A thesis that only says “AI” is a weather report. Everyone already has one. We need a map of where intelligence stops being software and starts being a system.

Here is ours. Five terrains. Same operating system in each: find the edge, make it real, build the team, compound.

Compute

The model is not the company. The constraint around the model is.

Power, packaging, custom silicon, compilers, the ugly software that makes a new chip usable by a small team. We look for people who treat compute as a scarce physical good, not an API credit. If your advantage disappears when the frontier lab ships a cheaper model, you are renting someone else’s thesis.

At zero: a belief about where the bottleneck actually is, and a prototype that proves you can move it. At one: a customer who will wait in line because the alternative is worse.

Intelligence

This is the software layer of the same story. Products that act, adapt, and compound. Agents in the critical path of value. Tools that give a formidable team the surface area of a much larger one.

We do not fund “a copilot for X” unless X is a workflow deep enough that the model cannot be peeled off. Durable companies here look like systems of record that happen to think, or networks that get smarter because they are used.

At zero: a non-obvious workflow and a demo that does the job. At one: a loop that is already teaching the product something the public model does not know.

Biology

Biology is becoming an engineering discipline whether the industry is ready or not. The interesting companies turn measurement and design into software: sequencing, spatial omics, closed-loop experiments, intervention that can be iterated.

The tell is a data flywheel that gets more valuable with every sample or every assay, not a model trained once on a public dataset. We are not here for a chatbot that summarizes papers.

At zero: a measurement or design primitive that was painful and is about to be cheap. At one: a lab or clinic that has changed how it works because of you.

Energy and matter

Intelligence that does not touch watts, grams, or miles is leaving the largest markets on the table. Generation, storage, fuels, materials, logistics. The physical plant.

Most of these companies look too hard for a software fund and too early for an infrastructure fund. That gap is the point. We will not underwrite a plant at pre-seed. We will underwrite the team that can make the next plant obvious.

At zero: a mechanism, not a slogan. At one: a path to atoms that does not require a miracle and a ministry on day one.

Control

The last mile. Robots, vehicles, drones, industrial autonomy. A model that cannot act is a briefing.

Control companies die from integration, not from model quality. We look for teams who have already touched the messy interface: hardware, regulation, a real environment that does not reset. Taste in what to automate first matters more than a general robot story.

At zero: one job, one environment, one machine that works. At one: a second customer who wants the same job done.

How we use the map

We are not collecting one company per box. We are refusing companies that have nowhere on this map. If the product can live forever as a thin interface, it belongs to someone else.

If you are on this map and still early enough that the logo is not the point, we should talk.