Six months ago we sat down to redo our thesis for the next two vintages. The short version: most general-purpose AI bets are now too crowded for our cheque size, and the interesting risk-adjusted returns are sitting in vertical AI — companies that take a foundation model and turn it into the operating system for one specific industry.
Here's what changed our minds.
The capability ceiling stopped mattering
Models from the major labs are good enough at most language tasks that further capability gains rarely move revenue for an applied startup. What moves revenue is everything around the model: the workflow it slots into, the data it learns from in production, the integrations that make it sticky. Those are domain problems, not ML problems. Founders who know an industry inside-out have an advantage no GPU budget can match.
Incumbents are slow to ship and faster to lose
Across legal, healthcare admin, manufacturing scheduling, and parts of financial back-office, we're seeing 18-month-old startups winning seven-figure contracts against vendors that have been in market for two decades. The incumbents have distribution but not product velocity. The pattern repeats: pilot with one team, expand horizontally because the agent already knows the workflow, then displace.
The unit economics finally pencil out
Inference costs for the workloads we care about have dropped roughly 70% in the last 12 months. A year ago, 30% gross margin on AI-heavy SaaS felt like a ceiling. Today we're underwriting 65–75% gross margins at scale on the right architectures. That changes which companies are venture-backable.
What we screen for in a first meeting
- The founder has run this workflow themselves. Not advised on it, not consulted to it — done it. The fastest companies in our portfolio are run by people who used to do the job they're now automating.
- A data loop the incumbent can't replicate. The product gets measurably better the more customers use it, in a way that compounds. A wrapper around someone else's API doesn't qualify.
- Sales motion already proven once. Even at pre-seed, we want to see one pilot that closed without the founder being in the room. If you can't sell it without you, you don't have a product yet.
- Time-to-value under 30 days. Long implementation cycles kill velocity at this stage. The companies growing fastest in our book deploy inside two weeks.
We've made four investments under this thesis since January and have term sheets out on two more. If you're building something that fits, our inbox is open.