YC's Hard Tech Share Doubled to 20% in a Year — Robotics and Defense Are Leading the Charge

Source: Y Combinator | Published: 2026-09-17T17:58:14Z

YC startups are now hitting seven-figure monthly revenue in just three months, down from 18 months — and hard tech has grown from 8% to 20% of the batch in a single year.

Robotics, Defense, Manufacturing, Semiconductors, Power: Five Curves Breaking Out at Once

Drill into that increase and the numbers get more concrete. Robotics companies went from 1% of a YC batch to 6–7%. Industrial manufacturing and companies building things on American soil went from 4% to 10%. Defense went from 1.5% to 5%. Semiconductors and photonic chips went from 1% to nearly 4%. Power infrastructure similarly went from 1% to nearly 3%. Across every one of these physical categories, growth has been three to five times over.

In this summer's batch, one in six founders holds a PhD. That's not a coincidence — silicon photonics requires genuine research depth.

Defense Startups Are More Than "Selling to the Government"

Jared Friedman, YC Group Partner, has invested in two companies that illustrate the point. The first is Icarus, which builds solar-powered reconnaissance drones capable of sustained coverage that also double as communication relays. In the era of drone warfare, maintaining contact with ground-based drones has become a genuine operational bottleneck. Icarus has already secured seven-figure contracts. The second is Nine Mothers, focused on counter-drone defense — essentially a computer vision-guided shotgun turret.

His argument for why this matters: special forces soldiers represent years of elite training, and the US simply doesn't have many of them. A cheap drone can pose a real threat. Protecting those soldiers with software-driven hardware isn't optional.

Defense is also producing a wave of "dual-use" companies — supply chain players selling to both the private sector and government simultaneously. Nox Metals is one example: it's rebuilding American metal manufacturing out of vacant legacy factories in Detroit. Its customers are the new wave of defense tech startups — companies that move at a pace traditional suppliers simply cannot match. That makes Nox Metals the "Stripe" of the new defense ecosystem: a modern vendor built for companies that need to iterate fast.

The Bottleneck in Data Centers Isn't the GPU

The hourly cost of Nvidia's H100 GPUs is rising, not falling — a rare phenomenon in hardware history, signaling that demand is outpacing supply by a wide margin.

That gap is generating a new category of companies. Dipole Labs is building all-optical switches: when GPU A and GPU B communicate, they currently go through electronic switches, and those switches can no longer keep up with GPU throughput — making them the actual bottleneck for many data center workloads. Dipole wants to replace that entire layer with photonics. Another company, Baud, is going even deeper, exploring new hardware architectures based on ternary representation. The logic: every generation from Nvidia's A100 to H100 to B300 has reduced floating-point precision, and it turns out LLMs don't need that much precision — FP2 is sufficient in certain contexts.

The Most Underestimated Business: Selling Data to Foundation Models

There's a structural information asymmetry at play here: the companies doing this well have almost no incentive to talk about how well they're doing.

YC recently pulled data from the past two years and found it had invested in over a dozen companies, each generating more than $10 million in annual revenue — by selling data or reinforcement learning environments to major labs. Some have already scaled to hundreds of millions of dollars, most with only two or three years of history.

Reported total annual spending by major labs in this category runs into the billions. That's not a widely known figure.

The data itself is fragmenting. Finance requires specialized RL environments that demand deep vertical expertise combined with systems engineering. Robotics demand is climbing fast — labs want to advance AI's ability to operate in the physical world, and they need egocentric-perspective data and long-horizon task data. Current batch companies are building industrial operations networks around the world to capture this, with some using on-the-ground local operators to collect it.

Monthly Revenue: From $8K to $20K

In the past, the median YC company went from zero revenue at batch entry to $8,000 in monthly revenue by Demo Day. That number is now $20,000.

The outlier cases are more striking: this year, companies have gone from zero to seven-figure revenue within the three-month batch. The same milestone used to take 18 months.

Diana Hu, YC Managing Partner, analyzed what's driving this acceleration: many companies are no longer building "point-feature solutions" — they're building "end-to-end, do the whole thing" products. That category has grown from 10% of the batch to over 25%. AI recruiting agents, clinical intake, end-to-end medical billing — let the agent do the work, rather than giving a person a tool and letting them do it.

Juicebox is the concrete example. It started as an LLM-powered talent search: describe the candidate profile, get a list. But reaching out was still on you. They recently launched an agent product: search, outreach, interview scheduling — all handled by the agent. Per-account revenue is projected to double or triple. Recruiters, meanwhile, end up focusing on what still requires a human: cultural fit and judgment — the part that hasn't been automated yet.

Solo Founders: From 5% to 19%

YC tracks this internally: the share of accepted companies with a solo founder has risen from roughly 5% to 18–19% — nearly one in five.

There's historical precedent. Instacart's Apoorva and Coinbase's Brian Armstrong both entered YC as solo founders. Parker Conrad came in solo too, and Jared later interviewed his future CTO, Prasanna Sankar, separately. In that era, the bar for a single person to simultaneously figure out what to build, sell it, and execute on it was extraordinarily high — only a rare few could manage all three.

Now the bar for "build it" is dropping. Knowing what to build has become the genuinely scarce capability.

This explains why experienced founders — people in their 30s, 40s, even 50s — are having a resurgence right now. Garry Tan, YC President and CEO, pointed to Peter Steinberger: founder of PSPDFKit, an early and deep user of AI tools — someone who knew exactly what to do and went and did it. In this environment, industry experience is an advantage: you know where the problems are, what customers are thinking, where the genuine demand is.

Managing a fleet of coding agents and managing an engineering team aren't actually that different. People with years of engineering management experience already know how to decompose tasks, set boundaries, and evaluate output quality.

SaaS Won't Die, But It Has to Become the Front Door

Salesforce quietly hit all-time highs amid all the "SaaS is dead" discourse. Snowflake just delivered a blowout earnings report. Jared's read: these two facts aren't contradictory.

Agents will call software more frequently than humans do. If you're a system that agents want to use, you're actually in a stronger position — because usage will be far higher than before. Part of Salesforce's growth is agents using Salesforce.

The challenge is that every "system of record" now faces a choice: open up MCP interfaces so data can be called from anywhere — but surrender the data moat in the process — or become the front door, the place where both humans and agents do their work rather than just read and write data. Slack is taking the second path; Benioff has built AI workflows directly into Slack. Which systems successfully make this transition remains an open question.

The next "front door war" is starting. Claude Code wants the position. Codex wants it. Salesforce is trying. This won't end like the browser wars with a single winner, but which environment agents most naturally operate in will be a real dimension of competition.