Verdict: Y Combinator's Fall 2026 Requests for Startups is a barbell — four nation-scale bets on the physical world (defense, sea compute, real-world data, work-site operating systems) on one end, and nine person-scale software ideas (single-user clouds, self-healing APIs, multiplayer AI) on the other. Classic team SaaS is absent for the second list in a row. The headline: a sitting U.S. Secretary of the Army co-authored an entry, a first in YC's 20-year history. Five of the 13 are buildable by a two-person team inside a quarter; the rest need domain depth, hardware, or regulatory moats that keep competitors out. The full accepted company gets the standard $500K SAFE.
Last verified: 2026-08-05
Total requests: 13 · Application deadline: July 27, 2026, 8pm PT
Theme: "AI is moving into the physical world"
Source: Y Combinator RFS — Fall 2026
What is YC's Requests for Startups (RFS) and why does it matter?
The RFS is YC's public list of problem areas its partners want founders to attack. It is not a homework assignment — YC funds hundreds of companies each batch, and most work on ideas not on the list. But it is the single clearest demand signal in early-stage venture. If you build one of these, you know the smartest money in the room already thinks the problem matters.
The Fall 2026 batch closed applications on July 27, 2026 at 8pm PT, with accepted companies joining the October–December batch in San Francisco. Demo Day is December 2. Every accepted company receives the standard $500K deal.
How is Fall 2026 different from Summer 2026?
Summer 2026 was almost entirely about software and AI infrastructure — agent plumbing, SaaS displacers, enterprise knowledge systems. You could start every idea from a text editor and ship meaningful progress in weeks.
Fall 2026 pivoted hard. The stated thesis: AI is now moving into the physical world — education, healthcare, defense, infrastructure, aging, the ocean. The bottleneck is no longer code. It is access, proximity, trust, regulatory approval, and willingness to work on problems that take longer than a weekend hackathon to understand.
The barrier to entry is higher, the domain knowledge deeper, and that is precisely why these categories are on the list. If they were easy to build from the outside, they would already be built.
All 13 Fall 2026 RFS Ideas at a Glance
Here is every request, who authored it, and what kind of team can realistically ship it:
| # | RFS Idea | Author(s) | Category | Tiny-team fit |
|---|---|---|---|---|
| 1 | The Primer | Andrew Miklas | AI education | Medium |
| 2 | Future of American Defense | Daniel P. Driscoll (Army Secretary) | Defense | Low |
| 3 | A Cloud for Small Software | Pete Koomen | Dev infrastructure | High |
| 4 | Multiplayer AI | Aaron Epstein | Collaboration tools | High |
| 5 | Compute at Sea | Francois Chaubard | Hard tech / infrastructure | Very low |
| 6 | AI Consumer Products for 1B People | Raphael Schaad | Consumer | Medium |
| 7 | AI for the Aging Population | Max Kolysh | Healthcare / consumer | Medium |
| 8 | New OS for the Physical World | Charlie Warren | Field operations | Low |
| 9 | The Best Time to Build in Crypto | Nemil Dalal | Crypto / fintech | Medium |
| 10 | Data for the Real World | Austin Tindle & Diana Hu | Hard tech / data | Low |
| 11 | Proving You're Human | Max Kolysh | Trust / identity | Medium |
| 12 | AI-Native Compliance Infrastructure | Daivik Goel | RegTech | High |
| 13 | Self-Maintaining APIs | Harsha Gaddipati | Developer tools | High |
Source: Y Combinator Fall 2026 RFS · Author names confirmed against the official YC page and Modelence's RFS index.
The 5 Ideas a Tiny Team Can Ship This Quarter
Of the 13, four to five are buildable by a two-person team inside 30–90 days. These are the plays where the bottleneck is product, not capex, maritime engineering, or security clearances.
1. A Cloud for Small Software (Pete Koomen)
AI makes it trivial to build a personal tool — a script that tracks your expenses, a bot that triages your inbox, a dashboard for three people on your team. What is still genuinely hard is deploying it, sharing it, setting permissions, and customizing the environment. AWS was designed for Big Software that serves millions; it carries enormous complexity for a tool three people will ever use.
Koomen — co-founder of Optimizely and a YC partner — wants a cloud built from scratch for Small Software, where sharing is as simple as a Google Doc. The hard problems are auth, tenancy, and sandboxing arbitrary code safely. If you've ever tried to deploy a two-user internal tool and spent the afternoon fighting IAM policies, you already understand the wedge.
