Y Combinator just dropped its Requests for Startups for Winter 2027, and the list says a lot about where YC thinks the next wave of companies will come from. The short version: the interesting problems are no longer inside the model. They are in the steel, silicon, paperwork and money needed to put AI to work in the real economy.
We went through all 11 requests, watched every partner video and wrote down what each one is actually asking for. If you are thinking about applying to the YC W27 batch, or you just want to know what the best early-stage investors in the world are looking for right now, this is the page to bookmark.
Quick reminder before we start: YC is clear that the RFS are only a small slice of what it funds. You do not need to work on one of these ideas to get in. But if your idea overlaps with one, treat it as extra validation and say so in your application.
YC RFS Winter 2027 at a glance
Request for Startups | YC partner | In one sentence |
|---|---|---|
Startups in a World of Abundant Space Access | Ankit Gupta | Build the businesses that only work once launch costs drop 100x. |
The $15-an-Hour Expert | Brad Flora | Use AI to turn smart people without credentials into experts. |
Robots Made in America | Diana Hu | Build robots and their parts in the US, from magnets to grippers. |
LLMs That Write Well | Jared Friedman | Train a model people actually enjoy reading. |
Plaid for Government Benefits | Gustaf Alströmer | Verify once, apply for every benefit program at once. |
Wall Street for the Physical Buildout | Brad Flora | Invent the financing for robots, GPUs and data centers. |
Vertical Foundation Models | Diana Hu | Small, cheap specialist models for manufacturing, energy and biology. |
AI for City Hall | Gustaf Alströmer | Let AI handle permits, zoning checks and 311. |
Data Center Supply Chain | Jared Friedman | Fix the transformers, switchgear and cooling bottlenecks. |
AI for the Air-Gapped Economy | Brad Flora | Bring frontier AI to customers who can never touch the cloud. |
Own Your Intelligence | Jon Xu | Help companies own and improve their own models instead of renting them. |
Key facts about the YC Winter 2027 batch
Deadline: November 2, 2026 at 8 pm Pacific Time. Late applications are still read, but YC does not promise when you will hear back.
Program: three months in person in San Francisco, January to March 2027, ending with Demo Day.
Deal: $500,000 per company, split into $125,000 for 7% on a post-money SAFE and $375,000 on an uncapped MFN SAFE.
Where: you apply on ycombinator.com/apply. The full RFS list lives on ycombinator.com/rfs.
All 11 Requests for Startups for Winter 2027, explained
Every request comes with a short video from the partner who wrote it. We embedded all of them below, in the same order as on YC's page.
1. Startups in a World of Abundant Space Access (Ankit Gupta)
Ankit's argument is simple math. SpaceX expects to fly Starship thousands of times a year, and even the expendable version already costs about as much as a reusable Falcon 9, at a far lower price per kilogram. Once Starship is fully reusable, the cost of getting a kilo into orbit could fall by orders of magnitude.
The request is not "build rockets". It is the question every founder should ask about any falling cost curve: which business model looks crazy today but obvious when launch is 100x cheaper? YC wants those companies started now, before the curve arrives.
Good fit if: you have a concrete business that breaks even only at much lower launch prices, from in-orbit manufacturing to services for the satellites everyone else will be launching.
2. The $15-an-Hour Expert (Brad Flora)
This is the most optimistic request on the list, and probably our favorite. Most of the AI-and-jobs conversation is about replacement. Brad flips it: AI can turn a sharp person with no credentials into someone who does expert work. YC is already seeing companies hire people like a great customer service rep, give them strong AI tools and put them on work that used to need a specialist.
The target markets are the boring, enormous ones: benefits, taxes, insurance claims, medical billing and immigration paperwork. The rule of thumb is anywhere the credential, not the skill, is the bottleneck. Brad's comparison is the factory jobs that built the last middle class. This time it could be expert jobs.
Good fit if: you are building a services company with software margins, where your edge is the AI tooling and the training that makes non-specialists reliable.
3. Robots Made in America (Diana Hu)
Robot foundation models are starting to move skills from one robot body to another, so Diana thinks the hard part is shifting to hardware. The numbers back her up. The US installed 38,400 industrial robots in 2025, 12 percent more than the year before, yet most are imported from Japan and Europe. Even locally assembled robots often depend on Chinese rare-earth magnets, and China produced 94 percent of the world's sintered rare-earth magnets in 2024. When it restricted exports in April 2025, some automakers had to slow production.
Regulation is now pushing in the same direction. The FCC stopped authorizing new connected mobile robots unless they are made in the US or get an exemption, with a domestic share of component cost above 65 percent today and 75 percent from 2029. Screwing together imported parts will not clear that bar.
Good fit if: you build robots or any part of their supply chain in the US: magnets, motors, strain-wave gears, encoders, force sensors, grippers, chips, battery packs, or the machines that make them.
4. LLMs That Write Well (Jared Friedman)
If you have felt that AI writing got worse lately, Jared says you are not imagining it. As models got better at code and math, their prose became flatter and more robotic than the models from 2024 and 2025.
His take is that this is structural, not a passing bug. The reward signals that make a model great at coding also sand down voice and style, and the big labs will always pick coding and agents if they have to choose. That leaves a clear gap for a startup whose whole mission is a model people genuinely like reading and talking to.
