YC Just Backed 235 Startups. The Money Moved Down A Layer.

Two hundred and thirty five companies pitched 1,500 investors in San Francisco last week. That is the largest batch Y Combinator has ever run, roughly 20% bigger than the one before it.

The headline number is not the interesting part. The mix is.

Of those 235 companies, 52% are business-to-business. Consumer startups came in at 5.5%, close to a rounding error for a program that once launched Airbnb and Reddit. AI companies made up around 60% of the cohort, up from about 40% two years ago. Fintech, a YC staple for a decade, slid from 10.2% of the previous batch to 6.8%. Industrials nearly doubled, from 12.8% to 23%, and are now the second largest group.

Zoom out across the last seven batches and the trend is cleaner still. Infrastructure and developer tools went from roughly a quarter of each cohort to more than two fifths.

Here is the short version. The previous wave of AI startups built agents. This wave is building the systems agents need to work in production: routing, memory, identity, evaluation, and cost control.

The apps came first. Now come the pipes.

Because the apps kept breaking in the same places, and the same fixes kept getting rebuilt by every team independently.

Look at what got funded. Inkbox gives an AI agent its own identity: an email address, a phone number, an iMessage handle, a two factor vault. Sounds boring until you have watched an agent try to log into a third party service using a human's borrowed credentials. Archal checks whether an agent actually did the right thing and opens a pull request when it did not. Glen and Egoist Machines both work on memory. Agentcard handles payments. Understudy Labs captures your production traffic, fine tunes a smaller open weight model on it, gates quality with automated evals, then routes routine work to the cheap model instead of the frontier one.

Every one of those is a problem teams were solving badly, by hand, in-house.

That is the pattern worth learning. A category becomes a market when enough builders hit the identical wall and start writing the identical workaround. When four separate companies in one batch attack different slices of the same inference cost problem, that is not a coincidence. It is a signal the roll-your-own answer stopped scaling.

You do not need a YC badge to notice that happening in your corner of the world. You need to be paying attention to what your users keep building themselves.

No. It means the easy version of the app bet is gone.

Infrastructure looks appealing because it feels defensible. But it is brutal in ways founder Twitter rarely mentions. Your buyers are engineers, which means they can build a worse version of your product in a weekend and often will. Sales cycles run long because you are asking to sit inside someone's production stack. And the moment a model provider ships your feature as a native capability, your roadmap belongs to them.

Applications have the opposite trade. They get cloned fast, but your buyer is a person with a budget and a headache rather than a platform team with opinions.

So the real read is not "go build infrastructure." It is that value in the AI stack is settling into layers, and each layer now has different economics. If you know an industry cold, the app layer is still yours. If you have lived inside the failure modes of production systems, the infrastructure layer is open.

Three things most first time founders underestimate.

Reliability you can prove. Infrastructure buyers do not evaluate your demo. They evaluate what happens when your thing fails at 2am on a Saturday. Status pages, error budgets, and a support commitment matter more than your landing page copy.

A wedge narrow enough to say in one sentence. "Agent observability platform" is not a product. "We tell you which of your agent's tool calls silently failed" is. The companies that got attention in this batch had a sentence, not a category.

Distribution that does not depend on being clever. Developer products spread through docs, open source, and the few communities where your buyer already reads. Slow, unglamorous, and it compounds.

There is a fourth thing, and it is the one that kills people. You need to survive long enough for the category to mature. YC's standard deal is $125,000 for 7% plus $375,000 on an uncapped note, which buys most teams a year and change. If your product requires the market to grow up before you have revenue, count your runway before you count your addressable market.

Because the same pattern is spreading out of software and into physical systems.

The batch includes humanoid robots, warehouse and data center robotics, and companies building training infrastructure for robots. They sit next to agent tooling because they share a problem. A robot arm making a decision in a warehouse needs what a software agent needs: a way to evaluate whether the action was correct, a memory of what happened before, an identity the rest of the system can trust, and a cost model that does not bankrupt the operator.

The tools built for software agents transfer. That is the bet.

