State of the radar
Across 20,495 tracked companies, the center of gravity is still Vertical AI agents at 25.7% of the latest YC batch — but that share fell 3.3pp over the last two batches, the largest drop on the board. The share it lost went to things with physical or regulated surfaces: robotics (+2.2pp), fintech (+1.8pp), AI infra and compute (+1.7pp), healthcare (+1.7pp), defense (+1.5pp). The vocabulary data says the same thing from another angle: "data centers" now appears in 10 one-liners since 2025 having barely existed before, alongside "ai-native insurance" (4) and "prediction markets" (5). Founders are moving from writing agents to owning the balance sheets, buildings and liability that agents run on.
Rising
- Robotics and physical world — 13.5% of the latest batch, +2.2pp, 130 companies since W24. Cheap perception plus cheap actuation has finally made non-lab hardware a normal seed-stage bet.
- Fintech — 4.3%, +1.8pp, 96 total. After two flat years, teams are back on payments, credit and crypto rails, largely because AI-driven volume needs somewhere to settle.
- AI infra and compute — 7.4%, +1.7pp, 82 total. The "data centers" bigram is the tell: power, cooling, scheduling and capacity brokerage are being treated as startup problems, not hyperscaler problems.
- Healthcare and bio — 5.7%, +1.7pp, 74 total. Rising from founders, and rising much harder from the government (see below).
Crowded
- Vertical AI agents — 547 companies since W24 and still a quarter of every batch. Any pitch of the form "AI agent for [industry]" now competes with a hundred near-identical ones.
- Agent infrastructure — 256 total, 16.5% of the latest batch, and 199 companies already sit against the "Self-Maintaining APIs" RFS. This is the single most oversupplied request YC has open.
- Developer tools — 155 total but sliding (-1.6pp, 3.5% of the latest batch) with 85 companies chasing "A Cloud for Small Software." Declining share plus heavy incumbency is a bad combination.
What happened to last season's startups
Against a snapshot of 5,737 YC companies: 654 added, 63 acquired, 56 inactive, 209 delisted, and 547 materially rewrote their pitch. The rewrites concentrate brutally: Vertical AI agents accounts for 188 of them on 710 tracked companies — more than one in four — versus 34 rewrites in Consumer and 36 in healthcare. B2B SaaS took the most exits (10 acquired) and Vertical AI agents 14. Three repositionings show the direction of travel: Sciloop went from "AI Co-Scientist that automates ML experimentation" to selling "Expert STEM reasoning data for frontier AI labs" — from application to picks-and-shovels. Valgo moved from "algorithmic safety validation tools for autonomy" to an "insurance risk layer for physical AI" — from tooling to underwriting. Tesora narrowed from "AI-Native Software for Operations and Finance Teams" to "Frontier AI for Actuaries." The pattern: leave the crowded horizontal middle, go either upstream into data or downstream into a licensed profession.
Money on the move
- Thrive Holdings — $2.0B, B2B SaaS. Rollup-scale capital, not venture-scale.
- Lovable — $400M series D+, developer tools. The exception to a declining cluster.
- Moove — $250M, robotics and physical world.
- ClearJet — $25M, commerce and marketplaces.
- Cytix — $7.0M series A, security and compliance — the only round in a cluster with two open RFS themes.
- Metal Morph — $0.9M seed, climate and energy.
Open windows
- The Future of American Defense — only 37 companies mapped since 2025, the thinnest supply against any request, while defense share is up 1.5pp. Least-contested RFS on the board.
- Proving You're Human / AI-Native Compliance Infrastructure — two requests sharing 38 companies. Identity verification and audit-grade compliance are underbuilt relative to how fast agents are shipping.
- Multiplayer AI — 41 companies. Nearly every assistant built so far assumes one user and one model; shared-context, multi-human workflows are open.
What the state is funding
There are 1,111 open federal opportunities, 282 closing within 30 days, plus 4,885 NSF awards worth $3,582M. The gap analysis is lopsided: healthcare and bio shows 131 open calls and 905 awards against 252 startups since 2024 — a 42.7pp gap, by far the largest mismatch between public money and founder attention. AI infra (+5.4pp) and data for AI (+3.5pp) trail far behind. NIH alone has paid 677 small companies across 829 awards worth $451M, 294 of those Phase II — that is non-dilutive product money most software founders never look at. Two concrete targets: the SBA's Manufacturing and Small Business Cybersecurity Resilience Program 2026, closing 2026-09-04, answerable by a two-person security team; and the NIH Parent SBIR (R43/R44), open until 2027-04-05, which funds prototypes, not papers.
Idea candidates
- An underwriting and claims data layer for autonomous physical systems — the Valgo move, applied to warehouse and delivery fleets rather than to cars.
- Proof-of-human attestation as an API for marketplaces and support desks, priced per verification, targeting the 38-company gap behind YC's "Proving You're Human."
- Grant-to-prototype tooling for hardware and diagnostics teams: a service that turns an NIH SBIR Phase I into a shipped device, capturing part of the $451M the state already pays small companies.