What YC Startups Are Building: 1,025 Companies Across 6 Batches

Why the question is worth asking carefully

"The YC bubble" is usually said with a smirk. It means a few hundred founders in the same three neighborhoods, reading the same essays, arriving at the same conclusions in the same week. And the joke lands, because the batch really does converge. But dismissing it misses what the directory is actually good for.

YC is the closest thing the industry has to a leading indicator. It funds companies eighteen to thirty-six months before the rest of the market decides whether that category exists. If you want to know which problems will have four funded competitors on your buyer's shortlist in two years, the batch list is a better instrument than any trend report — as long as you read it as a set of bets, not a set of predictions.

How we counted

Every number below comes from Y Combinator's own public company directory, mirrored as structured data by the open-source YC API. We pulled the full company list on August 4, 2026 and grouped it by batch, industry, sub-industry and tags. "AI company" means the company's own one-line description or its YC tags reference AI, agents, LLMs or machine learning — their words, not our interpretation. That's a deliberately conservative definition, and it undercounts: plenty of companies build on models without leading with it. YC company directory · YC open-source API

The shape of the last six batches

YC now runs four batches a year instead of two. That change matters more than it sounds: the portfolio turns over twice as fast, themes propagate faster, and a category can go from novel to crowded inside a single calendar year.

The dip to 56% in Winter 2026 is tempting to read as cooling. It probably isn't. It reads more like founders getting self-conscious about the label — when every company says AI, saying AI stops being information, so some stop saying it. The underlying products didn't change much between Fall 2025 and Winter 2026. The vocabulary did.

That's the first thing the data teaches you about hype cycles: the language moves before the substance does, in both directions.

Theme 1: the agent stopped being a product and became a job title

The most common company shape right now is an agent pointed at one back-office workflow. Not a chatbot, not a copilot — something that owns a task end to end. In Winter 2026, 57 companies used the word "agent" in their description. In Spring 2026, 81.

Read the descriptions and the pattern is oddly specific: agents for corporate accounting teams, agents that automate customer success, agents that build personal injury cases, agents for therapists, agents for enterprise IT, autonomous insurance brokerages. Nobody is pitching a general assistant anymore. The wedge is always a role, and increasingly the pricing follows the role rather than the seat.

This is what maturity looks like inside a hype cycle, and it's also what commoditization looks like. Both are true at once. Narrowing to a role makes the product buildable and sellable; it also makes it describable in one sentence that four other founders in the same batch can write.

Theme 2: the picks and shovels caught up

The second-largest cluster is infrastructure for the first cluster. B2B infrastructure became the top sub-category in Winter 2026 (23 companies), Spring 2026 (19) and Summer 2026 (23), overtaking engineering and product tooling, which led the 2025 batches.

In practice: control planes for agents, knowledge search built for agents rather than humans, evaluation and observability layers, runtime security for models that keep changing, compliance and governance for agents touching regulated systems, and human-expert-as-API services that patch the gaps agents can't cover.

When a market starts funding the tooling around a category, the category has stopped being speculative — you don't sell governance to something that doesn't exist in production. That's the strongest signal in the whole dataset, and it's a quiet one.

Theme 3: atoms came back

This is the real story of the 2026 batches, and it's the one that gets the least attention. Industrials went from 15 companies in Spring 2025 to 37 in Summer 2026 — second only to B2B, and larger than fintech, healthcare and consumer combined in that batch.

Inside industrials: drone defense operating systems, counter-drone companies with real defense revenue, autonomous robots for solar farm construction, US-made robot parts, in-space manufacturing, AI chips, aerospace suppliers. Defense — a category most YC founders wouldn't touch five years ago — now appears in every batch.

There's an obvious cynical reading (geopolitics made it fundable) and an obvious optimistic one (models finally got good enough to control machines). The more useful reading is structural: when software gets easy to build, software stops being scarce, and scarcity moves to the things that are still hard — supply chains, certification, physical deployment, tolerance for slow timelines. Smart founders follow scarcity, not enthusiasm.

