YC's Request for Startups:
Winter 2027 vs Fall 2026
Y Combinator has published the ideas it most wants founders to apply with for Winter 2027. Here are all 11, with the partner behind each and the argument in brief, set against the 13 from Fall 2026, and what the shift says about where YC thinks the next companies are.
Updated: October 28, 2026 · Ideas: 11 new, 13 prior · Source: ycombinator.com/rfs
At a glance
- What an RFS is: a list of ideas YC partners want to fund, each written up by the partner who cares about it. You do not need to work on one to get in, but an application that fits a fresh RFS lands with a partner already looking for it.
- Winter 2027 has 11 ideas from six partners. Brad Flora wrote three; Diana Hu, Jared Friedman and Gustaf Alströmer two each; Ankit Gupta and Jon Xu one each.
- The big turn: consumer is gone (four Fall 2026 ideas, zero now) and models are in (zero before, three now). The physical buildout holds at four, and government and regulated work grows to three.
- No idea carries over by name, but some read as follow-ons: Compute at Sea widens into the whole Data Center Supply Chain, and The Future of American Defense points toward AI for the Air-Gapped Economy.
Winter 2027: the 11 ideas
In YC's order, with the partner who wrote each one. Quotes are from the partners' own write-ups on ycombinator.com/rfs , which are short and worth reading in full.
1. Startups in a World of Abundant Space Access
SpaceX's projected 10,000 Starship launches a year would cut the cost of reaching orbit by orders of magnitude. YC wants businesses that only become viable once launch is 100x cheaper and access to space is plentiful.
“What business models maybe don't quite work today but could work when launch costs fall 100x or more”
2. The $15-an-Hour Expert
AI can lift an ordinary worker into a specialist without the credential, doing complex work in benefits, taxes, insurance and immigration. Done at scale, that builds a new middle class out of expert jobs.
“AI that turns an ordinary worker into an expert”
3. Robots Made in America
Robot foundation models are moving fast, but hardware manufacturing lags and most robots used in the U.S. are imported. New domestic-content rules open room for American magnets, motors, sensors and robot factories.
“more than 65 percent of a robot's component cost has to be domestic, rising to 75 percent in 2029”
4. LLMs That Write Well
Today's models sound more robotic than the 2024 and 2025 generation, a side effect of training tradeoffs that favor coding over style. There is room for models built first for expressive, natural writing and voice.
“the current models really do sound more robotic than the models from 2024 and 2025”
5. Plaid for Government Benefits
Applicants prove the same facts again and again across disconnected benefit portals. A shared data layer would speed verification and approvals and cut fraud, and it matters more as states expand Medicaid re-checks.
“Verify once, apply everywhere”
6. Wall Street for the Physical Buildout
Robots, GPU clusters and data centers need large upfront capital that ordinary bank lending does not serve well. YC wants specialized lenders and securitization built for physical AI assets.
“Every great buildout…created giant financial companies alongside it”
7. Vertical Foundation Models
Lean models built for manufacturing, energy or biology can match general-purpose models on their domain with far less compute. Winning takes deep vertical expertise, rights to good data and a clear, measurable performance edge.
“matching or beating general-purpose models at orders of magnitude less inference compute”
8. AI for City Hall
Most of the 19,000 U.S. municipalities still process permits, zoning, licensing and inspections by hand, and the delays cost billions in lost tax revenue. AI can automate the paperwork end to end.
“if all building permits in America were approved in 10% of the current time, US cities would make $16 billion a year”
9. Data Center Supply Chain
The AI data center buildout is the biggest capital project in American history. Past GPUs and power, the bottlenecks are transformers, switchgear, cooling equipment and raw materials such as electrical steel.
“The AI data center buildout is the largest capital investment project in American history”
10. AI for the Air-Gapped Economy
The most valuable data at defense contractors, banks and hospitals can never touch a cloud API. The openings are on-premises models, deployment tooling, security accreditation and dedicated hardware.
11. Own Your Intelligence
Open-weight models now compete with the frontier, so companies would rather own specialized models than rent them. The winners turn a customer's workflows, expertise and failures into a loop of continuous model improvement.
