A quarter of the companies presenting at Y Combinator's latest Demo Day are building hardware, up 40 times from three to five years ago. Garry Tan, YC's president and chief executive, called it the atoms side of an AI boom that started in bits.
Everyone assumed cheap intelligence would make software easier and hardware no more accessible than before. Tan's numbers say hardware is where the biggest shift actually is.
"It's time to do atoms, guys."
Garry Tan runs Y Combinator, having previously co-founded the YC-backed Posterous and the venture firm Initialized Capital before returning to lead the accelerator itself.
I listened to the full segment so you can skip it. 20 minutes of audio, 9 minutes of reading.
Here are the 6 takeaways that matter.
👤 Guest: Garry Tan, president and chief executive of Y Combinator, who previously co-founded Initialized Capital and the YC-backed Posterous
🎙️ Hosts: John Coogan and Jordi Hays, who present TBPN live on X and YouTube every weekday
📰 Published: 10 September 2026 on YouTube (TBPN)
🔴 YouTube | 🟣 Apple Podcasts | 🔗 Show notes | ⏱️ 20 min | ✅ Time saved: 11 min
Key Takeaways
A quarter of this YC batch is hard tech, up 40x from three to five years ago
Tan credits AI for making atoms-based startups as tractable as software ones
YC's standard check is now $500,000 plus more than $1 million in cloud credits
OpenAI separately gives any YC company $2 million in tokens, no ask required
Tan wants the AI-safety debate redirected at shutdown strategy and provenance, not personalities
He said little tech's job is building the cybersecurity, not talking on podcasts
Defense contracts of seven figures now go to teams of two to four people
Those teams can scale to 20 or 30 people plus hundreds of agents, he said
He predicts 20-person AI-native startups will out-execute companies with thousands of engineers
Each of those 20 people, in his framing, does the work of 100
A new wave of data companies is beating labor-marketplace incumbents like Mercor without scaling headcount
More than 10 make over $10 million a year; the top two are in the hundreds of millions
1. Hard Tech Now a Quarter
Hays asked Tan to open with the numbers behind the batch before going into the wider landscape.
Two hundred companies are presenting, and Tan gave the mix. "About a quarter of them are hard tech. That's up 40 x since our low bar, maybe like three or five years ago."
Tan's explanation is that AI removed the reason hardware stayed niche. "Hard tech is in, it's big, and it's here because of AI." Intelligence, he said, is now on tap, and "it's time to do atoms, guys"
Coogan pointed to a biochip startup, Frontier, using neuronal cells, which Tan said had already appeared on the show once and gone viral before landing in this batch
Tan framed the acceleration as two-sided. AI makes new hardware ideas possible that weren't before, and it also compresses the traditionally slow, manufacturer-dependent hardware development cycle — getting quotes, managing supply chains, tracking inventory
2. Checks Now Include Tokens
Coogan asked what solving a hard technical problem used to require versus now.
Before, it took a specialist you had to hire. Tan said a two-person garage team's odds of having the person who "really understands some hard tech thing" in-house were "so close to zero," and the alternative was raising $100 million just to convince that person to join
Now a frontier model substitutes for that hire, at least at first. "You say, hey, look, this is what I'm trying to figure out and you're iterating with it. You actually have a demo that's 10 times better than what you could have done without hiring that person"
YC's own check has grown past cash. "We're giving you $500,000. What can you actually get done? You know, these days, it's $500,000 plus more than a million dollars in cloud credits."
A separate, unsolicited grant comes from OpenAI. "OpenAI will come and just give any YC company $2,000,000 worth of tokens." Tan said any team doing something hard should take it: "if you can't figure out how to turn $2,000,000 worth of tokens into a $100,000,000 worth of enterprise value." His follow-up: "What are we doing here?"
3. A Defense-Tech Tailwind
Hays asked what YC is seeing on distribution for hard-tech and defense startups, including selling within the YC network itself.
Tan said space and defense companies are increasingly customers of each other, citing Star Cloud as a buyer of other space-tech startups' output
YC is building shared infrastructure for its defense founders directly. "We're putting together a CRADA so that we can actually help all YC defense companies get in theater with real Department of War departments," meeting Pentagon officials at a DC conference the following week
The contract sizes have changed. John Coogan: "They're willing to give 7 figure contracts to teams of, like, two or four people who, like, suddenly become teams of 20 or 30 people plus, like, hundreds of agents"
Tan credited the current administration with recognizing the shift. "We got a drone war happening. We have an AI war," he said, arguing America can no longer just pay incumbent defense contractors cost-plus rates. He called it remarkable how many opportunities have opened up even after Palantir and Anduril, when a winner-take-all outcome once looked likely
4. 20 People Beat 2,000
Hays asked what trends Tan is seeing across software, given how easy AI has made it to build.
