Y Combinator Startup Podcast Sep 18, 2026 36m 14m saved
Companies that build physical products went from 8% of a Y Combinator batch to 20% over twelve months, on the firm's own analysis of everything it accepted in the last year.
The usual explanation is that investors went cold on software. The partners' explanation is the opposite: code generation has removed the engineering headcount that used to make hardware unaffordable for a startup, so the physical companies now work.
"That's the true bull case for hard tech — it's not just that people are shying away from funding software businesses, but it's actually that the super-smart models we have now are accelerating scientific research and making it possible for startups to have bigger research breakthroughs earlier."
Y Combinator works with thousands of founders a year and ran the numbers on its own portfolio for this episode of its Lightcone show, including revenue at acceptance and at the end of each batch, the share of solo founders and the composition of every category.
The full episode is covered here so you can skip it. 36 minutes of audio, 22 minutes of reading.
Here are the 16 numbers that matter.
Key Takeaways
Hard tech went from 8% to 20% of the batch in twelve months
The median company ends the batch at $20K in monthly revenue, against $8K before
One in six founders in the current summer batch holds a PhD
Solo founders went from 5% to more than 18% of accepted companies
An Nvidia A100 costs more per hour now than it did when it was new, because demand outruns supply
Companies that do a whole job rather than track it went from 10% to over 25% of the batch
More than a dozen YC companies each make over $10M a year selling data or reinforcement-learning environments to the labs
Some companies now go from zero to seven figures inside a three-month batch, against 18 months before
Defense went from 1.5% to 5% of the batch, and industrial manufacturing from 4% to 10%
The most powerful founders the partners are seeing are in their late 30s, 40s and 50s
1. Hard Tech: 8% To 20%
The episode opened with an analysis of every company Y Combinator accepted in the previous twelve months, and two figures came out of it.
The share building physical products more than doubled
One of the big ones is the number of hard-tech companies that are in the batch — it has gone from 8% to 20%.
A host
The second figure was about speed. The median company arrives with no revenue and no product.
And it leaves with two and a half times what it used to
and by the end of the batch, in the past, companies would get to about $8K median revenue, and now the companies in the median are getting to $20,000 monthly revenue, as opposed to $8K.
A host
The definition of hard tech being used
Things that actually touch atoms and not just bits.
A host
2. Where The Growth Sits
The 8-to-20 move breaks into five categories, each of which grew by a multiple rather than a margin.
Robotics, manufacturing and defense
Specifically, robotics has been a big one — it has gone from 1% of the batch to about 6, 7% of the batch. Industrial manufacturing, building things back in the US, has been a huge trend — it has gone from about 4% to 10% of the batch. The other one is defense — a big one, we've all been working with a lot of defense startups — it has gone from about 1.5% to about 5% of the batch.
A host
Then the two layers underneath artificial intelligence
The other big one is there's this compute need that the world is getting into with AI, so there's a lot of companies building the semiconductor stack or photonics — that's gone from about 1% of the batch a year ago to close to 4% of the batch. And the other one, even below the stack of compute, is power — there's a lot of power infrastructure as well, gone from also 1% to close to 3% of the batch.
A host
The scale of the change across all of them
So all these numbers, across the physical, atom stacks, have somewhere tripled or quintupled.
A host
3. 1 In 6 Founders Has A PhD
The composition of the founders changed alongside the composition of the companies.
The research credential is now common
We have this fun stat about the current summer batch — one in six of the founders actually has a PhD.
A host
The reason given was mechanical rather than cultural: silicon photonics requires a research background, so funding those companies means funding those people. The partners said that group has been performing disproportionately well.
4. Codegen Makes Atoms Cheap
The argument for why hardware startups work now is about engineering headcount. A host cited Palmer Luckey on how code generation has changed the pace of work at Anduril.
Top software engineers used to be the constraint on hardware
Even three or four years ago you would talk about software engineering and the top-tier software engineers as one of the limiting reagents to being able to do really top-tier, full-stack hardware, and that's less and less true.
A host
And a startup no longer has to win a hiring war
You still need one or two of them, or you need a small team, but you don't need to hire a thousand great engineers versus Google or Meta or whoever else, and that really changes the economics.
