Y Combinator (YC), the Silicon Valley accelerator behind Airbnb, Stripe and Dropbox, published its Request for Startups for the Winter 2027 batch on October 27, 2026. The Request for Startups is the list of problems YC wants to see solved in each batch, signed by its partners. This one has 11 requests from 6 partners, and applications close on November 2.
The line that frames it comes from YC’s introduction: AI’s biggest bottlenecks “are moving beyond the models,” toward hardware, infrastructure, financial systems and specialized software, “from factories and data centers to government and space.” Only one of the 11 requests mentions AI agents. The rest bet on what’s scarce around the models: energy, materials, credentials, permits and trustworthy data.
For each request we checked what YC says, what data supports it (bull case), what data contradicts it (bear case) and what our own read is. We flag as a YC claim anything the accelerator doesn’t back with a source.
Summary
- The 11 requests fall into four fronts: the physical world (space, robots, the data center supply chain), financing that infrastructure, the state as a customer, and proprietary or specialized models.
- The request with the strongest evidence is “The $15-an-Hour Expert”: a study published in the Quarterly Journal of Economics found that AI raised the productivity of less-experienced workers by 34%.
- The weakest are the space request, which depends on unverifiable SpaceX figures, and the one about models that write well, which assumes a decline that no benchmark actually shows.
- In nearly every request, the most likely rival is an incumbent: Equifax, Tyler Technologies, Google, Eaton or Apollo.
- Crypto and stablecoins, present on the 2026 lists, are gone.
The 11 requests, at a glance
The table summarizes each request with the YC partner who signed it, according to YC’s page.
| # | Request | Partner | What it asks for |
|---|---|---|---|
| 1 | Startups in a World of Abundant Space Access | Ankit Gupta | Business models that only work with a 100x lower cost per kilo to orbit |
| 2 | The $15-an-Hour Expert | Brad Flora | AI that lets workers without credentials do expert-level work |
| 3 | Robots Made in America | Diana Hu | Robot components manufactured in the United States |
| 4 | LLMs That Write Well | Jared Friedman | A model people enjoy reading |
| 5 | Plaid for Government Benefits | Gustaf Alströmer | Verified income and residency data, shared with consent across social programs |
| 6 | Wall Street for the Physical Buildout | Brad Flora | Credit, leasing and hedges for robots, GPUs and data centers |
| 7 | Vertical Foundation Models | Diana Hu | Domain-specific models for manufacturing, energy and biology |
| 8 | AI for City Hall | Gustaf Alströmer | AI for municipal permits, claims and licenses |
| 9 | Data Center Supply Chain | Jared Friedman | Transformers, switchgear, liquid cooling, electrical steel |
| 10 | AI for the Air-Gapped Economy | Brad Flora | On-prem AI for defense, banks, hospitals and nuclear plants |
| 11 | Own Your Intelligence | Jon Xu | Letting companies build proprietary models on open weights |
Front 1: the physical world
1. Abundant space
What it asks for. Ankit Gupta is asking for startups whose business model only works if the cost of reaching orbit drops more than 100-fold. YC claims SpaceX “projects” 10,000 annual Starship launches, its next-generation reusable rocket, and that a disposable Starship costs almost the same as a reusable Falcon 9.
Bull case. The global space economy reached $686 billion in 2025, up 12% from the prior year, with 79% of it commercial activity, according to the Space Foundation’s annual report, a nonprofit that tracks the sector. Elon Musk said, per Ars Technica, that SpaceX will aim for weekly Starship launches in the first half of 2027 and two a week in the second half. SpaceX says its Starfactory plant is sized to produce up to 1,000 Starships a year.
Bear case. SpaceX doesn’t publish a Starship price, so the comparison with Falcon 9 can’t be verified. Its own long-term target is $100 to $200 per kilo, and outside estimates put the cost of each test flight at around $100 million. The 10,000 annual launches are far off the pace Musk projected for 2027: weekly in the first half, twice weekly in the second. Full orbital reusability has also never been achieved, as Ars Technica notes. The track record isn’t encouraging: Virgin Orbit, Richard Branson’s launch company, filed for Chapter 11 in April 2023, according to CNBC. And the main customer for that cheap capacity is SpaceX itself, through Starlink.
Our read. This request is an option on a single company’s execution curve. For a fund, the reasonable test is whether the business model holds up at $500 per kilo, without waiting for the $10 per kilo Musk floats as an aspiration.
3. Robots made in the USA
What it asks for. Diana Hu is asking for American manufacturers of robot components: magnets, motors, gears, encoders, sensors, grippers, chips and batteries.
Bull case. The United States installed 38,400 industrial robots in 2025, up 12%, ranking second behind China, according to the International Federation of Robotics (IFR), the sector’s global trade body. YC’s figure checks out. On July 28, 2026, the FCC, the U.S. communications regulator, added advanced robots made overseas (humanoids, quadrupeds and autonomous mobile robots) to its list of restricted equipment, and requires a domestically made product to have more than 65% U.S. components through 2028 and 75% from 2029 on. That creates captive demand. On top of that, MP Materials, the leading U.S. rare-earth miner, signed a deal with the Department of Defense for a 10-year price floor of $110 per kilo for NdPr (neodymium-praseodymium) and a commitment to buy 100% of the output from its new magnet plant.
Bear case. China installed 354,226 robots in 2025, 59% of global demand, more than half from Chinese brands, according to the IFR’s executive summary. That’s nine times the U.S. volume. Unitree’s G1 humanoid, from the Chinese manufacturer leading the budget segment, costs $13,500, according to Ars Technica. The FCC rule applies to connected mobile robots and leaves out classic industrial arms, which still come from Japan and Europe. And the 65% threshold is a regulatory decision a future FCC could reverse.
Our read. Industrial-policy money is flowing toward components (magnets, actuators, gears). Building full humanoids is still a cost war against Shenzhen.
9. The data center supply chain
What it asks for. Jared Friedman is asking for startups that manufacture what’s missing to build data centers: transformers, switchgear, breakers, liquid cooling and electrical steel. YC says large transformers now face delays of about three years and that grain-oriented electrical steel comes from a single U.S. plant.
Bull case. Large power transformers had an average lead time of 128 weeks in the second quarter of 2025, according to Wood Mackenzie, the energy consultancy. YC’s three-year figure is consistent with that. Cleveland-Cliffs’ Butler Works plant is, according to the company itself, the sole U.S. producer of grain-oriented electrical steel. And demand is already locked in: Anthropic’s S-1, the prospectus the AI lab filed to go public, discloses fixed or minimum commitments of $265.6 billion, mostly for compute.
Bear case. Eaton, Schneider Electric, Vertiv, Siemens Energy and Hitachi Energy dominate these categories and are expanding capacity. A startup making transformers faces its own multi-year qualification cycles. Demand also carries risk: models that need less compute to serve could reduce the need to build. YC’s comparison to the railroads, the highways and the Manhattan Project combined comes with no source at all.
Our read. The opportunity for a startup is selling to those incumbents: testing, design software, refurbishment, and alternatives to electrical steel.
Front 2: financing the infrastructure
6. Wall Street for the physical buildout
What it asks for. Brad Flora is asking for credit, leasing, securitization and hedging products for robots, GPUs and data centers, for example put options on GPU capacity.
Bull case. CoreWeave, the GPU-focused cloud computing provider, closed an $8.5 billion facility, which it describes as the first investment-grade-rated GPU-backed financing. Formic, a robots-as-a-service startup, raised $50 million in October 2026 between equity and debt. The market exists.
Bear case. The biggest tickets are already going to private credit funds and bank syndicates: SpaceX closed a roughly $40 billion GPU leasing commitment with an Apollo entity, the private credit manager, according to TechCrunch. There’s also a maturity mismatch: CoreWeave’s March loan runs through 2032, while a high-end chip can lose roughly half its resale value within three years, according to an analysis published by Forbes. And OpenAI and Anthropic cut their prices by as much as 80% within a few months, according to developer and analyst Max Woolf. Lower revenue per GPU-hour means less value backing the collateral.
Our read. A startup stands a better chance in the data and risk layer (GPU residual-value curves, usage telemetry, capacity hedges), selling to balance sheets the size of Apollo’s.