Theme 4: consumer is smaller than it has ever been

Eight of 174 companies in Summer 2026 were consumer. The Airbnb-and-Dropbox era of YC is, statistically, over. The consumer companies that do get in are AI-native personal software — health companions, personal agents, voice-first devices — rather than social apps or marketplaces.

It's worth sitting with that. The batch that built the consumer internet now funds it at roughly 5%. Partly that's distribution: acquiring consumers costs more than it did, and the platforms that made 2010-era growth possible closed. Partly it's economics: a B2B agent charging four figures a month reaches default-alive faster than any consumer app. Rational, and still a loss. Fewer people are trying to build things ordinary people love.

So is the bubble just building the same thing over and over?

Partly, yes. Read two hundred descriptions in a row and the sameness is undeniable: an agent, a vertical, a workflow, a claim about hours saved. That's what a hype cycle looks like from the inside, and it's why "YC bubble" works as an insult.

But the honest answer has a second half. The batch is converging at the idea level and diverging at the execution level. Two companies with the same one-liner now differ in the things a directory can't show: which regulated buyer let them in, whose data they can legally touch, what breaks at scale, who will still be there in year three of a hardware timeline. Convergence in language, divergence in difficulty.

The other half of the honest answer is that most of these companies will not work, and that's the design of the thing. A batch isn't a forecast. It's a few hundred parallel experiments where the portfolio only needs a handful to be enormous. Reading it as consensus about the future is the mistake. Reading it as a map of where the smartest available talent thinks the difficulty has moved is the use.

What this actually means if you're a founder outside the batch

The takeaway isn't which idea to copy. It's that hundreds of well-funded teams are shipping near-identical one-liners into the same buyer's inbox this quarter, and doing it with YC's distribution behind them. "AI agents for X" is no longer a differentiator; it's the price of entry.

When products converge, the tiebreaker moves upstream of the product. Buyers pick the company whose thinking they already recognize — a post that named their problem before the demo, a teardown that showed real judgment, a public opinion about the category that turned out to be right. That's not a marketing trick; it's the only durable asymmetry left when four companies can build the same thing in a quarter.

Which is the uncomfortable conclusion for technical founders: the work of explaining what you see is now part of the work of building. Not because content is magic, but because in a market this crowded, being understood early is the closest thing to a moat you can start building today.

Frequently asked questions

What percentage of YC startups are AI companies?

Between 56% and 73% depending on the batch, based on how companies describe themselves in YC's public directory. Across the last six batches the average is roughly 66%. Counting companies that use models without leading with the label, the practical share is higher.

How many companies are in a YC batch now?

Recent batches range from about 143 to 198 publicly launched companies. With four batches per year, YC is funding roughly 700 companies annually — a large increase over the two-batch era.

How many batches does Y Combinator run per year?

Four: Winter, Spring, Summer and Fall. The public directory shows companies tagged across all four seasons for 2025 and 2026, replacing the historic Winter/Summer-only cadence.

What sector is growing fastest in YC batches?

Industrials — robotics, manufacturing, drones, aerospace and defense. It more than doubled from 15 companies in Spring 2025 to 37 in Summer 2026, while consumer shrank to single digits.

Is the YC AI bubble going to pop?

No dataset answers that honestly. What the directory does show is concentration: most companies in a batch sell AI software to businesses, and many describe near-identical wedges. Concentration like that compresses pricing and raises the cost of distribution long before it resolves in either direction — so the risk shows up in go-to-market economics first, not in headlines.

References and sources

Y Combinator's official company directory, filterable by batch, industry and tag.ycombinator.com/companies

The open-source structured mirror of the YC directory used for the counts in this post.yc-oss.github.io/api

Y Combinator's Requests for Startups, which explains the areas YC actively encourages founders to work on.ycombinator.com/rfs

All figures were pulled on August 4, 2026 and reflect companies publicly listed at that time. Batch pages continue to update as companies launch out of stealth, so counts for the newest batches will rise.