2. Self-Maintaining APIs (Harsha Gaddipati)
When Stripe ships a breaking API change, customers find out because their code breaks in production. What if an agent scanned affected codebases, identified the specific usages, and opened a pull request with the fix — before anyone noticed? This is a genuinely new category of developer infrastructure, and the timing works because agentic coding tools like Claude Code have normalized giving external tools read access to repos. That normalization did not exist two years ago.
Gaddipati worked with 50+ API vendors before writing this RFS. The wedge is narrow (start with one API provider, one framework), the value is immediately measurable (fewer production incidents), and a two-person team can build a credible MVP.
3. Multiplayer AI (Aaron Epstein)
Google Docs beat Word. Figma beat Photoshop. The pattern is consistent: the multiplayer version wins. AI has not had its multiplayer moment. You open a chat, type a prompt, get an answer in a box only you can see. Sharing means sending a read-only transcript link. Epstein — a YC partner — wants shared live agent sessions that a whole team can drop into, redirect, and hand off, like working with a human colleague.
This is graph engineering meets product UX: org graphs, shared state, permissions, audit trails. A small team that ships a focused multiplayer layer on top of an existing coding agent or research agent can charge for it immediately. If you're building an agent OS for your business, multiplayer is the feature your team has been asking for.
4. AI-Native Compliance Infrastructure (Daivik Goel)
Financial compliance is still built on spreadsheets, siloed tools, and headcount. As businesses expand across jurisdictions, complexity compounds faster than revenue. The AI-native version rethinks what compliance looks like when AI is the default worker: regulatory change monitoring, anomaly flagging, report generation, audit trails — all tasks AI handles faster and cheaper than humans.
A small team can start with one jurisdiction and one framework (say, SOC 2 or EU AML directives) and expand from there. The buyer is already budgeting for compliance; you're replacing a six-figure consultancy retainer with software.
5. The Software Slice of AI for the Aging Population (Max Kolysh)
The full aging-population vision includes robotics, in-home sensors, and medical-grade monitoring. But the software layer — voice interfaces that hold real conversations, caregiver coordination across appointments, and medication-reminder apps designed for people who find Alexa frustrating — is buildable by a small team today.
The demographic math is brutal and inevitable. By 2030, one in five Americans will be 65 or older — about 71.6 million people, up from 62.7 million in 2025 (S&P Global, Claritas Pop-Facts 2024). Roughly 53 million Americans already provide unpaid care to an older adult (AARP / National Alliance for Caregiving, 2025). Almost no consumer technology is actually designed for older users. If you want to understand how an AI agent can help a caregiver coordinate care, see our guide to building an AI agent OS for daily work — the orchestration pattern is the same.
The 4 Hard-Tech Bets (Higher Bar, Bigger Moat)
Compute at Sea (Francois Chaubard)
AI needs more compute. Compute needs data centers. Data centers need land, power, and water — all three are running out where data centers want to be built, and local opposition kills projects before they start. The ocean is 70% of Earth's surface: no permitting, natural cooling, abundant sunlight, nobody using it for compute. YC wants modular floating data-center flotillas operating as a distributed cloud.
This is not theoretical. Microsoft's Project Natick proved the concept — in Phase 2, 855 servers ran submerged off the Orkney coast for two years with a lower failure rate than on land (Microsoft Research, Project Natick). But Microsoft discontinued Natick in 2024 because sealed underwater pods can't be upgraded to meet surging AI compute demand (IT Pro, July 2024). The engineering is solvable. The business model and operations are the hard part, and that's what YC is betting someone can crack.
The Future of American Defense (Daniel P. Driscoll)
This is the entry that made the whole batch historic. Daniel P. Driscoll — the 26th Secretary of the Army, confirmed February 25, 2025 — is the first sitting Cabinet secretary to author a YC RFS entry (U.S. Army official bio). A former Armor officer who deployed to Baghdad with the 10th Mountain Division and a former venture capital COO, Driscoll came from the startup world, not the traditional defense procurement world.
What he is asking for: low-cost interceptors that lower the cost per kill, next-gen sensors and payloads built for open system architectures (not proprietary lock-in), cutting-edge drones, resilient logistics, and advanced manufacturing — all of it able to survive extreme climates. The math problem is stark: a Patriot PAC-3 MSE interceptor costs about $4.2 million per unit based on the FY2025 Army budget request (Norsk Luftvern, FY2025 budget analysis), while a Shahed-style drone costs roughly $20,000–$50,000. That exchange ratio does not work at scale. The Army's xTechDisrupter competition, run with YC, is the procurement door opening to seed-stage startups — a structural shift, not a press release. For a deeper look at why AI infrastructure spending is reshaping markets this year, our analysis of why US tech stocks are falling in 2026 despite the AI spending boom connects the macro picture.
