YC: we're all in on these directions — if you're building one, come get funded
- YC, the world's best-known startup accelerator (Airbnb and Stripe both came out of it), has laid out the 13 directions it most wants to fund in fall 2026. The subtext is simple: we love these, so if you're building one, come get funded. And if you're not, you can still apply.
- This batch has one unifying theme: AI stepping out of the screen and into the real world — defense, elder care, job sites, the ocean. Pure generative AI (text-to-image, text-to-video, the models themselves) is almost entirely absent this time.
- For the first time ever, one direction was written by the sitting U.S. Secretary of the Army himself. He's openly asking founders for low-cost interceptors and for sensors and drones that plug into the military's open architecture — the goal being to drive down the "cost per kill."
- Why bet now: intelligence is finally good enough (you can treat an Agent like a person you delegate to), and costs are falling 10x a year. YC figures the consumer tipping point is right around the corner.
- For founders, it's a ready-made map of what to build. For everyone else, it's an official signal of where the world's top accelerator is putting its money in 2026.
First, what exactly is this YC list
YC is the world's best-known startup accelerator — Airbnb, Stripe, and Coinbase all came out of it. It makes money by betting on the next company like those. So every so often it publishes a list of the directions it most wants to fund, with one simple subtext: if you're building this, come get funded.
The list is called RFS (Requests for Startups). The fall 2026 batch has 13 entries, and there's a single theme: get AI out of the screen and into the real world. Each one is written by a YC partner or a working founder, spelling out the pain point, the opportunity, and what they'd like someone to build.
So the list works on two levels: it points founders toward opportunities, and it's YC publicly shopping for companies. These are directions it's genuinely excited about — build one and come get funding and support. YC also says up front that the list is just a way of sharing what's on its mind: even if you build none of these, you can still apply.
All 13 directions, laid out on one table
Thirteen sounds like a lot, but they group into four buckets. Scan the whole map first to get your bearings.
Why this batch is all about the physical world
String the 13 together and a through-line emerges: for the past few years, AI's battlefield was almost entirely inside the screen — writing code, generating text and images, answering questions in a chat box. This batch, YC pushed its chips beyond the screen.
Defense, moving compute out to sea, an OS for job sites, gathering physical-world data — all of these have to grapple directly with hardware and the offline world. Even the consumer-and-society entries carry a strong offline flavor: elder care depends on in-home monitoring and robots, education is aimed at real children, and "prove you're human" tackles exactly the trust that's breaking down both online and off.
Get AI to tutor kids like a private teacher
For this batch's headline direction, the author leans on a book. In the science-fiction novel The Diamond Age, there's a magical book that adapts to its reader, called A Young Lady's Illustrated Primer. It looks like a book for learning to read, but it's far more: it revolves entirely around one girl, using stories drawn from her own life not just to teach her letters but to teach her to think and reason, growing alongside her — and as she changes, the book changes with her.
The best education has always been one-on-one tutoring. Aristotle taught Alexander. But that privilege has always belonged to a tiny few. For the first time, AI makes giving every child a private tutor possible in principle.
What makes a good tutor great? Not drilling problems alongside the kid. It spends years figuring out how a particular child's mind works, then teaches the things drills can't: how to think, how to reason, even how to grow wise. That's exactly what today's education software is missing:
- One pace, one question bank for all
- Really just digital drills and flashcards
- Doesn't know who your child is
- Adapts to each kid like a tutor
- Teaches reading, writing & math at tutor quality
- Grows with the child; helps teachers, doesn't replace them
We're a long way from actually building the book from The Diamond Age, but you can start today. What YC wants now is a product that teaches kids reading, writing, and math like a devoted tutor, at a consumer price everyone can afford.
It'll start as a small thing parents pay for to shore up their kids' basic skills. But that's just the entry point. A company that gets it right can grow all the way toward that magical book. Even a small fraction of that vision realized would be enough to change society profoundly.
