Jeff Bezos raised $12 billion to build an artificial general engineer. The bet is that AI's largest compression opportunity is not in software, where cycles already take minutes, but in physical engineering, where a 10% thrust improvement on a jet engine is a ten-year program.
Jeff Bezos and Vikram Bajaj raised $12 billion for Prometheus on June 11 at a $41 billion valuation. The company has 150 employees. It launched in November 2025 with $6.2 billion in initial funding. The stated goal is to build an "artificial general engineer" that can take a complex physical product from concept through production.
Bajaj, who co-founded Alphabet's life sciences unit Verily and holds a professorship at Stanford's School of Medicine, described the ambition in specific terms. If you go to a jet engine manufacturer and ask for the same engine with 10% more thrust, the development program takes ten years. A new commercial jet engine family costs billions to develop. Pratt & Whitney's geared turbofan program ran to an estimated $10 billion. Military programs stretch to fifteen or twenty years. Prometheus wants to compress that cycle by a factor of ten.
The phrase Bezos uses is "the dream-build loop." An engineer imagines a design change. They simulate. They test. They revise. The speed of that loop determines how many iterations are possible. In software, the loop runs in seconds. Write code, compile, run tests. A developer working on a web application can execute hundreds of loops per day. Physical engineering runs on a different clock. Prototype fabrication takes weeks. Wind tunnel testing takes months. Metallurgical validation can take a year.
That asymmetry is the entire thesis. AI tools for software development compress a loop that already runs fast. The productivity gain is real but bounded. Reducing a five-minute compile-test cycle to thirty seconds changes the experience but does not change the economics of the industry. Reducing a ten-year jet engine program to one year changes the economics of aviation, defense, energy, and manufacturing. The compression ratio is what matters, and the compression ratio is largest where the existing cycle is slowest.
Prometheus builds AI systems that predict physical performance: stress analysis, fluid dynamics, thermal behavior, manufacturing tolerances. The approach sits closer to simulation than to generation. Traditional finite element analysis can take days of compute time for a single configuration. Neural surrogates trained on simulation data can approximate the same physics at speeds reported between one hundred and one thousand times faster. The tradeoff is accuracy. Physics does not tolerate hallucination. A turbine blade designed with a 2% error in thermal analysis will fail.
The company's investors are treating it like infrastructure. From $6.2 billion at launch to $12 billion seven months later, the capitalization trajectory resembles a semiconductor foundry or a hyperscaler data center more than a software startup. Physical-world AI has physical-world capital requirements: test facilities, materials databases, partnerships with manufacturers who provide the ground truth that training data demands.
The global computer-aided engineering market was $12.9 billion in 2025, according to Grand View Research. Ansys, the largest pure-play simulation vendor, was acquired by Synopsys for $35 billion in mid-2025. Prometheus at $41 billion is already valued above the dominant incumbent's acquisition price, with a fraction of the revenue. The premium is for the claim that neural simulation can replace, rather than supplement, the traditional solver.
Bezos has a pattern. Amazon Web Services was not the first cloud provider. It was the one that recognized the infrastructure layer would command more value than the applications running on top. Blue Origin builds rockets. Prometheus builds the tools that design them. The connective tissue across his portfolio is a conviction that platforms underneath capture more durable margin than products above.
The question the valuation embeds is whether physical engineering is compressible in the way that language and code turned out to be. Large language models work because natural language has deep statistical structure. Code works because programming languages are formal and constrained. Physical systems have structure too, governed by conservation laws and material properties. But the failure modes are different. A language model that hallucinates a citation produces embarrassment. An engineering model that hallucinates a material property produces a crash.
One hundred fifty employees, $41 billion in implied value, and a claim that a ten-year engineering cycle can become a one-year engineering cycle. Every number in that sentence is a bet on the same proposition: that the slowest loops in the economy are where AI creates the most value. The software industry spent the last three years arguing about which knowledge workers would be replaced. Prometheus is asking a different question. It is asking how fast a human engineer can think when the loop between imagining and testing drops from years to days.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)