Tan's prediction is a wholesale replacement of legacy engineering organizations. "Any company that's got thousands of engineers who are doing it the old way, like, they're gonna be replaced by a YC startup or a company like that has maybe 20 people working there. And that those 20 people each will do the work of, a 100 engineers themselves."
He said the advantage compounds because new companies carry no legacy habits. AI-native startups founded in the last two or three years "don't have any of the bad habits that were developed over the... ZERP era," and will out-execute SaaS incumbents that don't rebuild the same way
On infrastructure demand, he separated the growth rates. SaaS grows 10x; inference, data centers, semiconductors and GPU and CPU demand are headed "up like a thousand x," which he said is still not priced in
5. Fix Provenance, Not People
Coogan asked whether regulatory-capture risk from the AI-safety debate threatens little tech. Tan used the moment to reframe the debate itself, referring to the AI-lab whistleblower story the hosts had covered earlier in the show.
He wants the conversation off the individuals and onto the mechanics. "I think we should be talking less about this Jacob Coxon guy. We need to be talking a lot more about what is actually happening with Hugging Face."
His specific asks: a shutdown strategy and verified provenance. He said he wants people asking what the shutdown strategy is, how to ensure provenance, and "where is this agent actually located?" — not debating personalities
He named it as a business opportunity for little tech, not just a policy problem. Citing a former OpenAI chief scientist's post about "Neo Cloud security," Tan said he hopes multiple companies saw it and are already building the commercial defenses, alongside incumbents like Palo Alto Networks and CrowdStrike
He cited a YC company already selling exactly that. Tan described an unnamed portfolio company that can flag a prompt injection within about a hundred milliseconds and cut the request off — a kill switch that works regardless of how capable the underlying model becomes
His closing line dismissed the discourse track entirely. "I don't really care about science fiction, man. I saw Terminator two also... We need to actually talk about what's really happening with the servers, what's actually happening with agent swarms, how do we actually prevent that."
6. Data Startups Skip Headcount
Coogan asked what he's telling founders not to do, floating data-brokering as a fast but low-quality way to generate pre-Demo Day revenue.
Tan drew a line between two generations of data companies. He called Mercor and firms like it "body shops at some level," while a newer wave — he named Afterquery and DataCurve — "basically beat them head to head" without the headcount
The new wave's numbers, by his account. "More than 10 companies making more than $10,000,000 a year in data, and then the top two are making... hundreds of millions, on the way to billions," with 20 to 50 employees rather than thousands
He extended the same logic to gaming. He'd just funded a company called Summer Engine, which he compared to Roblox but built for a world where a small team can generate what once took hundreds of people years to build. Coogan and Jordi Hays pushed back a little — Hays said he'd expect Roblox itself to ship strong AI creation tools — and Tan said the incumbents were already responding, pointing to Astra
His view of company-building generally now centers on network effects layered on top of AI tooling. Getting friends onto a platform, knowing when they're playing and "designing all of that is actually the ultimate value," once the base capability is commoditized
Bonus Insights
Tan opened by revealing he already owns a Samsung tri-fold phone and said he's "holding out for the iPhone Quattro" — a four-panel accordion fold, which Hays said he'd also buy
Asked whether he'd found any new use for a folding screen, Tan joked the only one he'd found was "just watching TBPN"
Tan's bottom line is that AI has collapsed the cost of the hard problem — hiring the specialist, building the manufacturing relationship, staffing the engineering org — and that the winners will be small teams who use that collapse rather than the incumbents who built for the old cost structure.
Products, Companies & Tools Mentioned
Y Combinator (The accelerator Tan runs; this Demo Day batch is a quarter hard tech, up 40x)
OpenAI (Gives any YC company $2 million in tokens as part of the standard deal)
Star Cloud (A YC space company Tan cited as a buyer of other space-tech startups' output)
Palantir and Anduril (The incumbents Tan said made a winner-take-all defense-tech outcome look likely, before more teams found room to compete)
Palo Alto Networks and CrowdStrike (The incumbents Tan expects to build commercial defenses against agent swarms, alongside newer YC entrants)
Hugging Face (The subject Tan said deserves more attention than the whistleblower story dominating the timeline)
Mercor (The older-generation data company Tan called a "body shop," contrasted with newer, leaner competitors)
Roblox (Cited in the debate over whether AI-native gaming startups or incumbents with their own AI tools win the category)
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: