The State of Startups in 2026

Why does it matter? Because the founders winning right now look nothing like the founders winning five years ago.

YC's internal data from the last 18 months reads like a dispatch from a different industry than the one most founders think they're in. Hard tech more than doubled its share of the batch. Solo founders tripled. A billion-dollar market materialized in near-total secrecy. Some companies hit seven-figure revenue inside a single three-month batch.

• A dozen-plus YC companies each clear $10M+ annually selling data or RL environments to AI labs — most deliberately stay quiet about it
• Hard tech jumped from 8% to 20% of the YC batch in 18 months; robotics alone went from 1% to nearly 7%
• Solo founders hit 19% of accepted companies, up from 5% — AI coding broke the dependency on a technical co-founder
• Median batch revenue nearly tripled from $8K to $23K MRR; some companies cleared $1M in 90 days flat

There's a billion-dollar market hiding in plain sight — AI labs are spending it on companies you've never heard of

The most lucrative opportunity in AI right now isn't building the next app. A dozen-plus YC companies are each clearing $10M+ annually — in many cases, hundreds of millions — selling data and RL environments to frontier labs. "Reportedly, the big labs are spending about a billion dollars on this. It's not a very known fact."

The silence is deliberate. Once you have a nine-figure contract with Anthropic or Google, the last thing you want is competitors reverse-engineering your playbook. After Query and DataCurve get named on the podcast; most of the others would prefer not to be.

The mechanics: labs need proprietary data they genuinely can't synthesize. Finance-specific RL environments. Industrial teleoperation footage. Real-world robotic task data. Companies like Praxis Robotics and DeepReach are closing eight and nine-figure deals by supplying exactly that.

This category barely existed when Scale AI got funded in 2016. It's now a billion-dollar market that grew in near-total silence. If you have deep domain expertise producing data that's hard to replicate — industrial, medical, robotic, financial — there's an active buyer with a massive budget and very little competition.

AI removed software talent as the limiting reagent for hard tech — a small expert team can now build what previously required thousands

Palmer Luckey made this point about Anduril, and it keeps surfacing. Building full-stack hardware used to require a roster of elite software engineers that startups simply couldn't compete for against big tech. That constraint is structurally gone.

"Even three or four years ago, top-tier software engineers were one of the limiting reagents for full-stack hardware. You still need one or two, but you don't need to hire a thousand people."

The real bull case for hard tech's resurgence isn't that investors got tired of SaaS multiples. It's that "the super smart models that we have now are accelerating scientific research and making it possible for startups to have bigger research breakthroughs earlier." A small team with serious domain expertise and AI coding tools can now attack problems that previously required a large company's headcount and capital.

The PhD numbers make this concrete: one in six founders in the current YC summer batch holds a PhD — well above historical rates. When the execution barrier drops, domain knowledge becomes the scarce input. The people who spent careers building it are in the best position they've ever been.

Zero to $1M in 90 days — automating the whole job commands ten times more than software that merely records it

Eighteen months was once the benchmark for reaching seven-figure revenue. Some companies are hitting it inside a single YC batch.

"We have companies breaking from zero to seven figures in revenue during the batch. That is in a span of three months. In the past, that would have taken 18 months or more." Median batch MRR jumped from $8K to $23K. The average is higher still.

The explanation isn't hype. It's a product-value shift that was always going to arrive: "If they automate the whole job, they will actually just be more valuable. Then some like system of record that tracks the job, it doesn't do the job." Juicebox, an AI recruiting tool, started as LLM-powered candidate search — useful, but still a human-executed workflow. The agent product now contacts candidates directly and will eventually schedule interviews. Per-account revenue is set to double or triple. Recruiters are enthusiastic: mechanical outreach gets automated away; culture-fit judgment stays human.

Full end-to-end workflow companies grew from 10% to over 25% of the YC batch. Insurance broking, clinical intake, medical billing — agents running whole jobs, not logging them. Enterprises write big checks early because the value proposition requires no explanation.

Hard tech more than doubled its share of YC — three macro forces are still accelerating, not peaking

The jump from 8% to 20% wasn't one trend. It was three converging simultaneously.

The SpaceX generation wants to build in space. Exosat (a sovereign Starlink alternative) and Beyond Reach Labs (orbital solar panels for data centers in space) are in the current batch. Defense followed: from 1.5% to 5% of the YC batch. Icarus — a solar-powered drone providing overwatch and comms relay — closed seven-figure government contracts. Nine Mothers built anti-drone turrets to protect Special Forces behind enemy lines. A new administration approach to procurement, moving away from cost-plus primes, created a real opening for startups.

