Where the Ideas Went
I scraped eighteen batches of Y Combinator startups (2020–2026), had an LLM read all 367 descriptions, and turned them into one interactive map of where the ideas are going — and where there's still room to build.
I pulled every company from the last twenty batches — 367 startups across eighteen cohorts that actually had companies, from Summer 2020 to Winter 2026. Then I had an LLM read each description and tag it: what domain it's in, its business model, and how central AI is to the product. The result is the map below.
The map — tap anything
Each bubble is a theme. Across = how crowded it is (how many companies chase it). Up = how fast it's rising (its share in recent batches vs. older ones). Color = how AI-native it is. Size = how many companies. The quadrants do the work: top-left is rising-and-uncrowded — the white space.
Tap a bubble (or a row in the list) to drill into its sub-niches — "Government," for instance, breaks into procurement, public safety, permitting, and more, each with the actual companies building there.
Color of each bubble runs from grey (little AI) to green (fully AI-native).
What the map says
Three things jump out once you start tapping around.
1. AI stopped being a category and became the air. Across these batches, the share of companies where AI is the product went from 7% in 2020 to 86% in 2026. Today 53% of this slice touches AI in some way, and 38% are fully AI-native. The interesting question is no longer "is it AI?" — it's "AI for what?"
2. The marketplace era ended. The big red bubbles — Food & Delivery, Commerce & Retail — are crowded and cooling. The pandemic-era wave of on-demand and e-commerce startups has visibly receded. If you're building there now, you're building into a falling tide against a thick crowd.
3. The clearest white space is AI for Government. Public-sector tooling is rising fast, still relatively uncrowded, and already overwhelmingly AI-native — the rare corner that's both open and on-trend. Tap the green bubbles to see who's already there.
How it's built
The whole thing is deliberately simple, and it runs for free. No charting library — the map is hand-drawn SVG so it loads instantly on a phone and respects this site's security policy.
| Data | YC's public company directory, via its Algolia search index — pulled with a small Python script (standard library only, no dependencies) |
| Classification | An LLM read all 367 descriptions and tagged each one by domain, business model, and AI role |
| Analysis | Python — theme counts, momentum (recent vs. early share), and the crowded-vs-rising quadrants |
| Visualization | Hand-built inline SVG + vanilla JavaScript — no Plotly, no D3, ~80 KB total |
| Hosting | Static page on Vercel — same as the rest of this site |
That's the fun of public data: a few sentences per founder, multiplied by a few hundred founders, is enough to watch an entire market turn. The map is just a way to stand back far enough to see it.