The Secretary of the Army comes asking directly
The most unusual of the 13 comes from the sitting U.S. Secretary of the Army, Daniel Driscoll. His reasoning: warfare is at an inflection point, and the military's old procurement process can't keep up with today's shifting threats. So they "tore up the old procurement playbook" and came straight to founders to ask.
Modern warfare calls for commercially developed, modular open-system solutions — which is exactly why we're tearing up the old procurement playbook.Daniel P. Driscoll, U.S. Secretary of the Army
What he wants is specific, and for the first time it puts the military's math plainly in front of founders:
Two terms are worth defining up front:
Move data centers out to sea
This one sounds the most far-fetched, but it's aimed at a real problem. It starts from a genuine bottleneck: AI is running low on compute, and data centers are running low on power and land.
The trouble is building data centers on land: permitting drags on for years, you have to line up enough power, and a single objection from local government can kill the whole thing. Communities increasingly resist the land use — but the ocean is wide open. And the sea happens to solve all three chokepoints at once:
- Permitting drags on for years
- Short on power, short on land
- One local-government veto can kill it
- Skips land permitting (solves land & approval)
- Ample sun, solar power on-site (solves power)
- Seawater is a natural heat sink (solves cooling)
The author calls the vision "compute flotillas": fleets of standardized, modular vessels working in concert, used as one unified global cloud. Bold, yes — but aimed squarely at what's actually throttling AI right now: power and land.
Three more bets in consumer and society
Beyond the education headliner, the consumer-and-society bucket has three more directions. What they share: each faces an enormous market that's been badly underserved.
Every big platform shift births a consumer giant: the web gave us Google, mobile gave us Instagram. Yet three years into AI, the only new icon on your phone is still ChatGPT.
Why now? Intelligence is finally good enough (you can treat an Agent like a person you delegate to), and it's about to get cheap enough: making this magic run today burns roughly $1,000 in tokens per user per month, but that number drops 10x a year. Follow the curve down and the consumer tipping point is nearly here. Whoever moves now takes the ground.
By 2030, 1 in 5 Americans will be over 65. There aren't nearly enough caregivers — the shortfall in care jobs runs into the millions, and 53 million family members already do this work unpaid. Yet almost no technology is truly built for older people; even Alexa and Google Home leave them baffled.
AI makes a new class of products possible: voice interfaces that can actually hold a conversation, monitoring that helps seniors live safely and independently, robots that lend a hand with chores, and software that helps families coordinate care, appointments, and emergencies. It's one of the world's largest and most neglected markets — and it grows every day.
A real case: a finance worker joined a video call with "the CFO and several colleagues" and afterward wired out $25 million. Only later did it emerge that everyone on that call except him was a deepfake (an AI-swapped face and voice). Voice cloning and fake video calls are getting cheap and convincing fast, and scams like this are exploding.
The scary part: we no longer have a good way to tell whether the other party is a real person. Seeing a face and hearing a voice used to be enough — not anymore. Every trust signal we have was built in an era when impersonating someone was expensive, and that era is over. What YC wants is to rebuild the internet's trust layer: knowing the person on the other end of a call, message, or transaction is a verified human — ideally without anyone handing over their privacy. Whoever pulls it off becomes that layer: every bank, every app, every video call runs a check here before trusting the other side.
Two more from the physical world
Back to the physical-world bucket: beyond defense and sea compute, two more directions take aim at "offline work" and "real data."
80% of the world's workforce doesn't sit at a desk. Yet the software used on job sites, in repair, and by fleets hasn't changed much in 20 years — it does the same handful of things over and over: dispatch work, track people, manage assets, bill customers.
AI upends all of it. There are now three kinds of worker on a job site: AI Agents that quote and schedule, robots that actually do the work, and people wearing devices that record everything. Old software was never built to manage all three at once. Should a job go to an Agent, a robot, or a person? How do you handle safety when people and robots work side by side? These are all new openings. And these industries spend 10 to 100x more on labor than on software.
