Requests for Startups

AI models are improving really fast, and they're now able to do complex work far beyond engineering. Historically, services became SaaS software. More recently, they became AI copilots. That's what most startups from 2023 to 2025 built: tools that help people do their jobs. What we're excited about now is the next step: AI-native companies that don't sell software—they sell the service. Instead of giving you a tool, they just do the work. The reason this matters is simple. The total spend on services is many times larger than the spend on software. And a lot of these services are already outsourced, which makes them much easier to replace with an AI-native product. We're especially interested in areas like: - Insurance brokerage - Accounting, tax, and audit - Compliance - Healthcare administration If you're building a company that replaces a service—not just improves it—we'd love to hear from you.

Intelligent agents are enabling a new level of personalization in medical care. We can now use an agent harness like Claude Code to analyze personalized health data, whether that be a diagnostic test, genome scan, EHR data, or wearables information to get highly accurate, user-specific suggestions. At the same time, two big revolutions in science are occurring. First, the cost of generating personalized diagnostics is plummeting. The cost of genome sequencing has fallen at a rate faster than Moore's law, and a variety of new diagnostics are entering the market enabling early detection of a variety of health signals. Second, the cost of printing n of 1 genetic therapies is plummeting. We can now design and deliver personalized medicines through delivery vectors like mRNA, and the FDA has expressed more openness to letting patients try out these procedures. We think these factors collectively are going to bring about a revolution in care delivery. Abundant data and intelligence can help patients more accurately assess their disease risk and democratize access to treatments for the most serious illnesses. To enable this, we believe a wide variety of new startups will support every step of this ecosystem. If you're building for the future of intelligent personalized care, we'd love to hear from you.

The biggest blocker to AI automation of companies is no longer the models, they just got so good so quickly. Now the blocker is the domain knowledge. Every company has critical know-how scattered everywhere. Some of it lives in people's heads. Some of it is buried in old email accounts, Slack threads, support tickets, and databases. The company works because humans vaguely remember where that knowledge is and how to apply it. But AI agents can't operate like that. If we want every company to run on AI automation, we need a new primitive: a company brain. We need Garry's G-Brain, but for every business in the world. A system that pulls knowledge out of all these fragmented sources, structures it, keeps it current, and turns it into an executable skills file for AI. This isn't a company-wide search or a chatbot over documents. It's a living map of how a company works: how refunds get handled, how pricing exceptions are decided or how engineers respond to incidents. Then AI systems can use that skills file to actually do the work safely and consistently. The company brain becomes the missing layer between raw company data and reliable AI automation. I think every company in the world is going to need one. If you're building this, you should apply to YC.

Last month, a swarm of cheap Iranian drones took out an AWS data center. Nobody stopped them. Now imagine an incoming swarm of a thousand coordinated drones. We'd stand no chance. While everyone is racing to stop individual drones, the next wave is coming and it's worse: not one drone, but swarms. Cheap, autonomous, jam-resistant, and lethal. A Patriot missile costs three million dollars. An FPV drone? Five hundred bucks. All of the cost advantage lies with the attackers. We are not ready. Today, counter-drone defense is a messy pile of radars, cameras, jammers, interceptors, humans with binoculars, and systems that don't talk to each other. That might work against a hobby drone. It will not work against a swarm. I want to fund founders building the counter-swarm stack. That could mean high-capacity interceptors — a single platform that neutralizes fifty drones, not one. Software that fuses every sensor and every defender on a site into a single real-time picture. We need non-kinetic defenses that don't exist yet: aerosols that foul rotors, streamers that entangle swarms. We need new attacks on the autonomy stack itself now that radio jamming is becoming obsolete. The key insight is drone defense is looking less and less like operating a weapon and more like running a real-time distributed system. The winning companies will look more like Cloudflare than Raytheon. If you're working on the future of drone defense, and how we can restore the cost advantage to defenders — we want to hear from you.

Before AI, users of a piece of software all interacted with the same interface. There were only light customizations like a few different views or theme and color options. When users think about "personalization" like on Netflix, it still has the same layout for everyone, just different imagery. As a result, most software has a one-sized-fits-all feel rather than being hypercustomized to a user. As an example: the way I use an email is very different from how most college students use email, yet all email clients look basically the same. The exception is in enterprise software, where forward deployed engineers customize software for each customer to make it a great experience for them. We think that coding agents have now gotten good enough to allow users to become their own forward deployed engineers and more radically customize the software they consume. I'm imagining users designing widely different interfaces for their use cases — perhaps my email client looks more like a task list, and a students' looks more like an events calendar. But these two interfaces likely share some underlying primitives and design decisions that a software team can build and ship. We think that in the future, software companies will ship these shared primitives with full intention that users will heavily modify the final interfaces. To enable this future, we will have to rethink the whole stack of software delivery. How will a developer make software that can be accessed by the user's coding agents? Do they have to deliver source code rather than packaged binaries? Can they only modify front-end visual elements, or are there ways for them to modify middleware on the fly to enable more interesting use-cases? If you're a radical thinker looking to define the future of software, we'd love to hear from you.