New OS for the Physical World (Charlie Warren)
Eighty percent of the global workforce does not sit at a desk — construction crews, maintenance technicians, fleet operators, healthcare workers. The software running their workflows has not meaningfully changed in 20 years. What is changing: there are now three types of worker in these environments simultaneously — AI agents that quote and schedule, robots physically deployed in the field, and humans wearing devices that record everything they do. No current software was designed to manage all three. The company that builds the operating system for hybrid human-robot-agent work environments and captures the proprietary end-to-end data from doing so is sitting on something every incumbent will eventually need. This requires years of immersion in a specific vertical — construction, fleet, maintenance — before you can ship anything credible.
Data for the Real World (Austin Tindle & Diana Hu)
AI has become superhuman at code, language, and images because abundant high-quality training data exists in those domains. For the physical world, the data is sparse, designed for human readability rather than AI training, and collected by instruments built decades ago. The opportunity: companies that build new methods of collecting dense, machine-readable real-world data in specific domains — energy, agriculture, logistics, construction. Once you can model a physical system with enough fidelity, you can control it. Hu co-founded Saronic, an autonomous maritime company; Tindle comes from the atmospheric-data world. This is as deep-tech as the list gets.
The 4 Ideas About Rebuilding Human Life at Civilizational Scale
The Primer (Andrew Miklas)
Inspired by Neal Stephenson's The Diamond Age, in which a young girl receives an interactive book — the Primer — that adapts to her completely and teaches her not just to read but to think and reason. For the first time, building something like it feels possible. The best education has always come from one-on-one tutoring (Aristotle taught Alexander), but that has been reserved for the wealthy.
Andrew Miklas — a YC partner and co-founder of Pipejump — is not asking for the full Primer yet. He wants a product that adaptably teaches young children reading, writing, and arithmetic at the quality of a devoted private tutor, at consumer scale. The parent buys it, the child uses it, it remembers and adapts. The boss fight is curriculum fidelity, longitudinal memory of the child, safety, and parent trust — not chat-quality demos. This is technically buildable by a small team; the moat is pedagogy and trust, not model capability.
AI-Powered Consumer Products for 1 Billion People (Raphael Schaad)
Every platform shift produces consumer giants — the web gave us Google and Airbnb, mobile gave us Instagram and DoorDash. AI is the biggest shift yet, but three years in, the only new icon on most people's home screens is ChatGPT. Schaad's timing argument: models are finally good enough to treat an agent like a person, and token cost is falling roughly 10× per year. Whatever costs $1,000 a month per user in tokens today will cost $10 a month before the end of the decade. Whoever builds the consumer application before that cost curve finishes its descent owns the category.
AI for the Aging Population (Max Kolysh)
See the tiny-team section above for the software slice. The full vision is broader: voice interfaces that hold real conversations (not command-response), monitoring for independent living, robotics for physical tasks, and software that helps family caregivers coordinate across appointments and emergencies. This is one of the most demographically inevitable markets in the world — the question is whether the products are trustworthy enough for families to put them in a parent's home.
Multiplayer AI (Aaron Epstein)
Covered in the tiny-team section — but its ambition scales beyond a wedge. Anywhere a team already crowds around one problem, Epstein argues, there should be multiplayer agents the whole team shares. If you're already running a self-improving AI agent OS, you've felt the gap — agent work that takes hours or days is inherently multiplayer, even when the tooling isn't.
The 5 Trust-Layer and Systems Plays
Proving You're Human (Max Kolysh)
In 2024, a finance worker at the British engineering firm Arup joined a video call with what appeared to be the company's CFO and several colleagues and authorized $25 million in transfers. Every other person on the call was a deepfake (CNN, Feb 2024; Financial Times, May 2024). Voice clones and fake video calls are now cheap and ultra-realistic, and every existing trust signal was built for a world where faking a human was expensive — a world that no longer exists.
Kolysh — who authored both this and the aging entry — wants the company that rebuilds the trust layer of the internet: the layer every bank, app, and video call checks before trusting anyone. The hard constraints: the solution must work without destroying privacy. A centralized identity database is a surveillance architecture most people would reject. The right answer is almost certainly decentralized, verifiable, and built on cryptographic proof rather than centralized records. Consumer-tech distribution and platform partnerships are the bottleneck, not the crypto.