Good fit if: you have taste in writing, a view on how to measure it, and the research skills to train for it.
5. Plaid for Government Benefits (Gustaf Alströmer)
Anyone who has applied for public benefits knows the loop: prove your income, prove who lives with you, prove your address, then repeat it all for the next program and again at renewal. Every agency runs its own portal and none of them talk to each other. Honest people get rejected over one missing document, while caseworkers check pay stubs by hand.
The same broken system also invites fraud, which the GAO estimates costs the federal government up to half a trillion dollars a year, and AI can now fake a pay stub in seconds. Gustaf's answer is the Plaid playbook: with the applicant's consent, pull verified income, tax and residency data straight from the source and reuse it across every program. Verify once, apply everywhere.
The timing is concrete. From 2027, states have to re-check millions of Medicaid enrollees every six months, and AI agents that help people enroll will need data they can trust.
Good fit if: you can handle data integrations, identity and consent, and you are not scared of selling to government.
6. Wall Street for the Physical Buildout (Brad Flora)
Software is cheap to start. Robots, GPU clusters and data centers are not, and someone has to pay for all that hardware up front. Right now the financing mostly does not exist. Banks still underwrite a robot as if it were a forklift.
Brad points out that YC founders are already inventing the instruments themselves, like lending against robots that each earn around $500k a year, or put options on GPU capacity. Every big buildout in history, from railroads to electricity to telecom, created giant financial companies next to it. YC wants the lenders, leasing companies, securitization and hedging products for the physical AI economy.
Good fit if: you understand both the machines and the money. That combination is rare, which is exactly the point.
7. Vertical Foundation Models (Diana Hu)
Diana wants more research labs building specialized foundation models for fields like manufacturing, energy and biology. She uses Jev as an example of why developers care: instead of free text, it returns a typed decision with a probability, without anyone training a classifier for each task.
The bar she sets is high. These models should work zero-shot or from a handful of examples and match or beat general models at a tiny fraction of the inference compute, cheap enough to run everywhere. Google's TimesFM-3 shows the pattern for forecasting: a 330M-parameter model that forecasts related time series together without task-specific fine-tuning. It is trained once, and every user brings their own history as context.
The hard part is data. You need rights to high-quality, verifiable domain data and a clear win on accuracy and compute against existing specialist models.
Good fit if: you know one field deeply enough to build its model, and you can get the data to do it.
8. AI for City Hall (Gustaf Alströmer)
Want to build a house or open a business in the US? First you need your city's permission, which means forms, plan reviews, inspections and months of waiting. In some places the permit takes longer than the construction.
City staff spend their days on repetitive checks, like comparing drawings with the building code or matching applications to zoning rules, and AI can do most of that. YC ran the numbers: if every US building permit were approved in a tenth of the current time, cities would collect about $16 billion more in property tax every year.
Start with permits, then expand into 311, public records, licensing and inspections. There are around 19,000 cities in the US and nearly all still do this manually. Gustaf also takes on the old line that you cannot sell software to government. YC has funded several AI-for-government companies, and some are among the fastest growing it has seen.
Good fit if: you are ready for long sales cycles and want a market that almost nobody serves well.
9. Data Center Supply Chain (Jared Friedman)
Jared calls the AI data center buildout the largest capital project in American history, bigger than the railroads, the highways and the Manhattan Project put together. Everyone knows about the shortage of GPUs, power and electricians. The less obvious bottlenecks are where startups can win.
Large power transformers have lead times of about three years. Switchgear, circuit breakers and liquid cooling gear are sold out. Even raw materials are tight: the grain-oriented electrical steel used in transformer cores comes from a single US plant that cannot keep up.
Good fit if: you are working on any part of that chain, from components and materials to the software that schedules and procures them.
10. AI for the Air-Gapped Economy (Brad Flora)
The world's most valuable data is not in the cloud. Defense contractors, banks, hospitals and nuclear plants keep it behind walls and cannot send it to an AI API. That is a big reason so much of the economy is still waiting on AI.
Brad hears the same story from founders again and again: the customer loves the product, but wants it running entirely inside their own building with no internet connection. YC wants the companies that make that possible: on-prem models, deployment tooling, security accreditation and hardware. Selling is slower and harder, but that is where the biggest budgets and the most defensible companies are.
Good fit if: you know how security reviews and accreditation work, or you have already lost a deal because your product needed the cloud.
11. Own Your Intelligence (Jon Xu)
Almost every company building with AI today rents its intelligence: send tokens to a frontier lab, get answers back, pay the bill. That trade made sense while the best models were years ahead. Jon argues it no longer holds, because open-weight models are now competitive with the top frontier models on meaningful benchmarks.
He expects more companies to want models they own, improve on their own data, run where they choose and update on their own schedule, especially in regulated industries. But downloading weights is not the same as owning intelligence. Someone has to turn workflows into evals, expertise into training data and production failures into improvements, then carry all of that forward to every new base model.
Good fit if: you are building the tooling or the service layer that gets a company from "we downloaded an open model" to "our model gets better every week".
ChatGPT
Claude
Gemini
Perplexity