We are about to see an absolutely huge increase in the capacity that humanity has to put things in space because of reusable rockets from SpaceX and Stoke Space. This means that we're going to need enormous amounts of new compute capacity in space. The particular electronics in space we'd like to see is inference chips. There's going to be an absolutely enormous market for inference chips in space. What we mean by chips in space is something that is slightly optimized for mass, slightly optimized for thermal, and slightly optimized for radiation. If you are working at SpaceX or NVIDIA and you've been spending time doing chip design, then Y Combinator would love to hear from you.

We are funding more and more hardware companies, from medical devices to home robots to space companies. But building hardware in the US is still far too slow compared to China. In Shenzhen, a team can go from design to a new physical part in a day. In the US, that same loop often takes weeks. And that gap compounds. The problem isn't just the supply chain. It's iteration speed. China wins because hardware teams can move fast. They have dense supplier networks, rapid turnaround, and tight coordination between design and production. In the US, this system barely exists. A few startups are starting to build parts of it. For example, Hlabs (W26) is building actuators; Prototyping.io (P26) helps turn designs into mechanical parts in days. But the overall stack is still missing. We believe the next generation of great hardware companies will be built on much faster iteration loops. We're especially interested in startups that: - produce parts dramatically faster - enable rapid hardware iteration - tightly integrate design, manufacturing, and logistics If you're helping hardware teams move at an order of magnitude faster, we want to hear from you.

The broad idea we would like to see is developing industrial capabilities on the moon and in space, particularly extracting raw materials such as silicon, aluminum, iron, and titanium through electrolysis and 3D printing of complex structures from molten regolith on the moon, which should be more efficient than on Earth due to lack of supports. If you are working on something like that in the space, then Y Combinator would like to hear from you.

Most AI chips are designed for a world where inference means "prompt in, response out." Agents don't work that way. They loop: calling tools, branching, backtracking, holding context across dozens of steps. That's a completely different hardware problem. Current GPUs hit 30 to 40 percent of peak utilization on these workloads because the work is bursty, bouncing between memory-bound model calls, I/O-bound tool use, and CPU-bound orchestration. That gap is where purpose-built silicon wins. NVIDIA bought Groq for $20 billion because it saw this coming. Google built TPU v7 for inference specifically. But nobody's designing for the agent loop itself: fast context switching between models, native speculative decoding, memory built for KV caches that persist across an entire execution graph. Groq's real insight wasn't the chip. It was the compiler that made the chip work. We think that'll be true for whoever builds this next. If you understand both chip architecture and how agents actually execute, this is a rare moment where both halves of that experience matter. If you're building inference silicon for agentic AI, we'd love to hear from you.

Everyone's talking about how AI coding means the end of SaaS. Investors have wiped trillions off software market caps. Well, that might be bad news for incumbents, but it's good news for startups. If the incumbents really are this vulnerable, it should be the biggest startup opportunity in a decade. So, go build a challenger! The SaaS model won because custom software was too expensive. A five-person startup couldn't outbuild Salesforce. But AI has collapsed the cost of producing software by 10-100x, and that changes everything. The moat that once protected legacy SaaS — millions of lines of code, built over decades — is gone. There's a spectrum of ways to attack this. The simplest: clone an existing product and sell it for one-tenth the price. But you can go much further. You can build a product that's AI-native from the ground up. Not a chatbot bolted onto a 2010 UI, but software that fundamentally rethinks the workflow. You can take ten SaaS point solutions and bundle them into a single suite. You can build an open-source replacement for a product that costs $50K per seat and give it away, then monetize through services and hosting. Most people drawn to this idea start with simple targets like project management tools. We'd encourage you to think bigger. Go after the products that seem invulnerable: chip design software, ERPs, industrial control systems, supply chain management. The giant, 10-million-line codebases that have been untouchable for decades. The last generation of great software companies was built by replacing on-premise with cloud. The next generation will be built by replacing legacy SaaS with AI-native software.

