The Internet of Manufacturing Data Does Not Exist. Bezos Is Paying to Build It.

Noah Ogbi10 min readUpdated Aug 30, 2026

Tips, corrections, or questions? support@omniscient.media

TopicsCompute Economics Industry Strategy

CompaniesAmazon

What Bezos says he is building, and what he says he is not

Bezos describes the goal as an artificial general engineer: a set of tools that compress the loop from idea to manufactured object. "This is an age-old dream," he told CNBC, "the idea that you might build a set of tools that could actually do engineering, an artificial engineer," a dream people "have thought about for decades, but it's never really been possible."^[1] The examples are concrete. A jet engine that takes a thousand engineers ten years to design and build, done in one. A lithography toolmaker like ASML iterating its next machine faster. Chips, batteries, solar cells, drug compounds.^[1]

He is equally specific about what Prometheus is not. Asked why he had chosen to speak publicly at all, Bezos said the company had been quiet and the vacuum had filled with wrong guesses. "We heard, for example, that people thought we were building robots and doing world models and things like this, which isn't correct," he said. "We're not being secretive, we're just being heads down and trying to do the work." The alternative, as he put it: "If you'd let that just be a complete void, they'll fill it with nonsense."^[1] Bajaj drew the same line around the nature of the work. "You don't build a bridge or a jet engine through words, even the words that relate to mathematics," he said. "It's about shapes, assemblies, and to understand how something like that works inside the object, you have to understand the physics, which is very complicated, three-dimensional forces and fields and how they change over time."^[1]

That last sentence is the whole company in one breath. The thing Prometheus wants to model is not language about engineering. It is engineering itself, the physics of how a real object behaves, a different substance from the one every large language model has been trained on.

The corpus the labs used was free. This one is not.

The large language model labs were handed their training data. The text of the internet already existed, sat in public, and could be scraped at the marginal cost of bandwidth. GitHub supplied code, Common Crawl supplied prose, and the supply was, for practical purposes, infinite. Intelligence over text got built first because the corpus over text was lying on the ground.

The physical economy produces almost none of that. Faber put the word to Bezos directly, asking whether the data gave the company a moat, and Bezos neither reached for it nor refused it. "It's very differentiated," he said, "because you just need - you need access to that kind of data to be able to build a model like this."^[1] The distinction he drew is the operative one. The language models "have been trained on this giant corpus of humanity's knowledge that was already preexisting on the internet," he said. "What we're doing is very different because we have to create our data sets. We have to create our data sets and access data sets that are very hard to access. And so even the training data is completely different from what the LLMs that you're accustomed to have access to."^[1]

Bajaj said the same thing as a number. Prometheus operates in a physical economy that he put at $70 trillion, roughly 60 percent of world output, and that economy has nothing like Common Crawl or GitHub at its center.^[1] Speaking to Axios the same day, the founders would not say how the company trains its models, except to concede the precise gap: there is no "Internet of manufacturing data" for them to ingest.^[3] How a forging die wears over ten thousand cycles, how an alloy fatigues, how a process drifts on a real line, these are measured inside operating companies, held on sensors and in unstructured engineering documents, and never published. The data is not scarce because no one has scraped it. It is scarce because making it requires running the physical process that generates it.

The money is the entry fee to manufacture a dataset

Read against that gap, the $18 billion stops looking like a bet on an algorithm and starts looking like the cost of building an asset that cannot be downloaded. Bezos was blunt that the capital goes into the two expensive things a physical model needs. "This is a capital-intensive startup, there's no question about that," he said, citing the cost of compute and of building the specialized training data the company requires.^[4] A large share of the raise goes to compute, he said, and a separate large share to creating the datasets, which he called "also a very expensive investment."^[1]

This reframes the simulation work Bezos gestured at, the only progress he would claim. To learn the physics of manufacturing from first principles you simulate it, and high-fidelity simulation across "thousands of kinds of manufacturing processes," in his phrase, is among the most compute-hungry workloads in existence.^[1] The compute and the data are not two line items but one project: spend on compute to generate, by simulation and instrumentation, the proprietary physical corpus the open internet never produced. The eighteen billion buys the means of producing a dataset, not a model trained on one that already exists. Worth naming, too: this is the ideal story for a company raising eighteen billion dollars. A moat only capital and time can cross is at once a fair reading of the physical economy and the most flattering justification a founder could give for a check this size, and both can be true at once.

Why the checks came from banks, not just venture funds

The shape of the capital follows from the shape of the asset. The Series B investors are JPMorgan, BlackRock, Goldman Sachs, DST Global, and ARCH Venture Partners, with Bezos again participating after having been the largest backer of the launch round.^[3]^[4] Two of those names, DST Global and ARCH, are venture and growth investors of the usual kind; the tell is the other three, a bank, an asset manager, and an investment bank, balance-sheet institutions that did not write the early checks into the language model labs. A venture fund underwrites a software company that reaches the whole market through an API at near-zero marginal cost. Institutions underwrite capital-heavy, asset-backed industrial programs that pay out over decades, and a company whose central asset is a physically generated dataset, built with billions in compute over years of simulation and instrumentation, is closer to the second than the first. Bezos all but said so when asked about future raises: he has always looked, he said, for "investors who are interested in manufacturing and improving the manufacturing world."^[1] The capital that would build the physical-AI corpus and the capital that owns physical assets are pooling in the same institutions.

The reported next step, and why it would be consistent

If the constraint is that physical data lives inside operating companies and cannot be scraped, the most direct way to obtain it is to own the companies. That is precisely what the financial press has reported, and precisely what Bezos would not confirm. The Wall Street Journal reported that Bezos has sought investors for a roughly $100 billion fund to acquire industrial businesses and automate them, with project documents describing it as a "manufacturing transformation vehicle" aimed at sectors including chipmaking, defense, and aerospace; The New York Times reported that the fund would operate alongside Prometheus.^[5] Axios characterized the same reported effort as a vehicle to "buy legacy industrial companies that would then feed data into Prometheus."^[3]

None of this is confirmed, and the company went out of its way not to confirm it. Axios reports that Bezos and Bajaj "declined to discuss" the reported $100 billion holding company.^[3] On CNBC, Faber pressed twice, the second time proposing the private-equity framing outright. Bezos allowed only that "we may buy parts of companies and so on who could benefit from this technology, and then help them improve their processes," said the plan was "still in work," and would go no further: "It's premature to say too much about that."^[1] So the roll-up should be read as reported and unconfirmed, not as fact. It is noted here only because of how cleanly it would follow from the data problem the founders do confirm: if the corpus must be manufactured and the rest is locked inside incumbents, buying the incumbents is the logical continuation. That is a reason to watch for it, not evidence that it is settled.

Industry