A host
Which is the case, rather than a flight from software
That's the true bull case for hard tech — it's not just that people are shying away from funding software businesses, but it's actually that the super-smart models we have now are accelerating scientific research and making it possible for startups to have bigger research breakthroughs earlier, and that therefore these deep-tech companies will actually work better.
A host
5. Space After The SpaceX IPO
Three macro trends were named as pulling founders toward physical products. The first is space, and the trigger given was a listing.
One IPO created a generation of space founders
one is seeing huge companies like SpaceX have such a successful IPO has created a generation of founders wanting to build in space
A host
The examples were a company called Exosat in the current summer batch, building what a host described as a sovereign alternative to Starlink, and Beyond Reach Labs from the winter batch, building solar panels for satellites. The demand case offered for the second was that companies like Starcloud, which want to put data centers in orbit, will need power up there.
6. Two Defense Startups
The second trend is defense, and the explanation was generational: founders who have grown up with a war discussed constantly on social media and want to do something about it.
A solar-powered spy plane with seven-figure contracts
Icarus is doing a solar-powered U-2 spy plane that gives overwatch and can also do comms, which is actually really important — the future of drone war is being able to actually communicate with your drones on the ground and see what's going on.
A host
The second company, Nine Mothers, sells anti-drone defense to special forces.
A shotgun turret with computer vision
it's basically a shotgun turret with computer vision, but it's actually almost the only way that you could protect special forces deep behind enemy lines
A host
The argument for why that matters commercially is that the people it protects are scarce: a small, heavily trained force facing what a host called a commodity drone attack. On the buyers, the partners were direct about the incumbents.
The primes behave like consultants
classically there was just a lot of, frankly, capture from the big defense primes that are doing sort of cost-plus, they think of themselves as consultants
A host
And cannot build what the startups can
to build things that frankly the defense primes can't build — that's a really powerful mega-trend right now
A host
7. Metal, Back In Detroit
Beneath the companies that sell finished systems to governments sits a layer of dual-use suppliers. The example given was Knox Metals.
America lost the industry and cannot build without it
America has largely lost its metal industry — it got hollowed out over the last few decades and can't build stuff without metal.
A host
The company is rebuilding that supply chain from Detroit, in factory buildings that had been standing empty. What the partners found notable was the growth rate.
A metals company growing like software
I think I saw a PG tweet that Knox Metals is growing at software growth rates.
A host
Asked how, a host gave a demand-side answer.
Its customers are the new defense startups
So one reason is that a lot of their customers are these new defense-tech startups that have sprung up and need metal to build all their stuff, and the existing suppliers — these sort of sleepy old businesses, mostly run by old people — just can't keep up with the pace that the new defense-tech startups want to build at.
A host
He compared it to the early web era, when new startups preferred to buy from other new startups.
The analogy is Stripe against a legacy processor
Like Stripe, for example — you could use a legacy credit card vendor, but it's just way better to work with Stripe.
A host
8. Compute Is The Bottleneck
The third trend is the physical work of getting data centers running, and the price signal a host used is unusual for a piece of computing hardware.
An older chip is getting more expensive, not cheaper
there's a very interesting stat where GPUs from Nvidia — an A100 GPU per hour is actually appreciating in cost, which is unusual, because in the past when you get an A100, by now it's sort of old
A host
The reason is straightforward
The price is going up because there's just too much demand and not enough supply of compute.
A host
That has produced startups across the whole build: site construction, planning software, the buildout itself, and power and battery combinations. Below all of that, startups are attacking the chip. One, Lamb Labs, is building new processors; another, called Bot, is building custom hardware around a ternary representation of models, on a bet about numerical precision.
Each Nvidia generation uses less precision than the last
if you look at all the Nvidia architectures from A100s to H100s and now the B300s, each of these generations is actually going down in floating-point precision, going from FP32, to 16, to 8, etc. And it turns out that LLM architecture doesn't need the full precision — FP2 is even somewhat usable.
A host
A separate company, Dipole Labs, is attacking the network rather than the processor. Graphics processors inside a data center talk to each other through switches, which are electronic.
The switch, not the chip, is now the limit
the switches are not keeping up with the GPUs, the speed of the GPUs keeps going up, and the switches are actually the bottleneck for many data centers in many different workloads
A host
So it is building an all-optical one
And so Dipole Labs is building the first fully optical switch, where it's all photons from GPU A all the way to GPU B, so it will actually be much faster than the electronic switches we use now.
