YC's Requests for Startups

The 11 startup ideas Y Combinator wants to fund in its Winter 2027 batch, what its old lists teach about reading this one, and what it means for a founder in India.

Y Combinator has posted its Requests for Startups for the Winter 2027 batch: 11 ideas it would like founders to work on. YC sums them up in one line: "AI's biggest bottlenecks are moving beyond the models."

Most founders will read the list as a menu of ideas to pick from. To me, the old lists are more useful, because they show what happens after YC asks for something.

In this article, I want to cover:

• what is YC's requests for startups?
• do these lists come true?
• who writes the list?
• what is YC asking for this time?
• what happens to the ideas YC repeats?
• what does it mean if you're building from India?
• how should you use this list?

What is YC's Requests for Startups?

It's a list of ideas YC would like to fund but isn't seeing enough founders work on. YC partner Dalton Caldwell described them in February 2024 as "ideas that should be made real — and that we'd want to fund! — but that we aren't seeing many people work on or apply with."

Paul Graham, YC's co-founder, posted the first one in July 2008, with 30 ideas. From 2009 he posted numbered Requests for Startups, the name YC still uses. In 2014, when Sam Altman ran YC, it became one page of 22 areas, which YC updated or added to every year or two until December 2022. Since 2024, YC has posted nine lists for the last eleven batches, and every idea carries the name of the person who wrote it.

You don't need one of these ideas to apply. YC's page says so at the top.

Do these lists come true?

The first one mostly did, years later, and mostly through companies YC didn't fund. Of the 30 ideas on the 2008 list, 22 now have a company worth, or sold for, $1B or more doing what the idea asked for, by my count. The other 8 came true only in part. For 6 of the 22, the company I matched is one YC funded, like Stripe and Whatnot. Counting every $1B company I found, not only the one I matched, YC funded a $1B company in 10 of the 22 ideas.

17 of the 22 companies started after the list, typically four years later, and 6 of those 17 took five years or more. The other 5 already existed when it went up, including Reddit, which Graham named in the list, and Airbnb, which hosted its first guests in 2007.

The later winners I could trace were already in the problem. Stripe came out of Patrick Collison's own side projects. PhysicsWallah started in 2014 as Alakh Pandey's YouTube channel teaching physics.

To me, that makes the 2008 list a good forecast of where markets went.

Who writes the list?

Six YC partners wrote this one.

To me, the list follows where YC already is. Its 2014 list, which had no names on it, put energy, AI, robotics and biotech first. In three of those four areas, YC had already funded the companies people now know best: Helion, Oklo, Ginkgo Bioworks and Cruise all joined YC in the nine months before the list went up. YC did go on to fund a $1B company in 14 of that list's 22 areas by my count, or 16 counting two loose matches.

In 2026, the list and the batch moved together. Hardware and industrial companies went from 13.5% of the spring batch to 30.1% of the summer batch, 69 of 229 companies, by the labels founders choose on their YC pages. The Summer 2026 list, which asked for hardware, was posted only one to two weeks before the application deadline, so most of those founders applied before they could have read it. To me, that fits a list describing what YC was already seeing more than one steering it.

The people who write a list often lead a batch, but not always. The partners who signed the summer list were the primary partner, the partner a company works with during YC, for 213 of the summer's 229 companies. For spring and fall, the share was 65% and 33%. The fall batch, led mostly by other partners, was 30.8% hardware too, so the partners alone don't explain the shift.

The six partners who signed this list led 163 of the summer's 229 companies, but only 1 of spring's 193 and 7 of fall's 185. Partners seem to take turns, so I can't tell from this whether they will lead Winter 2027.

What is YC asking for this time?

YC's summary is that AI's bottlenecks are moving beyond the models. It has asked for 8 of the 11 ideas before, and for part of 2 more. Only one is completely new.

Hardware, and the money for it. Four ideas. Jared Friedman's note says large power transformers now have a three-year lead time. Brad Flora's says the financing for robots, GPU clusters and data centres "mostly doesn't exist yet", and that banks underwrite a robot "like a forklift". To me, this group asks for people who know supply chains, factories and asset finance.

AI that doesn't come from the big labs. Three ideas: models built for one industry, models a company owns, and AI that runs inside the customer's own building. Brad's note says defence contractors, banks, hospitals and nuclear plants can't send their most valuable data to an AI API, and that founders YC works with keep hearing customers want the product "running entirely inside their own building, with no internet connection". It calls this market "slower and harder to sell", and also "where the largest budgets and the most defensible companies are".

