Prometheus and the New Frontier of Physical AI

So, something interesting to note. Jeff Bezos’s new AI company is co-run by a chemist.

Vik Bajaj came out of Google X, trained as a physicist and chemist, and now sits across the table from Bezos as co-CEO of Project Prometheus. They launched the company in November 2025 with $6.2 billion in funding. In April it took in another round at a $38 billion valuation and pulled in over 120 people, many poached from OpenAI and DeepMind. It’s Bezos’s first operational role since leaving Amazon in 2021.

But that main hire is the whole bet. If you think the next decade of AI is mostly text and code, then you would put a text and code person in the chair. Bezos has instead put someone there whose career has been about how matter behaves. That’s not exactly a subtle move.

He’s not the only one chasing this shift. Tesla and Nvidia are working the same territory, and Elon Musk has called Bezos a copycat for it, which is gossipy for sure, but probably at least half right.

What Prometheus is really doing is hard to find in any press release. People close to the project describe it as a “manufacturing transformation vehicle.” Bezos is reportedly raising a separate $100 billion fund to acquire industrial businesses he expects AI to disrupt, then use AI to fix their margins. The technology is half of it, but the other half is owning the assets the technology operates on. Aerospace, semiconductors, automotive, drug development. The parts of the economy where small operational shifts compound into very large numbers.

For most businesses, what this means in practice is pretty unglamorous. The shift starts with AI becoming capable of reading how their operations really work.

Companies generally underestimate how opaque their own operations are. Most operational knowledge gets carried in people’s heads rather than systems, and the bits that do get written down often go stale fast. Leaders end up running on intuition because there’s often nothing else available. When AI starts interpreting the physical layer of a business, all of that unspoken activity comes into focus fast.

Early versions of this type of physical AI are already running. In property and facilities management, AI is reading patterns in maintenance logs and energy usage, spotting the small anomalies that tend to become large invoices a few weeks later. In mining and industrial settings, vibration and heat signatures are being used to predict equipment strain well before a breakdown. No futuristic robots are marching through the door, but intelligence is getting stitched into all the data that was already there.

The businesses that get value out of this wave will mostly be the ones that already understand themselves well enough to expose their operations to scrutiny. The ones running on improv will get read too, but the reading won’t be flattering.

Bezos has built his career around watching where physical things move and putting himself in the middle of it. Books, then everything else, then rockets. Putting a chemist in the chair this time is consistent. He didn’t come out of retirement for an LLM company and he didn’t come out of retirement for a robotics company. He came back for something in the seam between the two, and he hired a chemist to help him build it. In doing so, he’s pretty clearly telling us what kind of company he thinks this is, and what kind of problem he thinks is still worth a decade of his attention.