PROJECT PROMETHEUS · FINDINGS
DOSSIER · REF-AGE/2026 · UNVERIFIED OPEN SOURCES
Artificial General Engineer · Field Brief

The CAD of
the future.

Jeff Bezos's secretive AI lab wants to automate the intellectual work of engineering — designing, testing, and helping manufacture physical objects. Two questions, answered: what is it, and what happens to the timeline of a project once a tool like this exists.

FOUNDED Nov 2025 BASE San Francisco VALUATION ~$38–41B STATUS Pre-product
01 / IDENTITY

What Project Prometheus is

The founders' own pitch is an “artificial general engineer” — and Bezos has described the product as “a very, very modern version” of CAD. Not a marketing analogy bolted on afterward; it's the thesis.

Founders
Jeff Bezos & Vik Bajaj — co-CEOs. Bezos is the largest individual backer; Bajaj is a chemist/physicist formerly of Google X.
Founded
Late 2025 · San Francisco, with offices in London and Zurich
Funding
~$6.2B launch round → ~$12B Series B (Jun 2026) at a ~$41B valuation. Among the largest early-stage AI startups ever, before any revenue.
Team
120+ researchers poached from OpenAI, Meta, DeepMind, and xAI. Acquired agentic-AI startup General Agents.
The product
AI that designs physical objects. Trained not just on text but on real experimental data, physics simulations, and engineering workflows — so it can reason about how materials and systems behave.
Focus
Aerospace · automotive · manufacturing · new materials · drug discovery
Not
Robotics. Bezos has been emphatic — “nothing to do with robotics.” It's the design brain, not the robot body.

“Is it ChatGPT / Claude Code for CAD?”

Verdict — A fair elevator pitch, as long as you remember three things: it's physics-grounded (trained on real-world data, not text), it's aimed at hard physical engineering where mistakes are expensive, and it isn't a shipping product yet. The phrase describes the ambition, not something you can use today.

02 / TIMELINE

Before & after the tool

The biggest change isn't that one step gets faster — it's that the shape of the project changes. Traditional engineering is a slow physical loop. A tool like this turns most of that loop into a fast software loop and pushes physical work to the very end.

Time to a validated design — drawn to scale1 aerospace bracket
BEFORE · 4–9 MONTHS · 3–5 ITERATIONS
AFTER · ~2 WEEKS

Same problem, same physics. The compression comes from moving trial-and-error off the test rig and into the computer.

Why each chokepoint breaks — three distinct mechanisms:

MECH·01

Search breadth

A human conceives a handful of designs and anchors on familiar shapes. Generative tools explore thousands of candidates — including non-intuitive, lattice-like geometries no one would draw by hand. You don't just go faster; you find better optima.

↓ ~50% iteration time · ↓ ~38% mass
MECH·02

Evaluation speed

The real throttle. Verifying a design used to mean hours-long FEA/CFD runs, so you tested five. AI surrogate models predict performance in milliseconds — the difference between scoring five designs and scoring ten thousand.

hours/run → milliseconds/run
MECH·03

Physical validation

Cutting metal and running a test rig is the slowest, costliest step. If simulation is fast and trusted — the reason Prometheus trains on real experimental data — you reach a prototype later, fewer times, and closer to final.

many loops → one confirmation

The same project, phase by phase:

Traditional cyclesequential · physical
  1. Hand-model 2–3 CAD concepts1–2 wk
  2. Run FEA on the leading concept1–2 wk
  3. Machine + load-test a prototype4–8 wk
  4. Find a failure → back to step 1repeat
∑ 3–5 iterations · 4–9 months · lands on “good enough”
AI-for-CAD cycleparallel · virtual
  1. Specify the problem: loads, constraints, targetday 1
  2. Generator produces hundreds, surrogate scores allhours
  3. Review trade-off frontier, confirm top picksdays 2–4
  4. One physical prototype — already near-finalweek 2
∑ thousands explored · weeks · often 30–50% lighter
03 / PROOF

It already happened next door

The bracket is the toy version. The real bet is on domains where one physical iteration costs millions and takes a year — so the payoff is bigger. Adjacent fields already show the pattern.

Protein structure · AlphaFold

Months of lab work → minutes of compute

Structure determination once meant months of crystallography per protein. A model collapsed it by orders of magnitude and predicted the lot.

200M+structures predicted
Generative design · documented case studies

Half the iteration time, a third less mass

Topology-optimization tools cut design-iteration time against conventional CAD while hitting large weight reductions at equal stiffness — today, in shipping workflows.

~50% / ~38%faster iterating · lighter parts
04 / CAVEAT

The bottleneck moves — it doesn't vanish

Trust

Certification & validation don't compress

An FAA jet-engine certification takes years no matter how fast the part was designed. For safety-critical hardware, a confidently wrong simulation is worse than a slow one — so physical validation never fully disappears. The whole bet rides on whether the model's physics can be trusted.

Make

Manufacturing reality is a constraint, not an afterthought

A beautiful generative lattice that can't be machined or printed affordably is useless. Manufacturability has to be in the objective from the first run.

Net: design exploration compresses from months to days and gets better results, while validation, certification, and manufacturing become the new long poles — roughly the same pattern as Claude Code, where writing the code got fast and testing, review, and deployment are now where the time goes.