23% Of YC's New Batch Builds Physical Things Now

The composition moved, and it moved fast.

Industrials went from 12.8% of the previous batch to 23% of the Summer 2026 cohort. That is the largest category shift in the group, and it happened in one cycle. B2B is still the biggest bucket at 52.3%, down from 59.2%. Fintech slipped to 6.8%. Consumer stayed where it always is, around 5.5%.

The batch itself grew about 20%, to roughly 235 companies.

Industrials is not one thing. Inside that 23% you find humanoid and home robots, warehouse and data-center robotics, robot training and evaluation infrastructure, drones, defense hardware, energy, and space. At least eight companies are building defense systems outright, up from six last batch.

So the tidy version of the story is this: a year ago the batch was full of companies building AI that talks to people. Now a large slice is building AI that moves things.

Because a bottleneck moved.

For about three years, the interesting constraint in AI was the model. Everyone built on top of a model and the product was whatever the model could say. That produced an enormous number of chat interfaces, and most of them competed on the same axis, which is a rough place to build a company.

The constraint now sits somewhere else. Models are good enough that the hard part is connecting them to something that has consequences: a robot arm, a warehouse, a supply chain, a piece of regulated paperwork. That is harder, slower, and much less crowded.

You can see the same logic in the software half of the batch. The previous cohort was full of agent applications. This one is full of the plumbing underneath them: routing, memory, identity, evaluation, payment rails for agents, runtime debugging. Companies are building the parts that make an agent survive contact with production.

And a third pattern runs through both halves. Instead of an AI copilot for accounting, several companies are just building an accounting firm run by agents. Same for law, ERP implementation, insurance, government affairs. The framing shifted from assisting a professional to replacing the function.

One bottleneck cleared, so founders went looking for the next one. That is the entire mechanic.

Almost certainly not, and this is where the number gets misread.

A batch composition tells you what a specific selection process rewarded in a specific window. It is a lagging indicator of what founders decided to work on twelve to eighteen months ago, filtered through what one accelerator found fundable. By the time you read about it, the people in the chart have a year of head start, a batch behind them, and in many cases a hardware background you do not have.

Chasing a category because it grew is the single most reliable way to arrive late and undifferentiated.

There is also a practical filter. Hardware has real capital requirements, long feedback loops, and supply chains that punish inexperience. The batch does include a few one and two-person teams shipping hardware-adjacent products, which is a genuine change and worth noticing. It is not an invitation for a first-time software founder to start machining parts.

The useful question is not what category is growing. It is what constraint you are unusually positioned to remove.

That the crowded layer moves, and you should know which layer you are standing on.

Here is a way to run the check on your own idea. Ask what would have stopped this from working two years ago. Then ask whether that thing has changed.

If the honest answer is nothing has changed, you are building in a layer where the constraint cleared a while ago, which means the competition already arrived. That is not fatal, but it means you win on distribution, pricing, or service quality rather than on the product being possible.

If something did change recently, and you can name it specifically, you are closer to where the batch data is pointing. The change might be technical. It might also be regulatory, or a shift in who holds the budget, or a price that fell far enough to open a market.

The founders in the industrials slice are not smarter than the founders in the chat interface slice. They just picked a spot where the answer to that question was yes.

Mapping this out is dull work and most people skip it. You can do it on a whiteboard, in a doc, or in a planning tool like Foundra that gives you a structure for competitive analysis and market positioning. Any of them beats holding it in your head, because the assumptions you never write down are the ones you never test.

Four tests, and they are quick.

Who is complaining? Real shifts show up as customer complaints before they show up as categories. If the people you would sell to are already working around the problem with spreadsheets, contractors, or duct tape, that is a signal. If nobody is complaining yet, you are early in a way that costs money.

What does the workaround cost them? Put a number on it. A problem that costs a customer $400 a month supports a very different company than one that costs them $40,000.

Why now, specifically? You should be able to finish the sentence "this became possible when..." with a date and an event. If you cannot, the timing story is decoration.

What happens if a big company notices? Some shifts favor incumbents, because they own the distribution and the data. Others favor new entrants, because the incumbent's existing business punishes them for moving. Know which one you are in before you raise money on it.

None of these require a research budget. They require a few conversations and the willingness to hear an answer you do not like.

In a good spot, as long as you are honest about which software.

B2B is still more than half the batch. The difference is that the winning B2B companies have stopped positioning as a better interface and started positioning as a replaced function. That shift shows up in pricing. You can charge a software price for a tool, or you can charge a headcount price for an outcome, and the second number is much larger.

It also shows up in what you have to be good at. Replacing a function means owning the failure cases, the compliance, the edge conditions, and the customer's anxiety about all three. That is service work wearing a software margin, and plenty of founders find they do not want it once they see it up close.

The other opening is infrastructure. When enough agents run in production, someone has to sell identity, memory, evaluation, payments, and debugging for them. Those are unglamorous businesses with durable demand, which is a description of most good software companies.

So the batch is not telling software founders to leave. It is telling them the easy layer filled up.

• Industrials went from 12.8% to 23% of YC's Summer 2026 batch, the biggest single-cycle category shift in the cohort.
• B2B still leads at 52.3% of roughly 235 companies, so software is not going anywhere.
• The pattern underneath is a moving bottleneck: models stopped being the hard part, so founders moved to connecting them to physical and regulated work.
• Batch composition is a lagging indicator. Copying it puts you eighteen months behind the people in the chart.
• Ask what would have blocked your idea two years ago, and whether that thing actually changed.
• Positioning as a replaced function rather than a better interface changes what you can charge.

Is Y Combinator's batch composition a good signal for what to build?

It is a good signal for what a selective process funded a year ago. It is a poor instruction for what to start today. Use it to understand which constraints founders are chasing, not which category to enter.

Should a non-technical founder avoid AI startups now?

No, but the opportunity has moved toward domain depth. The agent-native services pattern rewards people who know how a specific industry actually works, which is often a non-technical advantage rather than a disadvantage.

What does "physical AI" mean in practice?

Software that takes action in the physical world rather than producing text or images. In this batch that spans warehouse robots, home robots, drones, and the training and evaluation tooling those systems need to be trusted.

How much money does a hardware startup need to start?

More than software, though less than it used to. The batch includes small teams shipping hardware-adjacent products, but prototyping, certification, and inventory still consume capital on a schedule that software founders tend to underestimate.

Is defense tech a realistic category for a first-time founder?

It is a real and growing category, with at least eight companies in this batch. It also carries procurement cycles measured in years and compliance requirements that are difficult without relevant experience or a co-founder who has it.

1. New Economies: Y Combinator Summer 2026 Batch
2. Forbes: Meet The YC Startups Betting On What Comes Next
3. UrbanGeekz: Y Combinator Reveals 15 Startup Ideas It Wants Founders to Build in Summer 2026
4. Foundevo: Y Combinator Summer 2026 Demo Day, startup list and sector trends
5. New Economies: Y Combinator Spring 2026 Batch
6. PitchBook-NVCA Venture Monitor, Q2 2026