, Meta, Microsoft, Anthropic, Nvidia and Grammarly2

2

4

4

. The company raised $6.2 billion last year and now seeks another $6 billion to develop AI systems for real-world industries4

4

.

Ambitious Acquisition Strategy Targets Manufacturing and Aerospace

Bezos and Bajaj are personally leading efforts to raise tens of billions of dollars or more for a "permanent capital vehicle" that would acquire equity stakes in companies likely to be disrupted by AI. One person compared the structure to a "Berkshire Hathaway-type holding company". The startup plans to amass stakes in companies across sectors including engineering, aviation, architecture, design, semiconductors, defense and aerospace4

4

.

These acquisitions would serve a dual purpose: gathering training data from these companies to improve the startup's AI model while deploying "forward deployed engineers" to work within the acquired firms. The investment and staff input aims to improve margins and operations at these companies. "Prometheus wants to back the progress of these industries, which will happen eventually with AI, but they don't want it to take 10 years," one source explained. The startup is discussing investments with sovereign investment funds from Singapore and Gulf nations2

2

.

Competition Heats Up in Physical AI Sector

Project Prometheus marks Jeff Bezos's first operational position since stepping down as Amazon's CEO in 2021, though he remains founder of Blue Origin4

4

. The venture faces stiff competition in the emerging physical AI space. Rivals include AMI Labs, a Paris-based startup led by former Meta executive Yann LeCun that raised $1 billion in Europe's largest-ever seed round; World Labs, a Fei-Fei Li-led venture focused on world models that raised $1 billion in February; and Periodic Labs, which merged AI with physical sciences and counted Bezos among its backers during a $300 million seed round4

4

.

The AI industry has struggled to create models that truly understand physical space due to a lack of high-quality data representing the real world, rather than more readily available text and computer code. Competitors' current efforts involve training on video data and simulations to mimic real-world environments. Global venture funding for physical AI reached $26.7 billion in the first two months of 2026 alone, according to Crunchbase data, and is expected to jump to $33.5 billion by year's end—nearly double the total from 20254

4

.