Jeff Bezos has raised roughly $18 billion for Prometheus, the industrial AI startup he co-founded and co-leads with physicist Vik Bajaj, and staffed it with only about 250 people. The company carries a $41 billion valuation from a $12 billion Series B closed in June, and Bezos told a Semafor event on November 17 that most of the capital has gone to compute. The pitch is an "artificial general engineer" — an AI system that designs and manufactures physical products the way large language models generate text and code.
A Fast Company investigation published this week mapped the workforce. Reporter Rebecca Heilweil identified 130 of the roughly 250 employees from LinkedIn and personal research pages, and the hiring pattern reads like a technical roadmap. Prometheus did not recruit prototypical LLM researchers. It pulled engineers from aerospace, biotech, defense, chemistry, ocean and energy engineering, and high-performance computing — people from xAI, Meta Reality Labs, Waymo, Epic Games and Red Bull Advanced Technologies. Several came from General Agents, an applied AI lab Prometheus quietly acquired last year, whose former CEO Sherjil Ozair now sits on the technical staff.
What the company builds is unusually narrow in architecture for its scope. Bezos and Bajaj said at the Semafor event that Prometheus has built one single multi-physics model rather than a suite of narrow simulators, and Bezos has explicitly ruled out robotics, likening the product to a much more advanced version of CAD software. On stage he gave the compression math his own way: a small team using Prometheus tools designed a wireless device in 12 hours, work he estimated would take six to nine months by conventional methods. That figure is the company's own claim, not an independently verified benchmark.
The money trail runs from a $6.2 billion initial round in November 2025 — much of it Bezos's own capital, according to Bloomberg — through the June Series B backed by JPMorgan, BlackRock, Goldman Sachs, DST Global and Arch Venture Partners. Bloomberg estimates Bezos holds about 35 percent of the company after those rounds. Prometheus launched publicly in November 2025 after being incubated in 2024 at Foresite Labs, the venture studio run by Bajaj, a Stanford professor who previously led science operations at Alphabet's Verily.
Reported experiments hint at where the model is being pointed. The company has reportedly used neural surrogates for numerical simulation to redesign hardware, including an application to the Mars helicopter. A trademark application maps ambitions across physical simulation, digital twins, structural, thermal, fluid and acoustic analysis, 3D printing, semiconductor manufacturing and factory layouts. Semafor has reported that Bezos and Bajaj are preparing a proposed $100 billion acquisition fund for legacy manufacturers, which would let the company rebuild acquired plants around AI-native design workflows. Blue Origin, Bezos's rocket company, is already running two private projects using Prometheus technology, and the startup has retained advisers with unusual résumés: one serves as chief scientist for the US Army's ground systems research, another worked in Marine Corps innovation and at defense contractor Anduril.
The competitive field is young but crowded. Anima Anandkumar and Benedikt Jenik were approached about leading Prometheus before founding their own physical-AI startup, Accelerated Understanding, while P-1 AI and Neural Concept are building agentic engineering tools from other directions. None has anything close to Prometheus's war chest.
The open question, as skeptics put it, is whether 250 people and $18 billion of compute are enough to distill engineering itself into one model. Physical AI needs training data current models mostly lack — the language of physics, materials and tolerances rather than text — and a hallucination in a structural engineering model does not produce a buggy feature but potentially a building that fails. What the staffing page proves, regardless of outcome, is the thesis: if industrial AI is going to compress a decade-long jet engine program into a few years, it will be built by the engineers whose decade it is replacing.
Comments (0)
Log in to join the discussion
Log InNo comments yet