AI-Native Compliance Infrastructure (Daivik Goel)
Covered in the tiny-team section. The full-market version isn't just one framework — it's global compliance when AI is the default worker, across jurisdictions and across regulatory regimes. The company that builds this becomes essential infrastructure for any business operating internationally.
Self-Maintaining APIs (Harsha Gaddipati)
Covered in the tiny-team section. The full vision isn't just Stripe — it's self-healing integrations across every API vendor, with agents that maintain the connections so production never breaks silently.
The Best Time to Build in Crypto (Nemil Dalal)
The most contrarian entry. Dalal — a YC partner who previously led crypto investments — argues for building in crypto precisely when sentiment is worst: prices down, hot narratives collapsed, builders leaving. His case: regulatory clarity has finally arrived, stablecoins are being adopted by every major financial institution, and AI agents will need financial rails that no legacy payment system can provide. The bear market filters for founders building real things rather than chasing liquidity.
YC expects that eventually every company in its portfolio will use crypto rails, even if they never know it. For more on why AI agents are outgrowing simple LLM chatboxes and need real infrastructure — payment rails included — see our comparison of large action models vs. LLMs in 2026.
What This Means for You
If you're a founder preparing a YC application: the RFS is a demand signal, not a gate. Most funded companies work on ideas not on the list. But an RFS-aligned angle is a free credibility shortcut for cold emails and pitch decks — especially for the five buildable plays, where a working MVP inside 90 days beats a 40-page deck.
If you're a small-business owner or builder: the barbell tells you where the smart money expects value to compound. The person-scale end — small software, self-maintaining APIs, multiplayer agent sessions — is where you can start without raising. The hard-tech end (defense, sea compute, real-world data) is where the moats are deepest and where domain expertise is the unfair advantage. Classic per-seat team SaaS is the undefended middle, and that's where AI-cloned competitors will crowd fastest. For a framework on positioning yourself in the current AI stack rather than competing with frontier labs, see our analysis of the 5 levels of AI building in 2026.
If you're an investor or analyst: the signal is not the ideas themselves — anyone can write a list. It is who wrote them. A sitting Cabinet secretary in the pitch room is procurement policy showing up in a startup accelerator. That is a structural shift in how defense technology gets funded and built, independent of whether any specific Fall 2026 company works out. For a deeper look at how AI infrastructure investment is diverging from hype, our analysis of leverage vs. long-horizon AI hardware investment strategies covers the capital-allocation side.
FAQ
Q: How many RFS entries are in the Fall 2026 batch? A: 13. They were published on YC's Requests for Startups page ahead of the July 27, 2026 application deadline, with the batch running October–December in San Francisco.
Q: Which RFS ideas can a two-person team actually ship? A: Five are realistic for a tiny team inside a quarter: A Cloud for Small Software, Self-Maintaining APIs, Multiplayer AI, AI-Native Compliance Infrastructure, and the software slice of AI for the Aging Population. The rest require hardware, regulatory clearance, domain depth, or platform distribution that takes years.
Q: Why is a U.S. Army Secretary writing a YC RFS entry? A: Daniel P. Driscoll, the 26th Secretary of the Army (confirmed February 25, 2025), came from venture capital and is systematically opening defense procurement to commercial startups. His RFS entry calls for low-cost interceptors, open-architecture sensors, drones, and resilient logistics. The Army partnered with YC on the xTechDisrupter competition — the procurement door is open to seed-stage startups in a way it was not before. (U.S. Army official bio)
Q: What was the single biggest RFS shift between Summer and Fall 2026? A: Summer 2026 was buildable from a laptop — agent infrastructure, SaaS challengers, knowledge systems. Fall 2026 moved the bottleneck from software to the physical world: education, defense, aging, infrastructure, the ocean. The required domain knowledge is deeper, which is exactly why these categories are still open.
Q: Does YC require you to work on an RFS idea to get in? A: No. YC's own framing states the RFS represents "just a fraction of what we fund." Most funded companies work on ideas outside the list. Execution quality, founder-market fit, and early validation matter more than alignment with a published category.
Q: What is the cost-per-kill problem the Army is trying to solve? A: A Patriot PAC-3 MSE interceptor costs approximately $4.2 million per unit based on the FY2025 Army budget request, while adversary drones can cost as little as $20,000. That exchange ratio is unsustainable at scale. Driscoll's RFS entry asks for low-cost interceptors and modular components that plug into open system architectures to collapse the cost asymmetry.
Discussion
0 comments