A host
9. Robotics Waits Its Turn
The fourth driver is robotics. The partners described companies building across the stack, from vertical applications to deployment infrastructure to data sold to the robotics labs, in advance of the moment that makes any of it work.
The industry is waiting for its ChatGPT moment
because there's this moment that everyone in the industry is feeling — that we're going to get to the ChatGPT moment, just not quite there yet
A host
A host put the split in the problem at half software and half hardware, and cited a benchmark as evidence of how fast the software half is moving, with Physical Intelligence's model at around 10% and the Astra model at 60 to 70% of tasks.
The pace of it
So — you just wake up another couple weeks and another breakthrough happens, and we're a little bit closer.
A host
10. Why VCs Came Back
Y Combinator has funded hardware since 2014, and for most of that time could not get the companies financed downstream.
The answer they used to get
I remember, pre this recent resurgence, we would fund awesome stuff that we were super excited about — rockets and planes and chips and data centers and stuff like that — and then VCs would just be like, ah, we only do B2B SaaS.
A host
That matters because a rocket company cannot bootstrap; it needs investors willing to fund a full capital buildout.
Venture returning to what it was for
which historically, venture was set up to fund hard tech, but it drifted away from it for a decade or two, because it was so profitable to just fund SaaS companies
A host
The switch, a host said, happened quickly, around the middle of the winter batch, when software stocks were falling and coding agents were surging, and was visible in investor behavior by demo day.
With a caveat the partners raised themselves
It is worth knowing, though, on the other side, that since then a bunch of the SaaS stocks have actually recovered and are doing better than ever — Salesforce is the prime example of that.
A host
They noted Snowflake's results two days earlier in the same breath.
11. The Harness Wars
The explanation offered for Salesforce's recovery is that the system-of-record position still holds, provided software becomes something an agent wants to use.
The moats are intact for now
Like, if you have a thing that agents can use, that is actually valuable, and if anything you'll just have agents use software a lot more than humans will, and that seems to be driving Salesforce growth.
A host
Which reframes who the customer is
if you think of agents as your customers, and you make things that agents want, and your software is something that agents want to use, then that seems like the right type of software
A host
The competitive fight the partners expect is over the harness, the layer a model works inside.
And they named the entrants
I think we're right at the beginning of the next AI-harness wars — Codex wants to be it, Claude Code wants to be it, OpenClaw could be it, Hermes, OpenCode
A host
Unlike browsers, they do not expect one winner
it seems like there are going to be a bunch of them, and it's not going to be quite like the browser wars, where the browser wars tend toward one winner
A host
The advantage they credited Salesforce with is that the companies most committed to artificial intelligence still run on Slack.
A harness inside the place people already collaborate
if the harness is in there and it's your system of record for how people collaborate, then you have this mega-data moat. And SaaS can still be as valuable as it's ever been valued, if those moats hold.
A host
The alternative outcome for a system of record was stated as a binary.
Become a harness, or lose the data
It's basically, if you are a system of record, you either will be preyed upon — you'll release an MCP, and then the data — you lose your moat around the data, the data goes elsewhere, it becomes very trivial to switch — or you kind of have to be a harness
A host
The same logic was applied to models themselves, using a reasoning benchmark where the harness rather than the model explained the jump.
With the right harness, above 90%
With a custom harness, they claim Astra got to north of 90% on ARC-AGI V3.
A host
From a much lower base
I think a couple months ago this was in the low two digits, right? Which is an impressive leap.
A host
12. $8K To $20K In One Batch
The revenue acceleration has one measurable cause: the share of companies that perform a whole task instead of tracking it.
Full-stack companies went from 10% to over a quarter of the batch
the percentage of companies we accepted that do sort of full-stack, end-to-end work, or a task, has gone from just 10% to over 25% of the batch
A host
And the old model needed a human to run it
which old SaaS five, eight years ago was just a point solution and you needed someone to operate the SaaS software — right now it just runs by itself
A host
The named examples were insurance brokering, clinical intake and medical billing, done end to end.
What customers are buying
we right now are about almost a year since agentic coding started to work, since Opus 4.5, and we're seeing these workflows fully blossom, and the result is basically people want their job to just be done, and are willing to buy software that just gets the job done
A host
The partners raised the obvious objection against themselves.
