Jeff Bezos is already worth 41 billion… for an AI that does not exist yet
The founder of Amazon has just raised 12 billion dollars for a start-up with no product, no customer and no revenue. But the ambition is one of the most dizzying in tech. A breakdown, without hype.
What Bezos actually announced
On 11 June 2026, Prometheus announced a raise of 12 billion dollars at a valuation of 41 billion — around 18 billion gathered in total since its discreet launch in November 2025. At the helm: Jeff Bezos, co-CEO with the scientist Vik Bajaj, his first operational role since leaving Amazon in 2021. 150 employees in San Francisco, London and Zurich, at work since late 2024.
What Prometheus is meant to become
Not a chatbot, not a robot: an “artificial general engineer”. Where ChatGPT manipulates text, Prometheus wants an AI that reasons about physics in order to accelerate the whole “invention loop”, from the idea to a manufacturable product. Bezos describes it as “a very modern version of CAD” (computer-aided design), and sums up his conviction: “all the wealth of a society comes from invention”.
The targets are colossal: aircraft engines, spacecraft, chips, automotive, medicines. Vik Bajaj takes the example of an aircraft engine — a decade and entire teams to design it, prototype it and manufacture it; the idea is to compress those cycles. The playing field being targeted: “the whole physical economy”, or around 70,000 billion dollars.
How it is supposed to work
The company says little, but the approach is taking shape. These are “world models”: AI trained not on text, but on enormous volumes of multimodal data drawn from the real world, in order to simulate the physical behaviour of objects. Concretely, the AI is meant to reconstruct the airflow around a wing, or predict where a part will fail under stress before it is even manufactured — replacing physical prototypes with digital validation. All of it backed by proprietary data on materials, tolerances and processes.
The catch, acknowledged by Bezos: there is “no Internet of manufacturing data” to hoover up. So it has to be created (synthetic data, “very compute-hungry”) and bought. As early as November 2025, Prometheus acquired General Agents, a start-up founded by alumni of DeepMind and OpenAI, whose “vision-language-action” (VLA) model translates images into actions. And according to the Financial Times, Bezos is reportedly seeking up to 100 billion dollars for a separate holding company buying up manufacturers (aerospace, semiconductors and so on) in order to capture their operational data — a concentration that is already worrying regulators, against the backdrop of a European AI Act in retreat and questions about who really controls the data.
What it can do — and what it does not do
What it is aiming at: designing, simulating, predicting and optimising, then selling those capabilities as software for simulation and design to engineers, for engines, chips, cars, spacecraft or medicines.
What it does not do: it does not manufacture anything itself — it helps with design and pre-production, not the factory; it is not a robot (“nothing to do with robotics”, Bezos insists); and it is not a conversational assistant. One troubling nuance: the VLA architecture it acquired is, elsewhere, a key building block… of robotics. So the boundary is blurrier than the pitch.
What we still do not know: the exact architecture, the data and the training method, a single performance benchmark, a single published paper, and any product date — Bezos considers that “premature”.
41 billion for a promise with no date
Here is the dizzying part: at 41 billion, Prometheus has no public product, no customer, no revenue. Bezos talks about “remarkable” progress but “piece by piece, with no sudden moment”, and places real change “on a horizon of about ten years”. So what is being paid for today is a result hoped for in 2036 — a scenario that recalls SpaceX valued at 2,100 billion for a space data center that does not exist yet.
If it works… or if it fails
If the promise holds, the impact would be historic: medicines, chips, engines or energy technologies designed in a fraction of the time they take today, a leap in productivity, and a major industrial and geopolitical advantage for whoever holds both the tool and the data — at the price of an industrial fortune concentrated as never before. It is also the promise, heard before, of an AI that would create more jobs than it destroys — when “AI” is already serving as the argument for waves of layoffs.
If it fails, the fall would be brutal. Physics is not software: certification, supply chains, material realities… the gap between an impressive demo and an industrial product remains “substantial”, specialists warn. If the thesis is slow to arrive, or if a competitor proves the capability first, it is 18 billion dollars committed that wobble — and an alarm signal for the whole AI bubble. Bezos brushes it aside: “don’t worry about that”.
The figure to remember: 10 years. That is the horizon Bezos himself gives for real change. The valuation, meanwhile, is already 41 billion — with not a single product.
Our verdict. AI promises, Le Recul checks: here the promise is one of the biggest of the decade — automating the invention of the physical world — but the proof is, to date, strictly non-existent. The bet could redefine global industry… or become one of the biggest mirages of the AI wave. Worth watching: a first product, a first customer, and the fate of the 100 billion holding company before the competition authorities.