Prometheus Is 150 Engineers. LinkedIn Only Shows 120. Source the Gap.

On June 11, 2026, Jeff Bezos confirmed Prometheus raised $12B at a $41B valuation to build what Vik Bajaj calls an "artificial general engineer." The team is roughly 150 people across San Francisco, London, and Zurich. Prometheus's LinkedIn page lists about 120. That gap is not a rounding error. It is the entire thesis of physical AI sourcing.

If you are running a standard Boolean pass on "ex-OpenAI" or "ex-DeepMind" to build your Prometheus competitor shortlist, you are chasing the plumbing and missing the payload. The scarce hires here are not LLM infra people. They are the ones who can hold CFD in one hand and PyTorch in the other, and they have never been indexed under a title you would think to search.

The frontier-lab poach is a decoy

The headline hires are real. Kyle Kosic, one of xAI's first 11 employees and a key builder behind Colossus, is at Prometheus. Sherjil Ozair, ex-DeepMind, joined via the General Agents acquisition to bring video-language-action models into the stack. GeekWire reports the ~120 public hires came largely from OpenAI, DeepMind, Meta, and xAI. That is what gets written up.

But read Bezos's own framing. He described the product as "a very, very modern version of CAD." Bajaj told WSJ the target is AI that assists "end to end" through design, prototyping, performance analysis, and manufacturing. TechFundingNews lists the verticals: jet engines, medical devices, semiconductors, advanced materials, consumer electronics.

None of those verticals ship without simulation engineers. And simulation engineers are not on the ex-OpenAI list.

What the real target profile looks like

The nearest public analog to Prometheus is PhysicsX, the London startup backed by Atomico, Nvidia, Temasek, and Siemens. PhysicsX has raised over $155M to compress engineering simulations from days to minutes. Their own team page describes their 35 simulation engineers this way: experience across "aerospace, automotive, energy, and semiconductors, bringing expertise in CFD, FEA, electromagnetics, chemistry, material science." Their job specs require candidates to "set up CAE simulations independently" AND collaborate with ML researchers on foundation models for engineering.

That Venn diagram is the entire game. Physical AI companies need people who spent eight years running ANSYS Fluent on turbine blades and then, somewhere along the way, learned to write a CUDA kernel. Or a materials PhD who ships PyTorch. Or a Formula One aero engineer who moonlit on a Kaggle sim-to-real challenge.

No bootcamp produces this person. No CS pipeline produces this person. And no LinkedIn title tag flags them.

3.2x

Why LinkedIn undercounts Prometheus by 20%

The 150-vs-120 delta at Prometheus is the specific number to remember. About 30 people, 20% of the team, do not surface in a LinkedIn company-page count. That is not because Prometheus is running a stealth op. It is because senior aerospace, propulsion, and materials engineers do not curate LinkedIn the way a Series B SaaS engineer in SoMa does.

If you have ever tried to source a Rolls-Royce combustion engineer or a JPL propulsion lead, you already know this. Half of them have profiles that stopped updating in 2019. The other half list "Engineer" at "Rolls-Royce plc" with no team, no project, no keywords, and no skills tagged. Ansys and Dassault Systèmes power users are worse. The strongest CFD practitioner in your target city might have a two-line profile and a name you cannot search for.

A pure LinkedIn Boolean sweep on physical AI will systematically miss the highest-leverage hires on the market. This is why we built Refolk: you describe the person in plain English, across GitHub, LinkedIn, and the open web, and get a ranked shortlist that includes the profiles a Boolean would drop on the floor. When we ran a test query for US, UK, and Switzerland-based profiles combining PyTorch skills with CAD, simulation, or aerospace keywords, exact matches returned essentially zero. The candidates exist. They are just not indexed under any title or skill tag that a recruiter would type.

You cannot overpay a candidate you never surfaced. Bezos-level packages do not fix an identification problem.

The actual sourcing lanes for Project Prometheus engineers

Forget "ex-Meta AI" for a second. Here is where the real artificial general engineer talent is sitting right now.

Formula One aerodynamics groups

PhysicsX was co-founded by a former chief scientist for championship-winning F1 teams. That is not a coincidence. Mercedes AMG Petronas, Red Bull Racing, McLaren Applied, and Ferrari's Maranello simulation groups run some of the most sophisticated CFD in the world under absurd time constraints. F1 engineers are used to shipping simulation-driven decisions on a weekend cadence. Half of them can code. A meaningful fraction have quietly moved to PyTorch on the side. Almost none of them show up when you search "AI engineer."

Rolls-Royce, Blue Origin, and propulsion R&D

Bezos's own Blue Origin propulsion team is a target-rich environment for Prometheus's jet engine vertical. Rolls-Royce's UK R&D houses in Derby and Bristol have a decade-plus roster of combustion, turbomachinery, and materials specialists. GE Aerospace's Niskayuna research center is the same. These teams shipped the papers that current physics-informed neural network research is now trying to replicate.

