Hard Tech Is One-Fifth of the YC Batch: Software Companies Are Being Eaten by Agents
In YC's latest batch, hard tech companies went from 8% to 20%, with robotics, defense, semiconductors and power all doubling or more; meanwhile software companies are shifting from selling tools to selling "getting the work done," and median revenue rose from 8K to 20K.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Hard tech went from 8% to 20%
Diana analyzed companies YC admitted over the last 18 to 12 months, and hard tech's share went from 8% to 20%. Broken out: robotics from 1% to 6-7%, industrial manufacturing from 4% to 10%, defense from 1.5% to 5%, semiconductors and photonics from 1% to nearly 4%, power infrastructure from 1% to nearly 3%. Across the physical atom stack these numbers generally tripled to quintupled. Over the same period, the median YC company had zero revenue when admitted; in past batches median monthly revenue was about 8K by the end, and now it's 20K.
— DianaThe share of technical founders is rising
Jared mentioned that in the current summer batch, one in every six founders has a PhD, far above historical levels. The reason is that working on directions like silicon photonics usually requires a strong research background. Gary added that hard tech used to be hard because supply chains and software engineering were the limiting factors; now with codegen, top-tier full-stack hardware no longer requires hiring a thousand excellent engineers — one or two people, or a small team, is enough, and that changes the economics. The real bull case isn't that investors are avoiding software, but that the models themselves are accelerating scientific research, letting startups make research breakthroughs earlier.
— JaredDefense and space have become the new narrative
Gary said three macro trends are driving atom growth. First, after SpaceX's successful IPO, a generation of founders wants to build space companies, like Exosat doing sovereign Starlink solutions and Beyond Reach Labs doing satellite solar panels. Second, this generation of founders grew up in an environment where war was heavily discussed, and they want to do defense. Icarus is building a solar-powered U2 reconnaissance aircraft that can do surveillance and communications, and has already won a seven-figure contract from the new Department of War; Nine Mothers does counter-drone defense, a shotgun turret with CV, and special forces are procuring it. Third, compute itself is a heavy physical process: the A100's hourly cost is rising because demand is too great and supply is insufficient.
— GaryHardware companies are running at software growth rates
Nox Metals is bringing metal manufacturing back to the US, rebuilding the metal supply chain in empty Detroit factories. Its customers are emerging defense tech companies that need metal to build things, and existing suppliers are mostly slow-moving old businesses that can't keep up with the pace of new defense companies. Gary analogized to the early Web 2.0 era: new startups prefer to buy from new startups because they can collaborate at the same speed, like using Stripe instead of traditional credit card providers. PG tweeted that Nox Metals is growing at software growth rates.
— GaryThe compute stack digs from chips down to power
Diana said GPU prices are反常: the A100 is already old, yet its hourly cost is still rising because demand exceeds supply. Around data center buildout — from site selection and construction, software, the actual buildout, to power supply solutions — there are startups working on all of it. Going down to core compute silicon: Lamb Labs is building a new processor, Bot is building a new custom hardware architecture using ternary representation, because as Nvidia went from A100 to H100 to B300, floating-point precision dropped generation by generation, and LLM architectures don't need full-precision floating point — even FP2 can work. Dipole Labs is building all-optical switches, because electronic switches can't keep up with GPUs and have become the bottleneck in many data centers.
— DianaSystems of record either get eaten or become the harness
Gary proposed: if you're a system of record company, you either get plundered — once MCP is released, the data moat disappears and switching becomes easy; or you become the harness yourself, the place where people not only read and write but get work done. He cited the RKGI benchmark as an example: the same model plugged into different harnesses performs differently, and model plus harness equals output. Some claim Astra with a custom harness exceeds 90% on RKGI 3, when a few months ago that number was in the low double digits.
— GaryThe share of software doing complete tasks has doubled
Diana said the share of YC-admitted companies doing full-stack end-to-end tasks went from 10% to over 25%. The difference is that agents actually do the work, rather than being point solutions like SaaS five to eight years ago that still needed a human to operate them. The fastest-growing ones are exactly these companies doing complete workflows, like insurance brokerage, clinical intake, medical billing. Jared added that it's been almost a year since agentic coding became truly usable (Opus 4.5), and these workflows are now blooming across the board; people are willing to pay more for software that "gets the work done."
— DianaFrom zero to seven-figure revenue in three months
Diana said that this year, for the first time, she saw companies go from zero to seven-figure revenue during the batch, over a span of only three months, where this used to take 18 months or more. Part of the reason is that they're solving real problems, plus agentic coding raises product maturity, and founders can run 20 coding agent sessions at once. Another fast-growing category is selling data or RL environments to labs, and many of these companies keep a low profile. Over the past two years YC has invested in more than a dozen companies with annual revenue over $10 million selling data or RL environments, many reaching hundreds of millions of dollars, with the companies only a few years old.
— DianaIn their own words · checked verbatim
the number of heart tech companies that are in the bat. It has gone from 8% to 20%.
Diana1:04
having codegen means that suddenly even all the things that they do at Anderil uh can happen much much faster
Gary4:04
if you are a system of record you either will be prayed upon like you'll release an MCP and then maybe like the data you know you lose your moat around the data the data goes elsewhere like becomes very trivial to switch or you kind of have to be a harness
Gary16:15
people want their job just be done and are willing to buy software that just gets the job done
Jared19:19
we have companies breaking from zero to seven figures in revenue during the batch. And that is in a span of 3 months.
Diana22:20
just in the last two years, YC has funded more than a dozen companies that are each making more than $10 million a year selling data or RL environments to the labs and in many cases hundreds of millions of dollars.
Diana23:20
what a weird moment we are in history where you wake up in the morning you like wire up a new model and then these things that even a month ago you're just like why isn't it working it just starts working
Gary35:23
Figures
| Hard tech companies' share of the YC batch | from 8% to 20% | 1:04 |
| Robotics companies' share of the YC batch | from 1% to 6-7% | 2:04 |
| Industrial manufacturing companies' share of the YC batch | from 4% to 10% | 2:04 |
| Defense companies' share of the YC batch | from 1.5% to 5% | 2:04 |
| Semiconductor/photonics companies' share of the YC batch | from 1% to nearly 4% | 2:04 |
| Power infrastructure companies' share of the YC batch | from 1% to nearly 3% | 2:04 |
| Share of founders in the current summer batch with a PhD | one in six | 3:04 |
| Companies doing full-stack end-to-end tasks as a share of the YC batch | from 10% to over 25% | 18:19 |
| Annual revenue of YC companies selling data or RL environments to labs | over $10 million, many reaching hundreds of millions of dollars | 23:20 |
Glossary
- harness
- The software layer that wraps a model and lets it actually execute tasks; model plus harness equals output.
- RL environments
- Simulated or real interactive environments used to train models with reinforcement learning; a data category labs pay for.
- codegen
- Using AI to automatically generate code, reducing hardware companies' dependence on large numbers of software engineers.
- system of record
- The system in an enterprise that records core data, like Salesforce or Slack, whose moat comes from accumulated data.
- egocentric data
- Physical-world data collected from a robot's or person's own viewpoint, used to train robot models.
How to listen
Suited to people watching hard tech, AI agents and early-stage investing, especially founders and investors who want to read the YC batch's direction and are considering building hard tech or agent products.
The last stretch from 33:22 to 36:28, the chit-chat and encouragement section, is low in information density and can be skipped.