We see a lot of AI decks. Most of them describe a company that is, underneath the language, a well-designed interface sitting on top of a model somebody else trained and somebody else pays to run.
That is not an insult. Some of those businesses will work. But they are application companies, and they should be priced, funded and defended as application companies — not presented as deep tech because the underlying model happens to be sophisticated.
The question that separates them
There is one question we ask in every AI meeting, and the answer usually settles it:
What happens to you if the model underneath gets ten times better, ten times cheaper, and available to everyone on the same day?
For a genuine deep tech company, the answer is we get better too — the improvement flows into a product whose advantage lives somewhere the model cannot reach. For a wrapper, the honest answer is our product becomes a feature and our margin goes to zero, though it is rarely phrased that way.
Where the advantage actually lives
In the companies we find defensible, the moat sits in one of three places, and none of them is the model:
- Proprietary data. Not "we will accumulate data as we grow" — data that already exists, that took something unusual to assemble, and that a competitor with more money cannot simply buy.
- Hardware. If inference has to happen on-device, at low power, in a factory or a vehicle or a medical implant, the hard problem is a hardware problem. Very few software teams can follow you there.
- Domain knowledge. Sometimes the moat is that the founders spent fifteen years inside an industry and understand a workflow that is genuinely opaque from the outside. This one is real but decays fastest, so we want to see it paired with something else.
The uncomfortable part
The uncomfortable version of this argument is that a lot of the current AI cycle is being funded on the assumption that speed of execution is a moat. It is not. It is a head start, which is a different asset with a much shorter half-life.
We would rather back a company where the hard part is visibly, expensively hard — where you can point at the thing that took four years and say that. Those companies are slower to build and much harder to catch.