Front 3: the state as customer
2. The $15-an-hour expert
What it’s asking for. Brad Flora wants companies where AI lets workers without credentials do expert-level work: social benefits paperwork, taxes, insurance claims, medical billing, and immigration filings.
Bull case. This is the request with the strongest evidence on the whole list. A study by Erik Brynjolfsson, Danielle Li and Lindsey Raymond, economists at Stanford and MIT, published in the Quarterly Journal of Economics, found that AI assistance raised support-agent productivity by 14% on average, and by 34% for newer or less-skilled workers. The IRS, the US tax agency, published guidelines for AI use in tax practice that open the door for a licensed professional to review work done by AI and assistants. And the biannual Medicaid renewals (Request 5) are going to need more caseworkers.
Bear case. In many fields, the credential is a legal barrier. In immigration, telling someone which benefit to apply for or whether they’re eligible is already considered practicing law, reserved for attorneys or accredited representatives, according to the Department of Justice. Those same IRS guidelines require a professional to review what the AI produces. Klarna, the Swedish fintech, started rehiring people for customer service after its CEO admitted that AI-driven cuts had hurt quality, according to Bloomberg. And if the software doesn’t capture the value, these businesses end up with BPO-style margins.
Our take. The winners will be the operators who have a licensed professional sign off on the AI-assisted work. Those who try to replace the credential altogether are going to end up in court.
5. Plaid for social benefits
What it’s asking for. Gustaf Alströmer wants a “Plaid for Government Benefits,” referring to the company that connects bank accounts to financial apps: verified income, tax, and residency data, shared with consent across social programs. YC sums it up as “verify once, apply everywhere,” and it’s the only request that mentions AI agents.
Bull case. The GAO, the US Congress’s audit arm, estimates that fraud costs the federal government between $233 billion and $521 billion a year. YC’s “up to half a trillion” is the top of that range, which also covers all federal fraud, not just benefits fraud. The more concrete detail is the timeline: the 2025 budget law requires states to re-verify eligibility for Medicaid expansion adults every six months, for renewals starting January 1, 2027, according to CMS guidance, the federal agency that runs Medicaid. The volume of verifications for that population is set to double. There’s already proof of concept: Missouri uses SteadyIQ’s Income Passport for several programs, according to StateScoop.
Bear case. The incumbent is Equifax, with The Work Number, its employment and income database, which already sells verifications to states. Colorado paid $6.65 per query, according to a memo from its legislature, and Oregon’s multi-agency contract has a cap of $35 million through July 2028. The Seattle Times ran a headline saying Equifax “sees dollar signs” in the new Medicaid rules. And part of the point of the six-month reviews is to drop beneficiaries, so the buyer might not actually want renewal to be too easy.
Our take. The comparison that matters is with Equifax. The pitch is cheaper, consent-based verification, sold to states that are already budgeting for the 2027 renewals.
8. AI for municipalities
What it’s asking for. Gustaf Alströmer wants AI for permits, claims, records, licenses, and inspections across America’s 19,000 municipalities. YC says it “did some math” and that permits issued ten times faster would add $16 billion a year in revenue, and that almost all of those municipalities still work by hand.
Bull case. CivCheck is an AI permit-review startup. According to Honolulu officials and CivCheck itself, speaking to Stateline, the state-government news service, the average applicant’s wait dropped from 73 to 32.5 days, and for single- and two-family homes, review cycles went from 3.4 to 1.4. The city of Los Angeles is soliciting bids for an AI pre-review with the goal of responding to 90% of projects in under 30 minutes. There are already companies in the space: EffiGov (YC S25) runs an AI call center for local governments, and Polimorphic raised $18.6 million led by General Catalyst.
Bear case. “Almost everything by hand” is an overstatement. Tyler Technologies, the largest software vendor for local governments in the US, billed $2.3 billion in 2025, 87% of it recurring, according to its annual report, and its permitting product serves more than 700 agencies, according to the company itself. Accela is the other major incumbent. The $16 billion figure is unpublished YC math. And most of those 19,000 municipalities are small, with minimal IT budgets.
Our take. An AI layer that sits on top of Tyler or Accela and shortens review times has a real way in. Once results like Honolulu’s go public, incumbents will fold them into their own packages.
