State of the radar
20,012 companies tracked across YC, Techstars, 500 Global, Plug and Play, SOSV and a16z speedrun, with ten YC batches from W24 to S26. The headline is that the AI application wave has stopped expanding: Vertical AI agents still take 25.8% of the latest batch but lost 3.2pp over the last two batches, and Agent infrastructure is flat at 16.6%. What is picking up the slack is anything that touches atoms, money or compute — Robotics and physical world (+2.0pp), Fintech (+1.8pp), AI infra and compute (+1.7pp), Healthcare and bio (+1.7pp), Defense and space (+1.5pp). Meanwhile classic software categories are shrinking: Developer tools -1.6pp, Consumer -1.9pp, Data for AI -1.0pp.
Rising
- Robotics and physical world (13.1%, +2.0pp, 129 since W24) — the agent playbook is being re-run on machines, where deployment friction still protects margins.
- Fintech (4.4%, +1.8pp, 96) — money movement is where AI output finally gets measured in dollars, and crypto rails are back on YC's own list.
- AI infra and compute (7.4%, +1.7pp, 82) — supply constraints have moved from models to power, siting and serving.
- Healthcare and bio (5.7%, +1.7pp, 74) — the only regulated vertical adding share fast; 443 tracked companies, only 5 marked inactive.
Crowded
- Vertical AI agents — 546 companies since W24 and 188 of 710 tracked rewrote their pitch in under six months. When a quarter of a cluster restates what it does, the category is being sorted, not built.
- Agent infrastructure — 255 companies, and the "Self-Maintaining APIs" RFS already maps to 198 companies from 2025 onward. The most requested theme is also the most supplied.
- Consumer — 783 tracked but only 1 acquisition and 34 pitch rewrites; latest-batch share fell to 3.5%. Low churn here reads as low motion, not stability.
What happened to last season's startups
Since 28 February: 651 companies added, 63 acquired, 56 marked inactive, 209 delisted, 546 materially rewrote their pitch. B2B SaaS produced the most exits (10 acquired of 752) — the tidy outcome for tooling that AI-native entrants can absorb. Vertical AI agents lost the most positioning: 14 acquired, 7 inactive, and by far the largest rewrite count.
Three moves worth reading closely. StarSling (X25) went from "Cursor for DevOps" to "Self-Driving CI: fast AI-native GitHub Actions runners" — from a generic assistant framing to a metered infrastructure product. CLODO (S25) dropped "Vibe GTM… with just a prompt" for "Planet-scale people discovery," trading a workflow wrapper for a data asset. The Robot Learning Company (X25) shifted from "infrastructure that robot intelligence is built on" to "tools for the next era of physical automation," a step down the stack toward what customers will actually pay for today.
Open windows
- The Future of American Defense — the thinnest supply of any RFS at 37 companies, against a cluster growing 1.5pp per two batches.
- Proving You're Human and AI-Native Compliance Infrastructure — two separate requests sharing one 38-company pool. Security and compliance is losing share (-0.7pp) exactly as demand for it is stated twice.
- Multiplayer AI — 41 companies. Nearly every agent product still assumes one user and one model in one session.
Idea candidates
- Proof-of-human verification for live voice and video channels, sold to call centres and recruiters who now cannot tell candidates from agents.
- An audit-trail layer that records what enterprise AI agents did, why, and under whose authority, formatted for the evidence requests compliance teams already receive.
- A shared workspace where several people and several agents hold one persistent task state — the missing multiplayer primitive behind the 41-company gap.