AI has trained on software, language, and image data to superhuman levels, but physical-world data is still scarce: we're still gathering sparse readings with a pile of sensors designed for people, not for AI.
As foundation models get stronger and sensors get cheaper, densely gathering physical-world data is viable for the first time. Some use robots to collect data in places people can't reach and build predictive models; the author's own company, Sorcerer, uses autonomous weather balloons to gather atmospheric data the U.S. government uses to improve forecasts. The biggest industries — energy, agriculture, logistics, construction — still run on limited data and gut feel. Dense enough data lets you model; a model lets you control.
Four foundations the software era still needs
There's a whole bucket aimed at building infrastructure for AI-era software — all of it work that "suddenly became possible now and wasn't worth doing before."
Using an Agent to build yourself a little tool that only one or two people use is now easy and even fun. But once it's built, deploying and sharing it is still hard. Clouds like AWS and Azure are designed for big software that can carry lots of users — and the price of that is complexity.
A cloud built specifically for small software could cut away most of that complexity, making sharing a little tool as easy as sharing a Google Doc. There are hard parts, of course: every company wants a custom runtime, permissions and auth are tricky, and letting non-technical people safely share arbitrary code is thorny.
The best tools of the past 20 years won by going multiplayer: Google Docs beat Word, Figma beat Photoshop. But AI is still a single-player game: you open a chat box, send a prompt, and the answer lands in a box only you can see. To use it with a colleague, the most you can do is share a read-only link to the transcript.
When Agents start running long tasks over hours, days, even weeks, that work was never meant for one person. Anyone on the team should be able to drop into the same Agent session at any time — watch it work, correct its course, take the work over, just like collaborating with a human colleague. Engineering, sales, support, legal, analytics, marketing: anywhere a group of people gathers around one problem, there should be a shared Agent.
Financial compliance today is still cobbled together from Excel, a pile of point tools each covering one slice, plus expensively paid people. Companies hire a chief compliance officer and stitch together a stack of point tools just to figure out "what's actually going on underneath." Expand into a new market and complexity doubles, with compliance costs rising faster than revenue.
But most of compliance work — watching for regulatory changes, flagging anomalies, generating reports, keeping an audit trail — is exactly what AI does fast and cheap. The best version goes further: not content to automate existing processes, it reimagines what compliance should look like when AI is the default labor. The most fragmented corners — multi-state licensing, renewals, audits — are where the pain concentrates most.
Over the past year the author dealt with more than 50 API providers and found a common ailment: "communication" between APIs is broken. Breaking changes ship with almost no warning, useful new features launch quietly and go unnoticed, and nobody reads the changelog. Back at AWS, he saw more than 30% of service outages caused by an external API or dependency changing without anyone noticing.
Since people already trust tools like Claude Code and Devin with access to their codebases, an API provider shouldn't just "notify" you of a change — it should apply the change for you directly. The moment Stripe ships a breaking change or a new feature, an Agent should sweep the customer's code, find the affected spots, fix them, and submit the change to you (open a fix PR). It could be each provider's own updater Agent, or a neutral cross-provider service — like Dependabot (the GitHub tool that auto-upgrades your dependencies), except for APIs.
Why now is the best time to build crypto
The last one is the most counterintuitive. The picture really doesn't look good: coin prices are down, hot narratives are fizzling one by one, and plenty of people are heading for the exits. Yet YC says it's more bullish than ever. The reason has nothing to do with price — it's all about who's still in the game.
- Prices detach from reality, bubble noise everywhere
- You fight scammers promising infinite yield for users
- It attracts people who are just there to cash out
- No one's competing, so you can build in peace
- The founders who stay want to build, not cash out
- Regulation is finally clear, stablecoins adopted by big institutions
YC says it has already funded more than 100 crypto companies and expects that number to climb sharply. It even reckons that in the future nearly every YC startup will use crypto's underlying plumbing (from fundraising to payments), most of them without even realizing it. The directions it names include: fundraising, new stablecoins and their applications, Agents using crypto networks directly as financial rails, trading, institution-facing products, and blockchains that are both scalable and privacy-preserving.