Front 4: proprietary and specialized models
4. Models that write well
What it’s asking for. Jared Friedman wants a model people actually enjoy reading and talking to. YC argues that models write worse because the reward signals used to train for coding flatten out voice, and that the big labs will always prioritize coding and agentic workflows over style.
Bull case. A February 2026 paper on arXiv, the scientific preprint repository, found that LLMs show far less variation in creative writing than professional writers do. Another paper showed that fine-tuning on high-quality books produced text that experts rated as professional-grade, so the specialization path does exist.
Bear case. This is the weakest-evidenced request on the list. We couldn’t find any benchmark showing that 2026 models write worse than earlier ones; the evidence points to homogeneity, not decline. The claim of deterioration is YC’s own. What’s more, companies built around “pleasant to talk to” models have tended to end up in licensing deals with their founders hired away: Inflection with Microsoft ($650 million, according to Reuters) and Character.AI with Google, both in 2024. And style is a post-training decision a lab can change with a single release.
Our take. For a publication, the useful conclusion is that distinctive voice has become a scarce input, and that a publisher with a strong style corpus has valuable training data. It’s the same argument we made in our guide to choosing a content agency: producing is cheap, judgment is expensive.
7. Vertical foundation models
What it’s asking for. Diana Hu wants domain-specific models for manufacturing, energy, and biology that match general-purpose models with far less compute. YC cites two examples: Jev and TimesFM-3.
Bull case. TimesFM-3, from Google, is a 330-million-parameter time-series forecasting model with open weights, integrated into BigQuery, according to the Google Research blog. TypeSafe launched Jev on September 15, 2026, a model that only makes decisions and responds in 70 to 500 milliseconds. IBM’s CEO said open-weight models deliver about 90% of frontier-model performance at 10 to 25% of the per-token cost, according to The Register.
Bear case. The best-known precedent went badly: GPT-4 outperformed BloombergGPT, Bloomberg’s financial model, on most of the financial tasks evaluated in an academic study. Both of YC’s examples come from, or are distributed through, big players (Google built TimesFM, and Jev is offered via Vercel). And manufacturing and energy data sit with Siemens, GE Vernova, and utility companies, which can train their own models.
Our take. The advantage lies in proprietary data. It makes sense to back teams with exclusive data access; without that, Google can ship an equivalent model for free.
10. AI for the air-gapped economy
What it’s asking for. Brad Flora wants frontier-level AI that runs on-prem, disconnected from the internet, for defense, banks, hospitals, and nuclear plants that can’t send data to a cloud API: models, deployment tools, accreditation, and hardware.
Bull case. Cohere, the Canadian enterprise model company, offers North with air-gapped deployment on as few as two GPUs. Cohere itself reported that an Alibaba Qwen model tried to send a file to an Alibaba Cloud server in Beijing during an isolated test, and recommended treating open weights as untrusted software. That creates demand for isolation and accreditation tools.
Bear case. Google already offers Gemini in its air-gapped cloud, fully disconnected. Palantir and the classified regions of AWS and Microsoft already have the accreditations, which take years to obtain. And many of the best open-weight models are Chinese (Qwen, DeepSeek, Kimi), and defense and regulated-sector buyers may be barred from using them.
Our take. The opening for a startup lies in accreditation and assurance: proving that a model is safe to run offline.
11. Owning your own intelligence
What it’s asking for. Jon Xu wants companies to have their own specialized models built on open weights, with their own evals and training data, and to be able to keep those improvements when the base model changes.
Bull case. Kimi K3.1 Pro, from China’s Moonshot AI, scores 55 on the Artificial Analysis index versus 58 for Claude Opus 5.5, at half the input and output price, according to Max Woolf. Anthropic’s S-1 shows that more than 80% of its third-quarter revenue came from versions released in the prior six months. Anyone renting a model has to migrate constantly, which is exactly the problem YC wants solved.
Bear case. Renting a frontier model keeps getting cheaper: OpenAI and Anthropic have cut prices by as much as 80% within months, according to Woolf, narrowing the cost gap that would justify owning your own model. That same S-1 reports net revenue retention (how much more the same customers spend a year later) of roughly 900% as of September 30, 2026, not even counting Claude Pro and Max subscribers: companies are renting more, not less. And owning open weights means taking on their supply-chain risk, as the Qwen case showed.
