Every infrastructure cycle ends the same way. Early on, value sits with whoever wraps the new capability in something usable. Then the capability gets cheap, the wrapper gets copied, and value migrates to whoever owns the layer underneath. Rails, not the app on top of them. AI is now at that turn.
A more capable model is a stress test. It makes the durable company stronger and the wrapper obsolete, on the same day.
So we underwrite every company against one question. When the next, more capable model ships, does this company get stronger, or get erased?
A company that gets stronger owns something the model cannot recreate. Proprietary data no one else has. A system of record the workflow already runs on. A regulatory license that took years to earn. Distribution into a market that trusts it. Hand that company a better model and it compounds. The moat was never the model.
A company that gets erased is a wrapper. It packages a capability the next release will simply include. The traction is real, the demo is good, and none of it matters, because the thing it sells is about to ship for free inside something more capable. That line rarely announces itself. Finding which side a company sits on, before the price reflects it, is the whole job.
This is why we back three layers specifically. Vertical Intelligence, where the AI becomes the regulated function rather than a tool sold to it. Physical AI, the simulation and data layer every autonomous machine depends on. Agentic Commerce, the identity and settlement rails every agent transaction must clear. In each, a stronger model is a tailwind, not a threat.
It is a narrow filter on purpose. Most of what looks exciting in AI right now is a wrapper with good months ahead of it and a hard ceiling behind that. We would rather own less of what lasts.
Tell us what you are building. Every note gets a real read.