For a founder outside both categories, the takeaway is narrower. When an abstraction proves itself in one domain, watch the domains next door where nobody has applied it yet. That gap is often where a first time founder can compete, because incumbents in the neighboring industry are not reading the same news you are.

You do not need batch statistics. You need three lists and an hour.

List one: what do your users build themselves after they buy your product? Spreadsheets, scripts, Zapier chains, a shared doc somebody maintains by hand. Each one is a feature request they gave up on asking for.

List two: what do your competitors' support forums work around? Not the feature requests, the workarounds. People posting "here is how I got around this" are describing an unbuilt product.

List three: where does money leak in your customer's workflow? Understudy Labs exists because API bills got big enough that somebody would pay to shrink them. Cost is the most reliable wedge there is, because it is measurable on day one and nobody has to be convinced the problem is real.

Put the three lists side by side and look for the item that shows up twice.

This is planning work, not engineering work, and it belongs somewhere you can revisit. A spreadsheet is fine. So is Notion, or a structured planning tool like Foundra that keeps competitive research and product scope in one view instead of scattered across tabs. The tool matters less than the habit of rereading the lists next month.

Nothing here requires you to change your product. It does give you a free read on where attention and capital point for the next eighteen months, which is worth something when you decide what to say to investors.

Reread your pitch for layer confusion. If you describe yourself as both an application and a platform, pick one. Investors hearing 235 pitches learned to filter for clarity, and so did your customers.

Check whether a batch company just became your competitor or your supplier. Both are useful. A company solving your hardest internal problem is a build-versus-buy decision you should make on purpose rather than by default.

Look at what got funded in the layer below you. If the tools you depend on are about to get cheap, that changes your cost structure and your roadmap.

Then talk to five users about the thing they built themselves. That conversation has produced more real companies than any demo day ever has.

The batch composition is the signal, not any single company. 235 companies, 52% B2B, consumer at 5.5%, AI at roughly 60%. Infrastructure and dev tools now exceed two fifths of a typical batch.

Agent infrastructure became a real category. Identity, memory, evaluation, payments, and cost routing each got dedicated companies rather than sitting as features inside a bigger platform.

Convergence is the tell. When several independent teams attack the same problem at once, the do-it-yourself version has stopped scaling.

Infrastructure is not automatically the better bet. Engineer buyers can rebuild you, sales cycles are long, and model providers can absorb your feature.

Industrials at 23% shows the pattern transferring. Physical agents need the same scaffolding software agents do.

Run this analysis on your own market for free. What do users build themselves, what do forums work around, and where does money leak.

Does a shift toward infrastructure mean consumer startups cannot raise? Consumer was 5.5% of this batch, which reflects what YC selected rather than what is fundable everywhere. Consumer companies still raise, but the bar is traction rather than narrative, and a different set of investors writes those checks.

What is agent infrastructure in plain terms? The services an AI agent needs to work in production rather than in a demo: its own login credentials instead of a borrowed human account, memory between sessions, a check on whether its actions were correct, a way to pay for things, and routing that sends cheap tasks to cheap models.

Is it too late to build in this category? The category is roughly a year old as a distinct market. Too late would look like three funded incumbents with real distribution and an obvious default choice. That has not happened. What has happened is that the obvious slices are taken, so a new entrant needs a specific wedge rather than a general platform.

How much does YC actually invest? The standard deal is $125,000 for 7% on a post-money note, plus $375,000 on an uncapped note with a most favored nation provision. YC accepts roughly 1% of more than 10,000 applicants per batch.

Should I apply to YC because of this? Apply if the program's specific advantages fit you: fast feedback, a founder network, a Demo Day audience. Do not apply because a category looks hot. The trend you noticed today will have moved by the time you would start.

1. byteiota: YC S26 Demo Day, Agent Infrastructure Is Now a Category
2. New Economies: Y Combinator Summer '26 Batch
3. Little Might: YC Summer 2026, the full batch
4. Y Combinator: 2026 Demo Days
5. Y Combinator: The YC Deal