How you should read this list
Reading through all 13, don't treat them as a required answer key — YC's bets miss plenty of the time. Their real weight is this: an institution that lives by betting on the future has put its money-backed intentions out in the open. At the very least, you now know which way the smart money is pushing in fall 2026.
And on that direction, all 13 nearly speak with one voice: AI's liveliest battlefield is moving out of the screen and into the real world beneath our feet — defense, the ocean, job sites, elder care.
YC names the 13 directions it most wants to fund: this batch, the weight shifts beyond the screen
The world's best-known startup accelerator lays out 13 directions at once, with a single theme — get AI into defense, elder care, job sites, and the ocean — and, for the first time, includes a request from the sitting U.S. Secretary of the Army.
↓ The whole story in one page · with one animated chart
YC is the world's best-known startup accelerator — Airbnb and Stripe both came out of it. Every so often it publishes a list of the directions it most wants to fund.
The list is called RFS (Requests for Startups). The fall 2026 batch has 13 entries, each written by a YC partner or a working founder, spelling out the pain point, the opportunity, and what they'd like built. The subtext is simple: if you're building one, come get funded — and if you're building none, you can still apply.
Thirteen sounds like a lot, but they group into four buckets. Scan the whole map first to get your bearings.
String the 13 together and a through-line emerges: for the past few years AI's battlefield was almost entirely inside the screen; this batch, YC pushed its chips beyond it.
- Coding, generating text & images
- Q&A in a chat box
- Output stays on screen
- Defense, compute moved out to sea
- Job sites, elder care, education
- Landing in hardware & offline
This batch's headline direction is called The Primer, borrowing its image from a magical book in the science-fiction novel The Diamond Age that adapts to its reader — a book that revolves around one child, grows up alongside her, and teaches her both to read and to think.
The core is one line: the best education has always been one-on-one tutoring (that's how Aristotle taught Alexander), but the privilege has always belonged to a tiny few. For the first time, AI makes giving every child a tutor possible in principle.
- One pace, one question bank for all
- Really just digital drills and flashcards
- Doesn't know who your child is
- Adapts to each kid like a tutor
- Teaches reading, writing & math, tutor-grade
- Grows with the child; a helper for teachers
It starts as a small thing parents pay for to shore up their kids' basics, but that's the entry point — get it right and it can grow all the way toward that magical book.
The biggest news among the 13 comes from Secretary of the Army Daniel Driscoll. His reasoning: warfare has hit an inflection point, the military's old procurement can't keep up, so it "tore up the old procurement playbook" and came straight to founders to ask.
What he wants is spelled out plainly: low-cost interceptors that drive the "cost per kill" (what it costs to knock down one target) as low as possible — enemy drones are cheap to build, and you can't trade a sky-high missile for one every time; plus sensors, software, and drones that plug into the military's open-system architecture (open interfaces, parts swappable like Lego). The promise: capital, and a proving ground where you can scale freely.
The directions, figures, and judgments here come from YC's official RFS page and the individual authors — a public call for startups with a recruiting angle, not a neutral third-party assessment. Key facts like the Secretary of the Army's role have been independently verified by us.
the money going?
making text & images —
all in the screen.
elder care, job sites.
offline?
- × text-to-image
- × text-to-video
- × the models themselves
posting on an accelerator?!
to down a cheap drone?!
to cut the unit price.
of all times?!
to delegate to AI like a person.
is moving into the real world.
Members can copy or download the full Markdown, and let their AI subscribe and ask.
Found this useful? Leave your email — every new explainer, the moment it's out. Unsubscribe anytime.