Paperwork. Three ideas. The $15-an-Hour Expert hires smart people without credentials and backs them with AI for specialist work like insurance claims, medical billing and immigration forms. The other two are about government: AI for City Hall starts with US building permits, and Plaid for Government Benefits asks for one consent-based way to check the income, household and address details that every benefits programme asks for.

Writing. One idea, and the only one that is completely new. Jared's note says the big labs, if forced to choose, "will always choose to focus on coding and agentic workflows over style". OpenAI said its GPT-4.5 "excels at helping with writing and design". It lasted four and a half months in OpenAI's API before a cheaper model replaced it, though it stayed in ChatGPT for paying users until June 2026, and OpenAI said it would carry its writing quality into future models.

What happens to the ideas YC repeats?

It depends on the idea. YC has put robots in the title of three requests since 2024, this one included. By YC's labels, robotics and manufacturing companies went from 2.4% of the Winter 2024 batch to 14.4% of Summer 2026, 33 companies. Again, that matches the timing; it doesn't prove the lists caused it.

Government hasn't moved the same way. Since 2014, YC has posted eight requests with "government" or "City Hall" in the title, across six lists, this one included. By YC's labels, government companies rose to about 2% of what it funded in the late 2010s, peaking at 2.9% in 2017, and have been about 1% since 2020. Those labels miss a few, like Flock Safety below, but counting by hand adds only about one company a year.

To me, part of the reason is that most founders never meet these problems. YC partner Harj Taggar wrote in a Winter 2025 request: "It's hard to think of something that sounds less appealing as a place to look for startup ideas than the government but if you're up for digging in that space we'd love to hear from you." The sale is slow too. Cloud software sold to US federal agencies needs a security approval called FedRAMP, and Archon, a Winter 2025 YC company, says it takes 16 months.

Gustaf Alströmer, who wrote both government ideas on this list, sees it differently. His note says the idea that you can't sell software to government "used to be true. It isn't anymore."

Flock Safety's founders lived with the problem: it started after a run of break-ins in co-founder Garrett Langley's Atlanta neighbourhood, joined YC in Summer 2017, and by 2024 had cameras in over 4,000 US cities. Garry Tan, YC's CEO, asks for founders who have worked on the inside, in his Spring 2026 request on government fraud: "We're looking for teams where at least one founder has actually done work like this, whether that's a former FCA counsel, compliance lead or auditor."

To me, a repeat that still barely shows up in YC's batches, like government, is an opening for a founder who has met the problem, as a resident or from the inside.

What does it mean if you're building from India?

Four of the 11 requests are written for America. They cite its 19,000 cities and its Medicaid rules. Of the 71 companies YC has labelled government since 2014, its directory places 64 in the US, against about three in four of all its companies.

India already has much of what this list asks for.

The people. The kinds of paperwork The $15-an-Hour Expert names, like insurance claims and medical billing, are already done from India for US customers. Sagility, a Bengaluru company, does this work for US health insurers, healthcare providers and their partners, on claims, enrolment and billing, and over 22,000 of its 49,000-plus staff are in India.

The factories. In January 2026, CG Power won a ₹900 crore order to supply power transformers for a hyperscale data centre in the US, made at its factories in India and delivered over 12 to 20 months. Transformers are one of the shortages in Jared's note, which puts large ones at a three-year wait.

The rails. India already runs consent-based sharing of the kind Plaid for Government Benefits asks America to build. DigiLocker had 72.43 crore registered users by August 2026, and the Account Aggregator network lets a lender pull a person's bank data with their consent.

The rules. Since 2018, the RBI has required banks and payment companies to store payment data only in India, and SEBI's 2023 cloud rules keep the data of brokers and exchanges in India too. Jon Xu's note says owning your model matters most in regulated industries, "where strict data governance and locked-down infrastructure make renting even more expensive for the performance they get". These rules decide where data is kept, not who owns the model, but to me they make owning one an easier sell here.

If I were building from India, I'd look at three routes. Sell to the US what Indian factories and teams already do well, like transformers or claims work. Build the Indian version of a US request on rails that already exist, like DigiLocker and Account Aggregator. Or help Indian banks and brokers own their models.

How should you use this list?

1/ Start from your own idea, then check it against the list. YC's 2014 list said not to pick an RFS idea over one you'd rather work on: "don't apply with an idea described in an RFS if there's some other idea you'd rather work on."

2/ Look up the partner. Find the partner whose ideas sit closest to yours, then look at the companies they already work with on YC's company pages.

3/ Check whether a repeat shows up in the batches. YC's company directory can be filtered by industry and batch. If the share is climbing, as with robots, expect company. If it stays flat, as with government, the space is still open.

×