The bear case is that companies are buying because it is fashionable
I think when people hear these revenue numbers growing so fast, an easy knock on it is that maybe it's just AI hype, and these companies are just shelling out money for AI products because it's the cool thing to do — and to be fair, that's probably some of that.
A host
And the bull case is a pricing argument
if they automate the whole job, they will actually just be more valuable than some system of record that tracks the job but doesn't do the job, and therefore companies will just pay more money for the product
A host
The acceleration shows up at the top of the range as well as the median.
Zero to seven figures in three months
we have companies breaking from zero to seven figures in revenue during the batch, and that's in a span of three months. That's shocking — in the past, that would have taken a company about 18 months or more to get to that.
A host
Part of the explanation given was product maturity: founders running many coding-agent sessions at once arrive at something finished sooner.
13. A Dozen $10M Data Firms
A category the partners said is deliberately quiet is selling training data and reinforcement-learning environments to the model labs.
The companies have a reason not to talk
a lot of these companies are pretty stealthy — they tend to have a disincentive to talk about how well they're doing, unlike most companies that like to talk about how well they're doing
A host
When Y Combinator funded Scale in 2016 this was not yet a category. Mercor followed, then others.
The scale of it now
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
A host
And they got there quickly
It makes hundreds of millions of dollars, and these are companies that are just a couple years old. That's pretty fast to revenue, honestly.
A host
The two named were Afterquery and Datacurve; a host said there were too many others to list, and that several would prefer not to be listed.
The reason the category exists
data is one of the legs of the scaling law, and much has been made of compute, but without the data how are you going to make these models that much better?
A host
The environments are being built for specific domains, with finance as the example, and a host allowed that some of the spending may be aimed at benchmark scores. He added that it costs money either way and looks permanent.
The labs' reported spend
Reportedly, the big labs are spending about a billion dollars on this — it's not a very known fact, but there's actually a real business to be built around this
A host
The next version of it is physical
the labs also want to solve the problem of getting AI to work on the physical world, so they need a lot of environments in the real world — things with egocentric data, teleoptic tasks, starting to emerge as a big data category, where labs are spending eight-, nine-figure deals with these companies
A host
Companies named in that line of work were Practis Robotics, which collects data from a network of industrial production sites, Deep Reach, which sources data through local entrepreneurs, and Human Archive.
14. Fine-Tuning Everything
The prediction that follows from all of it is that systems of record will train their own models, not merely plug into someone else's.
Training a model is becoming a thing a harness can do
they might also need to start training their own models, and that's where things like River AI or Tinker start becoming really interesting
A host
Open weights plus proprietary data is the combination
I mean the open-weight models are really nearly frontier — if you can sort of special-purpose-train these things to do even better than what the frontier can do, that's going to be really, really powerful
A host
The partners think this matters more in robotics, where the modeling problem is different.
Language against physical space
the whole thing with an LLM is you model reality as language, and for robotics you model reality in physical 3D space, which has way more degrees of freedom
A host
There is also a latency requirement that language models do not have: a robot has to respond immediately, and a mistake has physical consequences. The example given was a company called Boost Robotics, which builds robots to run cabling inside data centers, where connecting a cable to the wrong plug is not a recoverable error. The pattern holds across the portfolio.
Nobody uses the base robotics model as it comes
My understanding is that all the YC companies that are using Physical Intelligence's models to deploy robotics are all fine-tuning the PI models — I don't think any of them are able to use the PI models out of the box.
A host
What the fine-tuning data looks like
Yeah, you start with the PI model, but then they have thousands of hours of footage of putting things in boxes, that makes it really good at putting things in boxes.
A host
The same flywheel was described for software and for video: transcripts from a coding tool can identify the best coders and train a better coding model, and ownership of a short-video feed supplies the signal for generating more compelling video.
And he expects it to keep running
I just think that trend is going to continue in a fairly spectacular way from here.
A host
15. Solo Founders Hit 19%
The partners said the firm has become the thing other funds say they want to be.
The line on solo founders
A lot of people seem to say that they want to be the YC for solo founders, but it turns out YC is the YC for solo founders.
A host
And the number behind it
we used to only have about 5% of the companies accepted be solo founders, and now we're over 18, 19%, which is huge
A host
The explanation was about which skill is now scarce. A founder used to need to sell and to build, and pairing a strong seller with a world-class engineer was the reliable way to get both.
