I — The substrate bet
Every platform shift looks, in real time, like an applications boom. The browser shipped and people built portals. The cloud shipped and people built SaaS. The smartphone shipped and people built apps. In each case, the loud money flowed to the surface. The durable money formed underneath.
We think the AI shift is no different. The headline companies will be the ones you can see — the chat windows, the agents, the copilots. The compounding companies will be the ones you don’t. The retrieval engine. The eval harness. The inference router. The data layer. The pieces every AI company will quietly write a check to, every month, forever.
The application layer wins the press cycle. The substrate layer wins the decade.
II — The five layers
When we map the AI stack as a venture surface, five layers consistently produce companies we want to own:
- Inference economics. Routing, caching, batching, distillation. The companies turning expensive tokens into cheap ones.
- Retrieval & memory. Vector stores are commodity; relevance is not. The next decade is fought on recall and ranking.
- Evaluation. If you can’t measure it, you can’t ship it. Every AI company eventually buys an eval product they wish they’d built.
- Data pipelines. Synthetic, curated, labeled, decontaminated. The boring layer that determines model quality.
- Runtime primitives. Sandboxes, tool calling, browser automation, codegen runtimes. The new POSIX.
III — Where we’re probably wrong
A thesis without falsifiable claims is just a vibe. Here’s where we could be wrong, and what we’d update on:
- If foundation labs absorb the substrate vertically, the layers we love become features, not companies.
- If inference cost collapses by 100× in 24 months, several of our portfolio bets become harder to defend.
- If open-source models reach parity at small sizes, the data and eval layers re-shape dramatically.
We size positions accordingly. Concentrated, not crowded. We’d rather own real percentages of fewer companies than spray and pray across a thesis we don’t actually believe.
IV — What we’re looking for
A working prototype. A benchmark we haven’t seen. A weird research direction. A founder who can explain the FLOP budget on a napkin.
We don’t need a deck. We don’t need traction. We need a credible answer to why this team, why this layer, why now — and a willingness to be wrong in public about the next two years.
If that’s you, send us something strange. We answer every pitch within 48 hours.