Artificial intelligence has models of the physical world, of language, of code, of proteins, of weather. It has no model of the computational world in which it runs. That is the last unmodeled environment, and it is the one AI lives in.

Every model that learned to predict an environment eventually learned to design one. Weather models now site the turbines. Protein models now write the protein. A model of compute will specify the compute.

We are building it.

A chip is only as useful as the team available to operate it, which means most of the compute humanity has built cannot be reached by the work that needs it.

Our approach is to model that environment directly: learned action-conditioned dynamics across millisecond, second, minute, and facility timescales, constrained planning under hard limits, and an independent safety boundary. We deploy these as AI compute engineers, agents that predict what a placement will do before it touches production, and that state what the next machine needs when nothing in the fleet fits.

Here is what that makes possible. A model is deployed once and runs on whatever silicon is available. Trying new hardware costs what the traffic costs instead of a year of porting. Stranded capacity becomes reachable. An institution builds frontier AI on the chips it can obtain rather than the chips one supply chain permits. And hardware competes on physics rather than on how many years of software a company could afford to fund.

Every piece of work placed on a machine returns two things: the result, and a measurement of what that machine does under that work. Nobody has ever kept the measurements. We keep them. Enough of them, across enough unlike silicon, and the record states what the work actually needs: a rack mix, a memory hierarchy, an interconnect, a power envelope, in units a designer can build to and a buyer can hold a vendor to. Machines built to that specification carry traffic from the day they ship. Their traffic feeds the record. Each turn of the loop is faster than the last.

We sit at one point on that loop and nowhere else: the model that reads the record and writes the specification. We own no silicon and hold no design seat. We build the Compute World Models that make this real.

We are a small group of engineers and researchers working on a short list of problems that decide whether compute can be modeled at all: whether a learned representation of one machine transfers to a machine it has never run on, what it costs to buy that knowledge in production traffic, and whether a record of what work does on existing silicon can state what the next silicon needs. If this sounds interesting, we would love to hear from you.

Open positions