“companies have become more aware of their dependency on model rental”
Fall 2026: the 13 ideas it replaced
The previous list, as YC summarized it. It leaned toward consumer AI and new kinds of software, with a strong thread of physical-world and defense ideas.
| # | Idea | In one line | Theme |
|---|---|---|---|
| 1 | The Primer | Adaptive AI tutors for young children. | Consumer and money |
| 2 | The Future of American Defense | Affordable interceptors, drones and sensors. | Physical buildout |
| 3 | A Cloud for Small Software | Effortless hosting for tiny agent-built apps. | AI and work |
| 4 | Multiplayer AI | Shared live agent sessions for entire teams. | AI and work |
| 5 | Compute at Sea | Modular offshore data centers on floating vessels. | Physical buildout |
| 6 | AI-Powered Consumer Products for 1 Billion People | Agent-native apps for everyday consumers. | Consumer and money |
| 7 | AI for the Aging Population | Monitoring, independence and caregiver support. | Consumer and money |
| 8 | New Operating Systems for the Physical World | Coordinate field workers, robots and AI agents. | Physical buildout |
| 9 | The Best Time to Build in Crypto | Stablecoins, trading and agent payments. | Consumer and money |
| 10 | Data for the Real World | Capture rich physical-world data using sensors and robots. | Physical buildout |
| 11 | Proving You're Human | Verify authentic users amid AI impersonation. | Government and regulated |
| 12 | AI-Native Compliance Infrastructure | Automate oversight, reporting and audit trails. | Government and regulated |
| 13 | Self-Maintaining APIs | Agents update customer code as APIs change. | AI and work |
What changed, theme by theme
We grouped both lists into five themes. The groupings are ours, not YC's, and a few ideas could sit in two places, but the counts show the shape of the shift.
| Theme | Fall 2026 | Winter 2027 |
|---|---|---|
| Physical buildout | 4 The Future of American Defense; Compute at Sea; New Operating Systems for the Physical World; Data for the Real World | 4 Startups in a World of Abundant Space Access; Robots Made in America; Wall Street for the Physical Buildout; Data Center Supply Chain |
| Government and regulated | 2 Proving You're Human; AI-Native Compliance Infrastructure | 3 Plaid for Government Benefits; AI for City Hall; AI for the Air-Gapped Economy |
| Models | 0 None | 3 LLMs That Write Well; Vertical Foundation Models; Own Your Intelligence |
| AI and work | 3 A Cloud for Small Software; Multiplayer AI; Self-Maintaining APIs | 1 The $15-an-Hour Expert |
| Consumer and money | 4 The Primer; AI-Powered Consumer Products for 1 Billion People; AI for the Aging Population; The Best Time to Build in Crypto | 0 None |
1. From consumer to the boring middle of the economy
Fall 2026 asked for AI tutors, apps for a billion people, tools for the aging and crypto. Winter 2027 has none of that. Its customers are city permit offices, benefits agencies, defense contractors, lenders and transformer makers: slow buyers with huge budgets and paperwork that AI can finally do.
2. Models are a startup category again
Fall 2026 treated models as something you build on. Three Winter 2027 ideas say build your own: vertical foundation models that beat general ones at a fraction of the inference cost, models that own a company's workflows instead of renting them, and models trained for good writing rather than code. The bet behind all three is that open weights have made training a narrow model affordable, and that frontier models leave real gaps in style, cost and control.
3. The buildout moves down the supply chain
Physical ideas hold at four, but they move one layer down. Fall 2026 wanted to put data centers at sea and give the physical world an operating system. Winter 2027 wants the parts and the money: transformers, switchgear and electrical steel; American magnets, motors and robot factories; lenders for GPUs and robots; and whatever cheap launch makes possible in orbit.
4. Regulation as the moat
Two Fall 2026 ideas, compliance infrastructure and proving you're human, give way to three aimed at institutions that cannot simply adopt cloud AI: governments verifying benefits, cities issuing permits, and banks, hospitals and defense firms that need models running behind their own walls. The hard part, accreditation and procurement, is also what keeps competitors out.
5. AI as a labor upgrade, not a labor replacement
Fall 2026's work ideas were about developers: hosting for agent-built apps, multiplayer agent sessions, self-maintaining APIs. Winter 2027 has one work idea, and it points somewhere else: The $15-an-Hour Expert, where AI turns a generalist into a benefits, tax or immigration specialist.
How to use this list
- Read the partner, not just the title. Each write-up names the partner who will be most interested. If your company fits, say so in the application and expect sharp questions from someone who has thought about the space.
- Do not pivot into an RFS. YC funds many companies outside the list. An idea you have lived with beats a borrowed one that happens to be on it.
- Look for the seams between ideas. Wall Street for the Physical Buildout could finance Robots Made in America; Plaid for Government Benefits could feed The $15-an-Hour Expert; Vertical Foundation Models could run in the Air-Gapped Economy.
- Fall 2026 ideas still count. An RFS rotating off the list does not mean YC stopped caring. Defense, AI for the aging and agent payments are all still very fundable.
Winter 2027 ideas, authors and quotes are from ycombinator.com/rfs as of October 28, 2026; summaries are our paraphrase. Fall 2026 one-liners follow YC's own summary of that list. Theme groupings and analysis are Serverless Ventures' own. Spot something out of date? Tell us.