National labs and European technical universities

JPL, Sandia, Fraunhofer institutes, ETH Zurich, EPFL, and Imperial College London mechanical engineering are producing exactly the hybrid PhDs Prometheus is hiring in Zurich and London. The Zurich office location is not accidental. ETH's robotics, materials, and applied math labs sit five minutes away.

CAE vendor R&D

Ansys, Dassault Systèmes, Siemens Digital Industries, and Altair employ the people who build the tools everyone else uses. Their R&D groups are quietly full of ML-curious simulation engineers who watched Nvidia release Modulus and PhysicsNeMo and thought "I should be doing that." They are.

DeepMind's materials and AlphaFold alumni

The GNoME team at DeepMind published on 2.2 million new crystal structures. Those researchers are already thinking in the physical-AI frame. Same for the AlphaFold diaspora now scattered across biotech.

This is a specific list of specific groups. Sourcing ML infrastructure engineers from these pools requires more than a title filter, because the titles do not agree with each other. "Senior Research Engineer" at DeepMind, "Principal CAE Analyst" at Rolls-Royce, and "Aerodynamicist" at Red Bull Racing describe candidates who could sit next to each other at Prometheus. No standard ATS taxonomy will surface them together.

What "artificial general engineer" tells the market

When Bezos and Bajaj said the phrase out loud in June, they handed every competitor the exact target. Expect a wave of BlackRock, Nvidia, or sovereign-backed clones inside twelve months. PhysicsX will accelerate hiring. Nvidia's own PhysicsNeMo team will grow. At least two frontier labs will spin up physical AI groups. Every one of them will chase the same sub-thousand-person global pool.

The recruiters who win in this window are not the ones with the tightest outreach cadence. They are the ones who identified the pool first. Aerospace AI recruiting is not a specialty yet. In eighteen months it will be. Build the map now, while your competitors are still copy-pasting "ex-OpenAI" into Recruiter Lite.

The mechanical part of this is straightforward. Pull the last five years of AIAA, SIGGRAPH, NeurIPS ML4PS workshop, and Journal of Computational Physics author lists. Cross-reference against active GitHub accounts. Layer on employer signal from the CAE vendors and OEM R&D groups above. You now have a shortlist that no LinkedIn Recruiter seat can produce.

This is exactly the workflow Refolk collapses into a single plain English query. Instead of maintaining seven separate Boolean strings across three tools, you describe the hybrid profile once ("materials science PhD who ships CUDA kernels, ideally with turbomachinery or semiconductor exposure") and the system returns candidates whose signal lives across GitHub commits, conference authorship, and LinkedIn history, weighted together.

The Kosic template versus the Bajaj template

Two hires define the two lanes at Prometheus.

Kyle Kosic is the compute plumbing lane. xAI cofounder, Colossus infrastructure, the kind of hire every AI startup wants and roughly forty companies are actively poaching. This is a hard search but a known one. You know the schools, the labs, the alumni networks. Your existing sourcing ML infrastructure engineers playbook mostly works.

Vik Bajaj is the other lane. PhD physical chemistry from MIT. Co-founder of Verily. The archetype of the physics-plus-biology-plus-org-building hybrid. There is no LinkedIn filter for "Bajaj-shaped." You surface these people by reading their papers, watching what they star on GitHub, and noticing who Bajaj himself follows.

The Prometheus org chart at 150 will end up roughly split between these two archetypes. Your pipeline needs to be too. If your current shortlist skews 90% Kosic and 10% Bajaj, you are going to lose every Bajaj-lane search you run for the next three years.

FAQ

Who is actually hiring for physical AI right now besides Prometheus?

PhysicsX in London is the closest public analog and is actively growing past 35 simulation engineers. Nvidia's Modulus and PhysicsNeMo group is expanding. Applied Materials, Siemens, and the sovereign-adjacent European industrial AI plays are all quietly building. Expect two to three new well-funded entrants inside twelve months once "artificial general engineer" becomes a category rather than a Bezos phrase.

How do I find the 20% of Prometheus hires who are invisible on LinkedIn?

Cross-source is the only workable approach. Conference author lists (AIAA, NeurIPS physics workshops, SIGGRAPH), GitHub activity on physics-informed ML repos, and OEM patent filings all name people whose LinkedIn is either stale or nonexistent. Refolk's query layer is built to blend those signals, but even a manual version of this workflow will out-produce Recruiter Lite for aerospace AI recruiting.

What does the actual JD look like for an artificial general engineer hire?

Read PhysicsX's public job specs as a template. They require independent setup of CAE simulations (CFD, FEA, or electromagnetics) AND collaboration with ML researchers on foundation models for engineering. Prometheus's roles will be similar, layered with whichever vertical the team is targeting that quarter (jet engines, semiconductors, medical devices, materials, consumer electronics).

Is the compensation gap really the bottleneck?

No. The bottleneck is identification. With a global 3.2-to-1 gap between open AI roles and qualified candidates, and the physical AI subset being a sliver of that 518,000, the winning recruiters in 2026 are the ones who can turn a plain-English brief into a real shortlist. You cannot overpay a candidate you never surfaced, and Bezos-level packages do not fix a Boolean that returns zero.