Requests for Startups

RFS is our tradition of sharing ideas we'd like to see founders tackle. These represent just a fraction of what we fund — if one excites you, take it as extra validation to dive in, but you don't need to work on these ideas to apply to YC.

Fall 2025

The last few years were about proving what AI can do. Now, it's about building with it. We've put together a new list of startup ideas we're especially excited about— ideas that treat AI not as a feature, but as a foundation. They reflect where we think some of the biggest opportunities are right now.

Retraining Workers for the AI Economy

By Harj Taggar
We talk a lot about the AI revolution in terms of models, chips, and software. But to make it a reality, there needs to be a huge buildout of physical infrastructure like data centers and semiconductor fabs. And that's where we have a problem. While we're focused on the race for AI talent, we also have a shortage of skilled tradespeople—the electricians, the HVAC technicians, the welders—who are essential to building this physical infrastructure. The government's new AI Action plan creates a forcing function to solve this. There's a big emphasis on a "worker-first" agenda, and directions to the Departments of Labor and Commerce to fund new rapid retraining programs for exactly these kinds of physical labor jobs. This creates opportunities for startups. We want to fund startups building a new kind of vocational school for the AI economy to train people for these jobs. We think you could use AI to create personalized training programs to get people job-ready in months, not years. The challenge is, how do you teach someone to weld or fix pipes via AI? Unlike coding, you can't learn these skills by typing on a keyboard — you need to learn by practicing them in the real world. This is where multimodal AI could create opportunities — for example, maybe a voice AI could coach someone through these tasks. Or perhaps some combination of AR/VR could let people practice the work in simulation with an AI tutor using vision models to watch them and give feedback. It's clear how you'd make money from this business — employers would pay to hire your well-trained workers. In the past, these types of training businesses — like coding bootcamps — have struggled to expand because it's hard to scale the quality of human tutors, but AI might solve this problem too. If you can make one effective AI teacher, it'll scale infinitely. This is a chance to build a huge business that lets everyone benefit from and participate in the new AI economy that's changing the world. If you want to work on this, we'd love to hear from you.

Video Generation as a Primitive

By David Lieb
Video generation models are getting really good. Google's Veo 3 already produces 8-second, photorealistic, sound-on clips for just a few dollars per video, often indistinguishable from reality. Soon, you'll be able to generate near-perfect footage of anything, on the fly, for a marginal cost approaching 0, and video will become a new basic building block for software. When this happens, a lot of new ideas become possible. It's definitely going to change media and entertainment: Imagine being able to create a brand new season of your favorite canceled TV series. Or a personalized kids cartoon starring your own family as the characters. Or an AI-native successor to TikTok, where every video is made for exactly one viewer. Video generation will also change how we buy things: Imagine you're shopping online, you'll be able to see yourself wearing the clothes or using the products you're browsing. And when you're apartment hunting, your stuff will be auto-staged in every listing you click on. This new primitive also opens up major changes to gaming and simulation: We'll have video games built with no game engine. We'll have APIs that return infinite robotic training data. And we'll definitely be able to have video calls with loved ones long after they're gone. And these are just the obvious ideas we can imagine today. We're interested in founders who treat generative video as a new computing primitive—not an output—and build new apps, tooling, and infrastructure for a world with limitless, low-latency video.

The First 10-person, $100B Company

By Aaron Epstein
Thanks to new AI tools, we believe it's now possible for small, high-agency teams – even solo founders – to build multi-billion dollar companies with as little as just $500k in funding from YC. 15 years ago, cloud computing came along and eliminated the need for spending tons of money on physical server infrastructure, making it easier to build a big company with way less capital. This is happening again now, with new AI tools that make it easier for ambitious founders to scale with far fewer people. That's why the best high-agency startups of the future will all optimize for one metric: revenue per employee. These small teams have incredible advantages over bloated incumbents, too. With smaller, efficient teams at scale, they won't get bogged down with the politics, excessive meetings, and lack of focus that grinds huge companies to a halt. They can just focus on winning with better speed and execution. We want to fund these high-agency founders to help them build the first 10-person, $100b dollar company.

Infrastructure for Multi-Agent Systems

By Pete Koomen
AI agents are evolving from single-threaded loops into distributed workflows that fan out many sub‑agent calls in a single run. These multi-agent systems are useful for everything from long-running workflows to agentic mapreduce jobs where hundreds of thousands of subagents apply human-level judgment to filter and search through large amounts of data in parallel. These systems are difficult to build. They require solving traditional distributed systems problems to ensure high throughput and reliability while controlling costs. They also introduce new problems that look familiar but can be solved at a higher level of abstraction, like: - how to write effective agent and subagent prompts - how to handle untrusted context - how to monitor and debug these agents We're looking for builders who have felt this pain in production and want to build tools to make these systems easier to build and maintain. If you want to make operating fleets of agents as routine and reliable as deploying a web service or running a Spark job, we'd love to hear from you.

AI Native Enterprise Software

By Andrew Miklas
Salesforce and ServiceNow are two of the world's largest enterprise software vendors, each making more than $10B a year in revenue, and each with a market cap of more than $200B. Both produce software needed to run the world's biggest businesses. Interestingly, both were founded about 25 years ago. Coincidence? I don't think so! Salesforce built the first cloud-native CRM, and ServiceNow built the first cloud-native ITSM system. Both rode the cloud computing wave to victory over their respective incumbents. The story is much the same for Netsuite, Successfactors, and a host of other B2B startups from the early two thousands. The founders of these businesses realized that SaaS offered a way to build a ten times better product. Just as important, these founders knew the incumbents would struggle to adapt to the new world of cloud computing, giving them the edge they needed to go up against them and win. AI presents the same once-in-a-generation opportunity for today's founders. Tomorrow's enterprise software systems won't just be the "system of record" for work done by humans. They'll have AI embedded deeply and thoughtfully throughout, and will help employees do their work faster and more accurately. Think Cursor for sales, HR, and accounting. Just as before, today's incumbents will struggle to rebuild their product around this new technology, giving today's startups the time they need to win. History doesn't repeat itself, but it sure does rhyme. Interested in building AI-native enterprise software? If so, we should chat.

Using LLMs Instead of Government Consulting

By Gustaf Alstromer
The U.S. government spends over $100 billion a year on consulting. As you might imagine, this isn't the most efficient or innovative part of our economy. But over the last couple of years, there have been a few big reasons we believe this will change. Most importantly, today, there is political pressure to cut wasteful consulting and spending. Every part of the government now runs on software, but usually custom software built by a consulting company, and anyone who has used this software knows we can do a lot better. Finally, LLMs today are so good that they can already do the jobs of many consulting firms. We've recently funded companies that help companies get approved to sell to the government, called FedRAMP. We've also recently funded companies that help the government cut regulation and use LLMs to help make sure the laws and policies coming out of the government are actually legal. We think there is a lot more work that government consulting firms like Deloitte and Accenture do for the government today, and we want to fund startups that build LLM software to do that work.