The State of Startups in 2026

Summary · 4 min read instead of 36 min

The YC partners use batch data to argue that startups are moving from bits to atoms: hard tech has gone from 8% to 20% of the batch, driven by robotics, defense, manufacturing, compute and power. They say AI coding lowers the engineering burden for hardware companies and lets startups that do whole jobs end to end grow faster. Median revenue by the end of the batch has risen from about 8K to about 20K monthly, and some companies reach seven figures within the batch. Solo founders now make up nearly one fifth of accepted companies, and the partners say experienced founders who know what to build are doing especially well.

Minute by minute

1. 0:58

   Hard tech jumps to 20% of batch. Diana Hu says her analysis of recently accepted companies shows hard tech rising from 8% to 20% of the batch. She also says the median company by the end of the batch now reaches 20,000 in monthly revenue, up from 8K in the past. The median company starts at zero revenue and pre-product.

2. 2:12

   Physical categories have tripled or quintupled. Hu gives the category shifts: robotics from 1% to about 6-7%, industrial manufacturing from about 4% to 10%, and defense from about 1.5% to about 5%. Semiconductors and photonics went from about 1% to close to 4%, and power from 1% to close to 3%. Tan calls this the age of the machine.

3. 3:24

   AI makes hard tech more feasible. Jared Friedman says one in six founders in the current summer batch has a PhD, far above the historical share. Tan says code generation means hardware companies no longer need a thousand great software engineers, only a small team, which changes the economics. He adds that smarter models are speeding up scientific research for deep tech startups.

4. 5:15

   Space, defense and metal drive atoms. Hu names three macro trends, starting with SpaceX's IPO inspiring founders to build in space, with examples such as Exosat and Beyond Reach Labs. Tan describes defense startups Icarus and Nine Mothers and says new companies can build what the defense primes cannot. Friedman says Nox Metals is rebuilding US metal supply and growing at a software rate because new defense startups need metal faster than old suppliers can deliver.

5. 9:44

   Compute is a physical infrastructure problem. Hu says A100 GPU prices per hour are rising because demand exceeds supply, which has pushed startups into data center buildout, power and new silicon. She cites Bot, which is building hardware around ternary model representation because each Nvidia generation has lowered floating point precision. Dipole Labs is building a fully optical switch, since electronic switches are a data center bottleneck.

6. 12:30

   Robotics nears its ChatGPT moment. Hu says people in the industry feel a ChatGPT moment for robotics is coming but is not here yet. Friedman says hard tech was hard to fund before, when VCs said they only did B2B SaaS, and he is glad venture is returning to its roots. Harj Taggar notes that investor interest shifted quickly, around the winter batch, as SaaS stocks fell.

7. 14:55

   Software moats may hold as harnesses. The partners note that many SaaS stocks have recovered, with Salesforce and Snowflake as examples. Tan argues that systems of record must become harnesses where people do their work, or else lose their data moat. He also says agents will use software far more than humans do, so products built for agents look like the right kind.

8. 17:49

   End-to-end AI products grow fastest. Hu says companies doing full end-to-end work have gone from 10% to over 25% of the batch, and that is where revenue growth is concentrated, in areas like insurance brokerage and medical billing. She says some companies now go from zero to seven figures in three months, which used to take about 18 months or more. Examples include Juicebox, whose agent product contacts candidates and schedules interviews.

9. 22:46

   RL environments are a hidden boom. Friedman says YC has funded more than a dozen companies in two years that each make over $10 million a year selling data or RL environments to AI labs, some reaching hundreds of millions. Tan names Afterquery and DataCurve. Hu says the big labs reportedly spend about a billion dollars on this and that eight- and nine-figure deals are emerging for robotics data.

10. 26:55

    Robotics will need specialized models. Hu hypothesizes that robotics foundation models differ from LLMs because they model 3D physical space and must respond in real time. She says a model fine-tuned on custom data for one setting, such as data center cabling, may work better. Tan says YC companies using Physical Intelligence models all fine-tune them rather than using them out of the box.

11. 29:24

    Solo founders reach nearly one fifth. Hu says solo founders went from about 5% of accepted companies a year ago to over 18-19%, the highest spike seen. Tan argues knowing what to prompt and build is now harder and more valuable than coding. Friedman says co-founders are still valuable and expects successful solo founders to add them as the company grows, as Instacart and Coinbase founders did.

12. 32:45

    Experienced founders are having a resurgence. Tan says people who have been around the block do better now, citing Peter Steinberger, who is in his early 40s and knows what to build. Taggar suggests managing coding agents resembles managing people, which helps those who led engineering teams, such as Toby from Shopify. Tan's advice to would-be founders is to start prompting, and he says each new model makes agents noticeably smarter.

Published Sep 17, 2026.