The next trillion users on the internet won't be people, they'll be AI agents. And now is the time to "Make Something Agents Want". Agents are already browsing the web, doing research, making purchases, and managing legacy CRMs – but they're doing it on top of software that was designed for humans clicking buttons in a browser, which is slow, inconsistent, and brittle. Agents need a completely different foundation. Instead of visual interfaces like forms, buttons, and dashboards, they need machine-readable interfaces like APIs, MCPs, and CLIs. Agents also need thorough documentation, to enable them to discover, sign up for, and instantly start using new tools programmatically, without needing a human in the loop. That means every major category of software that people use today needs to be rebuilt for agents. And the new agent-first software won't come from incumbents bolting on agent support, it'll come from startups that build explicitly for agents as first-class citizens. While everyone else is building agents, the biggest opportunity might be building the software those agents depend on. So if you're Making Something Agents Want, we'd love to hear from you.

One of PG's wisest pieces of advice has always been that startups should sell to other startups. It's always been a hack to quickly get really smart, forward-thinking users that can help you shape your product into something awesome and important. It turns out there's another type of company that has really smart, forward-thinking buyers but it's one that's been out of reach to startup founders until AI came along: Massive Enterprises. We're not talking about just "big" companies, we're talking about the biggest companies in the world, which, it turns out, are run by incredibly smart, forward-thinking people. It's just always been too hard to get a hold of the right people there, too hard to build a product with the right depth and mix of features for a big company to use quickly and too low ROI for these companies to take on the risk of working with an early stage company. AI has changed the score on all three of these fronts: the people running these companies, again not just "big companies," but "the biggest companies in the world" F100 size companies, aren't hiding behind their computers, they're out there looking for teams that can bend AI to solve key problems for them. In the last 3 years for the first time ever we've seen YC companies land pilots and actual multimillion dollar deals within their first year if not during the actual YC batch. It's not unusual at all to see a company's first customer be one of the largest companies in the world. The buyers are awake and want to talk. Second AI has made it possible for the first time for small teams to ship incredibly thoughtful, nuanced products to large orgs in months, not years. The meme where some startups want to operate in stealth for 3 years until they reach feature parity with incumbents before launching is dead. For the first time a 2-3 person team can ship something a Fortune 10 can find useful before the ink is dry on their incorporation docs. Finally, leaders at these companies understand where their companies should create value internally, what they should outsource and what will happen to their companies if they don't adapt. Again, these companies are run by really smart people. We think this is a terrific environment for founders to sell into and want to back more teams that want to go after deals with the most famous, important customers in the world.

A single advanced AI chip goes through about 1,400 process steps, crosses a dozen countries, and takes five months to build. This supply chain is managed with spreadsheets, SAP, and phone calls. In 2021 a $300 chip held up a $50,000 car, and $210 billion in vehicles didn't get built. Companies could see their direct suppliers but had zero visibility into the second and third tiers. It's gotten worse. TSMC's advanced packaging is the single biggest bottleneck in AI compute right now, and NVIDIA has locked up over 60 percent of it. HBM memory is booked through 2026. Export controls change quarterly. At the same time the CHIPS Act is standing up new American fabs in Arizona, Texas, Ohio, and New York, each needing a supply chain built nearly from scratch. Almost none of the tooling you'd expect exists: real-time allocation tracking, multi-tier risk monitoring, export compliance. You need to understand wafer allocation and packaging constraints at a deep level to build this, which is exactly why it's a startup opportunity and not a feature inside SAP. If you're building supply chain 2.0 for chips, we'd love to talk.

The best AI-native companies we're seeing have figured out something most haven't: they've made their entire company queryable. Every meeting recorded, every ticket tracked, every customer interaction captured, all legible to an intelligence layer that learns from it. This turns a company from an open loop into a closed loop. In an open loop, you make a decision and maybe check the results weeks later. In a closed loop, the system monitors what's happening, compares it to what should be happening, and adjusts. I've seen teams that do this cut sprint time in half and ship twice as much. The problem is building this today requires brutal integration work, stitching together Slack, Linear, GitHub, Notion, call recordings, and a dozen other tools with custom glue code. There's no product that connects all this context into a single intelligence layer that can reason across it, flag when engineering is building the wrong thing, or generate specs agents can execute on. We think there's a big opportunity to build the connective layer that makes a company legible to AI by default. Not another dashboard. The system that turns a company's own artifacts into a self-improving loop. If you're building this, we'd love to talk.