State of the radar
17,173 companies tracked across YC, Techstars, 500 Global, Plug and Play, SOSV and a16z speedrun; the YC read covers W24 through S26. The headline is that the AI application wave has stopped expanding and started rotating. Vertical AI agents still take 25.8% of the latest batch — by far the largest single cluster, 546 companies since W24 — but it is down 3.2pp over the last two batches, the steepest decline on the board. The share it gave up went to things with physical or regulated substrate: robotics, compute, health, defense, fintech. Meanwhile the outcome data shows how unstable the middle is: of 5,737 YC companies snapshotted in February, 546 materially rewrote their pitch in under six months, 209 were delisted and 63 were acquired.
Rising
- Robotics and physical world — 13.1% of the latest batch, +2.0pp. Founders are betting that the durable moat is hardware and deployment, not the model layer.
- Fintech — 4.4%, +1.8pp, after two years of quiet. Agents that move money need rails, custody and underwriting that nobody has built for non-human actors.
- AI infra and compute — 7.4%, +1.7pp. Compute scarcity is now a product category, not a procurement problem.
- Healthcare and bio — 5.7%, +1.7pp, and only 4 acquisitions and 5 inactive out of 442 tracked. Slow, but the least churn-prone cluster on the list.
Crowded
- Vertical AI agents. 546 companies since W24 and falling share. Any generic "AI employee for X" is now entering a market with hundreds of near-identical pitches.
- Agent infrastructure. 255 companies, 16.6% of the latest batch, and 198 of them sit against the single "Self-Maintaining APIs" RFS — the most oversupplied theme YC asked for.
- Developer tools. 155 total but only 3.5% of the latest batch, down 1.6pp. Demand is being absorbed upward into agent infra.
What happened to last season's startups
Vertical AI agents rewrote 188 pitches out of 710 tracked — roughly one in four teams changed their story in six months, and they also led on acquisitions (14) and inactives (7). Consumer, by contrast, tracked 783 companies with just 34 rewrites and 1 acquisition: less churn, less exit. Three moves worth reading:
- Closera (S25): "AI employees for commercial real estate" → "AI Shopping Agent for Your Ecommerce Store" — a vertical agent abandoning a slow-buying vertical for e-commerce volume.
- Normal (S25): "Deploying and integrating useful robots" → "Testing and certifying the hardware of the future" — from robot deployment to the picks-and-shovels of robot validation.
- Vela (W26): "The AI Scheduling Assistant that truly knows you" → "AI Recruiting Coordinator for Executive Search" — a horizontal assistant narrowing into one paid workflow.
Open windows
- The Future of American Defense — 37 companies of supply, the thinnest of any RFS, against a cluster growing 1.5pp. Defense procurement literacy is the real barrier, not the tech.
- Proving You're Human and AI-Native Compliance Infrastructure — 38 companies split across two distinct asks. Identity-of-origin and machine-readable compliance are both effectively unbuilt.
- Multiplayer AI — 41 companies. Almost every assistant built since W24 assumes one user and one agent; shared context across a team is open.
Idea candidates
- A compliance-evidence layer that emits machine-readable audit artifacts from agent runs, so regulated buyers can approve AI workflows without a manual review cycle.
- Hardware test-and-certification-as-a-service for small robotics teams shipping their first units, priced per device rather than per engagement.
- Shared-context workspaces where several people and several agents work the same thread, with provenance on who — or what — changed each artifact.