Compute became a physical-world problem. A100 GPU hours are actually appreciating in price — older chips cost more than they did because demand outstrips supply. That's driving startups across the full infrastructure stack: construction, cooling, power, photonics. Dipole Labs is building fully optical data center switches to eliminate the electronic interconnect bottleneck slowing GPU clusters.

Robotics hasn't hit its ChatGPT moment yet — but Astra's RKGI benchmark performance jumped from low single digits to 60-70% of tasks in months. "All these numbers across the physical atom stacks have somewhere triple or quintupled." Companies building the infrastructure now are positioning for that threshold.

Solo founders tripled their share of YC — AI snapped the co-founder dependency for technical execution

The traditional co-founder pairing had a clear logic: no single person could simultaneously sell, lead, and build at world-class level. AI coding tools broke the third leg.

Solo founders went from 5% to 18-19% of accepted companies — "the highest spike that we've seen" — and the trend is still moving.

"Knowing what to prompt and knowing what to build is so much more difficult and valuable than just knowing the CTO being able to code the thing." Exceptional solos existed before — Apoorva Mehta at Instacart, Brian Armstrong at Coinbase, Parker Conrad at Rippling all entered YC alone. But those cases required being exceptional across every dimension simultaneously. The execution ceiling has now dropped enough that domain knowledge and product judgment can carry a founder to traction without a world-class technical partner in the seat.

The pattern emerging from the data: start alone, build to traction, add co-founders once direction is proven. Equity dynamics shift accordingly — the founder who got the company moving holds a stronger position when bringing people in. YC still thinks co-founders improve your odds significantly. The argument isn't against them. It's that waiting for the right technical co-founder when you have strong domain knowledge and taste is now optional.

Founders in their 40s and 50s are having their best moment in a generation — taste and pattern recognition now outrank raw coding speed

The conventional startup wisdom — that youth and technical execution velocity are the decisive advantages — is inverting, and the mechanism is precise.

"Managing coding agents is, in some ways, not that different from managing people. People who have had whole careers managing engineering teams actually take to this super well and can spin up huge teams of coding agents and manage them more effectively than even a really smart 19-year-old without those years of experience."

Peter Steinberger is the example cited: early 40s, former dev manager, serial startup veteran, early AI convert. The edge isn't speed. It's taste — knowing where the dragons are before you start, recognizing what's worth building without spending six months discovering it isn't.

For anyone who's spent years in a domain and has strong opinions about what's broken: the execution barrier just dropped to meet the expertise you already built. Domain knowledge and management pattern-matching, previously locked out of the founding moment by raw coding requirements, are suddenly the inputs that matter most.

SaaS systems of record must become AI harnesses — releasing an MCP without making that shift hands your data moat to competitors

Salesforce's blowout recent performance looks like vindication for legacy SaaS. The explanation reveals a trap for anyone who reads it as simple confirmation.

"Agents will use software a lot more than humans will. And that seems to be driving Salesforce growth." The moat isn't the software itself — it's the data locked inside it. Agents need that data, so legacy systems of record temporarily benefit. But only if they adapt fast enough.

"If you are a system of record, you either will be preyed upon — you release an MCP, the data goes elsewhere, switching becomes trivial — or you kind of have to be a harness. You have to be the way people not just read and write, but actually do their work inside your system of record."

Slack AI is Salesforce's opening move in what looks like the next platform war. The harness market may support multiple winners — unlike browser wars, there's no obvious gravity toward one dominant player — but the structural logic is clear: the platform where agents do work captures the value that previously went to the humans doing it.

For founders: find the vertical where the system of record is passive and humans still execute the job. That gap is where the pricing power lives.

Every constraint that kept the wrong founders out collapsed at once — that doesn't happen often

What the data actually shows is a simultaneous failure of multiple gatekeeping mechanisms: co-founder requirements, software talent costs, hard tech unfundability, age penalties, capital barriers to atoms-based businesses. They weakened together, in the same 18-month window.

That's historically unusual. Convergences like this tend to produce outsized companies — and they tend to close as early movers lock in defensible positions. The robotics ChatGPT moment, the lab data market, the harness wars, the defense tech buildout: all still early enough that positioning now matters.

The window is open. It won't stay that way.

---

Topics: YC, startups, hard tech, robotics, defense tech, AI agents, SaaS, solo founders, experienced founders, data labeling, RL environments, compute infrastructure, founder trends, 2026