Our take. In practice, “owning” your intelligence means owning your evals and your data. Companies that can switch base models every quarter come out ahead either way.
What dropped off the list
Compared with the Spring 2026 (10 requests) and Fall 2026 (13 requests) lists, which we reviewed through third-party copies, the Winter 2027 list left out crypto and stablecoins (a topic we’ve tracked in our analysis of stablecoins as payment rails in Latin America), native AI agencies, native AI compliance, humanity verification, and the defense request. What remains are government as a customer, heavy industry, and the idea of making model training easier, which appeared in Spring 2026 as “Make LLMs easy to train” and now returns as request 11. A request coming back suggests YC still hasn’t found the company to solve it. The shift from crypto to AI as accelerators’ focus is a move we already analyzed in how to build a company that survives the hype.
What it means for incumbents and funds
Read from an incumbent’s perspective, the list is a map of where YC is going to fund competition. In nearly every request there’s an established company already charging to solve the problem: Equifax in income verification, Tyler Technologies and Accela in permitting, Eaton and Siemens Energy in electrical equipment, Apollo in GPU financing, Google in vertical and air-gapped models, Cleveland-Cliffs in electrical steel.
For those companies, each request points to which layer could end up in a startup’s hands: consent-based verification sitting on top of Equifax’s database, AI-powered review sitting on top of Tyler’s system, the risk telemetry a GPU lender needs. That’s where it’s worth deciding whether to partner, buy, or build.
For a fund, the list is useful for reading which requests YC has evidence for and which ones are a bet. Requests 2, 3, 5, and 9 rest on verifiable data and concrete regulatory calendars. Requests 1 and 4 depend on claims that can’t currently be checked. For a founder in Latin America, several requests are exportable: income verification for social programs, AI for municipal paperwork, and proprietary models built on open weights all have equivalents in the region. We’ve already analyzed how productized services are becoming a go-to-market model, which fits best with request 2.
The first test has a date: on January 1, 2027, Medicaid’s semi-annual renewals begin, and that’s where we’ll see whether states buy verification from a startup or renew their contract with Equifax.
Frequently asked questions
What is Y Combinator’s Request for Startups for Winter 2027?
Y Combinator’s Request for Startups for Winter 2027 is a list of 11 problems the accelerator wants to fund, published on October 27, 2026, and signed by 6 partners. It covers space, robots, data center supply chains, infrastructure financing, social benefits, municipalities, and specialized AI models. Applications for the batch close on November 2, 2026.
Is Y Combinator betting on AI agents for Winter 2027?
Only one of the 11 requests in Y Combinator’s Request for Startups for Winter 2027 mentions AI agents: “Plaid for Government Benefits,” which envisions agents managing enrollment in social programs. According to YC’s introduction, AI’s biggest bottlenecks are moving beyond models, toward hardware, infrastructure, financial systems, and specialized software.
Which YC request for Winter 2027 has the strongest evidence?
The YC request with the strongest evidence is “The $15-an-Hour Expert,” which proposes using AI so that workers without credentials can do expert-level work. A study by Brynjolfsson, Li, and Raymond published in the Quarterly Journal of Economics found that AI assistance raised productivity among newer or less-skilled workers by 34%. The main limit is legal, since in fields like immigration the credential is mandatory.
Which YC requests for Winter 2027 have the weakest evidence?
The YC requests with the weakest evidence are the one on space and the one on models that write well. The space request depends on 10,000 annual Starship launches and a cost figure SpaceX doesn’t publish. The writing request assumes models write worse now, but the studies available on arXiv show more homogeneous text, and no benchmark shows actual deterioration.
What dropped off YC’s W27 Request for Startups compared with 2026?
Compared with the Spring and Fall 2026 lists, YC’s Request for Startups for Winter 2027 left out crypto and stablecoins, native AI agencies, native AI compliance, humanity verification, and the defense request. What stays are government as a customer, heavy industry, and the idea of companies training their own models, which reappears as “Own Your Intelligence.”