Knowing what to build is the scarce skill now
it's becoming such that 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
A host
The partners pushed back on the idea that this is new, naming Apoorva Mehta at Instacart, Brian Armstrong at Coinbase and Parker Conrad as people who entered the program alone.
The bar used to be extraordinary
the bar for being able to have the idea, be able to sell it, and be able to build it all by yourself, was just really, really high, and that's actually totally doable.
A host
What has changed is how exceptional you have to be
now you can actually get going, so I think you just don't have to be quite that exceptional, at least on one of those dimensions, the building part, to be able to get going
A host
They were clear that co-founders still help, and that all three of those founders added them later.
The signal a co-founder sends has not gone away
it's still a measure of, if your co-founders are super elite, that means you're probably super elite, and it just increases the chance of success by a lot
A host
So the expectation is later pairing, not none
I think we'll see more single founders in the batch, which we're already seeing, starting the batch, but at least of the things that succeed, I still expect that they're going to be adding co-founders as the company progresses.
A host
16. The 40-Year-Old Founder
The last shift the partners described is in age, and it runs against the usual assumption about who starts companies.
The founders impressing them are older
some of the most powerful and badass founders that we've been seeing lately might be in their late 30s, 40s, even 50s — there's a sort of resurgence of the experienced founder
A host
The example given was Peter Steinberger, in his early forties, a former engineering manager who had worked on startups before.
Experience tells a founder where the problems are
basically, if you've been around the block, you know where the dragons are, you sort of have taste, and then those people in particular are unusually powerful right now
A host
The gatekeeping that used to apply has weakened
there's just so many classic gatekept things that happen — it's like, oh, you know, you have to have a co-founder, you need a certain set of cool investors to be into you — and now it's just less and less true
A host
A second reason was offered, which is that directing coding agents resembles a management job.
Managing agents is close to managing engineers
people like Peter, or you, or Boris Cherny, and Toby from Shopify, who have had whole careers managing teams of engineers, actually take to this super well, and can spin up huge teams of coding agents and manage them maybe more effectively than even a really smart 19-year-old who hasn't had those years of experience.
A host
With one difference, dryly noted
We can be a little bit less abusive to our agents, try to understand where they're coming from — you have to catch their emotions like 99.9% less.
A host
Bonus Insights
The advice, when asked for something concrete
I mean, just start prompting.
A host
The illustration was the arrival of a new model that morning. A host described pointing it at a backlog of bugs that a previous model had failed on, and having them resolved.
The experience he described
What a weird moment we are in history, where you wake up in the morning, you 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
A host
A host also made the case that recruiting agents do not replace recruiters, using a company called Juicebox, whose product moved from searching for candidates to contacting them and scheduling interviews, and which expects that to double or triple revenue per account.
Why the recruiters welcome it
The thing that makes the recruiter's job, I would say, more skilled and interesting, is culture fit — that's just going to be really hard for an AI to do a phone screen that assesses how well someone's going to be a culture fit, and the human element of it.
A host
The partners' bottom line is that the constraint on building a company has moved from engineering capacity to knowing what to build, which is why the batch now holds more physical products, more solo founders, more people in their forties and more revenue by the end of three months.
Products, Companies & Tools Mentioned
Y Combinator (Source of every figure in the episode, from its own analysis of the companies it accepted over twelve months)
Nvidia (Its A100 costs more per hour now than when it was new; each chip generation uses lower numerical precision, which is the opening the new silicon startups are aiming at)
Salesforce and Slack (The partners' test case for whether a system of record survives by becoming the place agents do the work)
Snowflake (Named alongside Salesforce as evidence that software stocks recovered after the rotation into hard tech had already begun)
Scale AI and Mercor (The first two companies in what is now a category of more than a dozen YC companies each earning over $10M a year from the labs)
Anduril (Cited via Palmer Luckey as the example of code generation compressing the timelines of a hardware company)
SpaceX (Its listing is credited with creating the current generation of space founders)
Stripe (The analogy for why defense startups buy from other startups rather than incumbent suppliers)
Knox Metals (Rebuilding a US metal supply chain out of empty Detroit factories, and growing at software rates because its customers are defense startups)
Physical Intelligence (Every YC robotics company using its models fine-tunes them; none, a host said, can use them as they come)
Juicebox (An AI recruiting tool whose move from search to an agent that contacts candidates is expected to double or triple revenue per account)
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: