9:49
Honestly, some of the strongest and coolest founders we've seen lately are people in their 40s, 40s, and even 50s. I mean, there's a bit of a resurgence of experienced founders right now. Many people say they want to create a "YC for solo founders ," but it turns out that YC is just that, a YC for solo founders. What a strange historical moment we live in: you wake up in the morning, plug in a new model, and something that just a month ago made you ask: " Why doesn't it work?" It just starts working. Welcome back to another episode of The Light Cone. At YC, we work with thousands of founders every year, which means we see things before they become obvious. So today we wanted to share this with you. What is the current state of affairs and what awaits
us next? What should you, creators, know? Let's begin. Diana, you have something to share with us. Yes, we did a little analysis of all the companies that have adopted in the last 18-12 months and we have some pretty shocking statistics. So, one of the main indicators is the number of hard tech companies in the set. It increased from 8% to 20%. There are many deep reasons for this. We will look at this in more detail. Another point is that the speed of growth of companies and how YC affects them has accelerated significantly. So, the median YC company has zero revenue at the time of adoption. This is the stage before generating income and creating a product. Previously, companies reached a median revenue of approximately
$8,000 by the end of the recruitment period. And now companies are averaging $20,000 in monthly revenue, instead of $8,000. So, those are the two main points we can delve into. So, let's deal with the hard tech first. What are these companies and what is the driving force behind this process? These are things that interact with atoms, not just bits. Yes. And I think you have a breakdown of hard tech companies by category, right, Diana? So, in particular, robotics has become a big trend. Its share increased from 1% to approximately 6-7% of the recruitment.
Industrial production, the return of manufacturing to the US-this is a huge trend. Its share increased from approximately 4% to 10% of the recruitment. Another important area is defense. We all work with many defense startups. The share increased from approximately 1.5% to about 5% in the recruitment. Another big change is the need for computing power as the world transitions to AI. So there are a lot of companies emerging that are building a semiconductor stack or photonics. This has increased from about 1% of the recruitment a year ago to almost 4%. And even lower down the stack is computing power. So, there is a lot of power infrastructure. This also increased from 1% to almost 3% in the recruitment. So all these
indicators in the fields of physical technologies have increased threefold or even fivefold. Yes, I think this is the age of machines. And, you know, as our motto on the T-shirts says: "do what people need," and people really need it right now. Another interesting factor about companies that are delving into physical technology is that we are funding more tech founders with more experience than ever before, right, Jared? We have some interesting statistics regarding the current summer recruitment. In the current summer intake, one in six founders has a PhD. This is significantly more than it has been historically, and that's because if you're doing, for example, silicon photonics, you're likely to need serious research experience. So we fund
a lot more of these founders, and they, in my opinion, show particularly good results. I think AGI is incredibly empowering, because in the past, complex technologies were difficult because of supply chains and the often complex software component required. Palmer Luckey talked a lot about this in relation to Anduril. Code generation means that now all the processes in Anduril can happen much faster, right? Just three or four years ago, you would have called software development and the best engineers one of the main limitations to creating top-notch, full-fledged hardware. But this is becoming less and less relevant. That means you still need one or two specialists or a small team, but you don't need to hire a thousand great engineers,
competing with Google, Meta or anyone else, and that changes the economics significantly. Actually, this is the main argument in favor of investing in complex technologies. It's not just that people are avoiding software funding, it's that the super-intelligent models we have now are accelerating scientific research, allowing startups to make big breakthroughs faster, and therefore these deep- tech companies will perform better. I think another factor is that there are three macro trends driving all this growth in the tangible sector, and the successful IPOs of giants like SpaceX have created a generation of founders who are eager to build in space. So, there are many such
companies that build solutions at all levels of infrastructure. Specifically, in the current summer of '26, we're working with a company called Exosat, which is trying to build what is essentially a sovereign Starlink-type solution. There is another company that I am working with in the winter of '26, Beyond Reach Labs, which develops solar panels for satellites in space. If you imagine that companies like Starcloud want to put all these data centers in space, they will need energy. So this is an obvious solution. Another macro trend is that we now have a generation of founders who grew up in a war that was constantly in the spotlight and actively discussed on social media, and they want to do something about it. For example, the two companies that I'm
most excited about and that I've funded over the past few rounds are Icarus last fall and Nine Mothers last spring, both of which are in the defense space. Icarus is building a solar-powered spy plane similar to the U-2 that provides surveillance and can maintain communications, which is very important. The future of drone warfare lies in the ability to effectively communicate with your drones on the ground and see the situation in real time. They were able to secure seven-figure contracts with the new War Department, and just as, because of the drone war, special operations forces began purchasing anti-drone defenses from Nine Mothers. It's essentially a shotgun turret with computer vision, but it's probably the only way to protect special forces
deep behind enemy lines. I mean, these are people who have trained for years in elite special forces, and America doesn't have thousands of such fighters; we have very few of them, so protection against conventional drone attacks is a matter of survival for the War Department that this new administration is approaching defense in a completely different way. You know, classically there was a lot of outright market capture by large defense contractors who simply operate on a "cost plus " basis. They consider themselves consultants, and you know, it's great to see new startups that can really use all the AI, all the technology, all the new manufacturing methods to create things that, frankly, the defense giants
can't create. You know, this is a really powerful megatrend right now. The fact is that defense is not just ready-made solutions that are sold to the government. There is also a large category of dual- purpose startups that sell their products to both the private sector and the government, and this applies to the entire supply chain. For example, bringing production back to America, creating custom solutions, it seems you had a company called Knox Metal. Yes, they are bringing metal manufacturing back to America. America has largely lost its steel industry. It has declined in recent decades, and without metal it is impossible to build anything. So Knox Metals is essentially
rebuilding the metal supply chain in America. And they're doing it in the heart of America, in Detroit, where there are all these empty factories that have just been idle. They are an example of a trend where people are not just creating hardware companies, but those companies themselves are growing faster than ever. I think I saw PG tweet about Knox Metals growing at a rate typical of software companies you understand this? Why are they growing so fast One reason is that many of their customers are new that have emerged and need metal to manufacture their stuff. And existing suppliers are mostly old, sluggish businesses run by elderly people, and they
simply can't keep up with the pace needed by new defense startups. It reminds me a bit of the Web 2.0 boom at the beginning of the YC era: we had new startups coming in, and they preferred to buy from new startups that could move at their speed and collaborate well with them. Like Stripe, for example. You could use a legacy credit card provider, but working with Stripe is much better. So it seems to me that they are becoming something similar for the entire defense technology ecosystem. So the third trend is that computing is a very energy-intensive physical process that is needed to quickly run all these data centers, as the demand for AI that we talked about is growing rapidly. And there is a very interesting statistic that
the cost of running an Nvidia GPU, say an A100, per hour is actually increasing, which is unusual. Previously, when you had an A100, they would now be considered a bit outdated. Outdated, yes. The price is rising because the demand is too high and the supply of computing power is limited. So many startups are emerging that are working on launching data centers: from building sites and planning software to deploying data centers and interesting power, energy, and battery solutions. So, there is a whole category of such startups that even go as far as creating basic computing silicon chips. There are a number of startups developing new silicon as an alternative to Nvidia. There's a
company that Tyler worked with called Lamb Labs that creates new processors for computing. There's another company I work with in this group called Bot that's trying to create a new custom hardware architecture using a ternary representation for models. Because it turns out, and this is a funny trend right now, if you look at all Nvidia architectures from A100 to H100, and now B200, with each generation they actually reduce the accuracy of floating-point calculations. They go from FP32 to 16, 8, etc. And it turns out that the LLM architecture doesn't need full accuracy. Even FP2 is usable to some extent That's right. So that's what Bot is trying to do. And I think you have an interesting project dealing with interconnects using photonics.
Yes, in the current group there is a company called Dipole Labs, which replaces switches in data centers-they are essentially routing systems between different GPUs. If GPU A wants to communicate with GPU B, they communicate through a device called a switch. And these switches are now electronic. There is a problem: switches cannot keep up with the speed of GPUs. GPU speeds continue to increase, and switches are actually becoming a " bottleneck" for many data centers across a variety of workloads. So, Dipal Labs is creating the first all-optical switch, where everything is transmitted by photons from GPU A to GPU B. And it will be significantly faster than the electronic switches we use now, the final factor driving this shift to physical objects is the
development of robotics. There are many companies building an ecosystem around this: from highly specialized robotics to robot deployment infrastructure and data sales to new labs, because everyone in the industry feels that we are approaching a "ChatGPT" moment in this area. We're not quite there yet, and I think we're on our way to discovering new scaling laws for this. So there's a lot of that, and we hosted Quan a few episodes ago, and we believe it's going to happen. I mean Pi robotics, right? It's half artificial intelligence, hardware. This morning I heard that even Astro is a big step forward for robotics. I forget the name of the test, but there was a comparison where Fable performed maybe 10%, and Astro shows that it can handle 60-70% of
the tasks. So the data just keeps piling up, a few more weeks go by, another breakthrough happens, and we'll be one step closer. I am very pleased to see the revival of hard tech technologies, because at YC we have been funding such companies since 2014. This is really when we started, but it used to be quite difficult to find funding for companies like this. I remember, before this recent renaissance, we were funding amazing things that we were excited about: rockets, airplanes, chips,
data centers, and so on. And the venture capitalists just said, "Oh, we only do B2B SaaS." And it's hard to grow such a company on your own, so you really need investors who can finance the full capital construction. So it's great that Silicon Valley, historically built to fund technology, is returning to that, after a decade or two of moving away from the more profitable SaaS. It's cool to see everything going back to its roots. Yes, regarding the fact that investors want to invest in "hard tech" again-it seems that this is indeed the case. I've never seen it happen so quickly. That is, it seemed to happen instantly: when SaaS stocks were falling at the beginning of the year
, Cloud Code was growing rapidly, and it became kind of, well, I think it literally happened even during Demo Day. I think it happened somewhere in the middle of the winter recruitment at the beginning of the year, and by Demo Day, investors had started to become much more interested in hard tech companies, and it was only getting stronger. It is worth noting, however, that since then many SaaS stocks have actually recovered and are performing better than ever. Salesforce is the best example of this. It seems that Snowflake recently, about two days ago, had some incredible financial reports. So, it's possible that all of this will work.
Yes, it is possible. That would be the ideal scenario. Although I still believe we are seeing real things. The software that will be created in the future and what will be valuable will obviously be different. It just has to be like that . So in a way, it seems like with Salesforce we're seeing the classic "system of record" argument come to fruition. The protection levels are still in effect. Yes. If you have a tool that agents can use and it's truly valuable, then you understand that agents use software much more often than people, and that's what seems to be driving Salesforce's growth. This brings us back to another trend we talked about:
if you treat agents as customers and build things that agents need, and if your software is something that agents want to use, then it's the right kind of software. Yes, Salesforce is very interesting, because it seems they have started releasing their own Slack interface, Slack AI harness. So I think we're at the very beginning of the next AI interface wars. It's like Codex wants to be it, Cloud Code wants to be it, Open Cloud could become it, Hermes, Open Code. It looks like there will be a lot of them, and it won't be quite like the " browser wars" where it usually comes down to one winner. Although, who knows, and, you know, Benioff has a pretty big
advantage, because many of the most AI-focused people and companies in the world still use Slack. And if the interface is integrated there and is your system of record of how people collaborate, then you get a real mega-level of data protection. And you know, SaaS can still be as valuable as it was before if those barriers persist. Yes, you had a very interesting tweet, probably a week ago, about how software or accounting systems companies should actually become " harnesses." Yes, they have-I mean, it was about Slack, I would say. So, essentially, if you are an accounting system, you either get absorbed, or you release the MCP, and maybe you lose your protective moat around the data, the data goes somewhere else
it becomes very easy to change the system, or you have to become a " harness". You need to become the thing that people use to not only read and write, but actually do their work inside your accounting system. And they get the maximum benefit from it. It's a bit like modeling companies. For example, was it RKGIV or some other benchmark where. RKGIV3? Yes, where the model before ChatGPT didn't show such results, but they say, "Well, that's just because it was connected to the wrong harness ." So the model plus the harness gives you the result. Oh yes. I mean, with the individual harness, they claim that Astra achieved over 90% in RKGIV3.
Yes. I think a few months ago it was in the low double digits, which is an impressive jump. And I think you're right about the software. It's not that software and SaaS are dead, as some claim on the internet. They simply transformed. We actually see this in our set. The percentage of companies we accepted that were doing full-time work or assignments increased from 10% to over 25% of the total recruitment. This is related to the direct performance of the work. The agent does the work, it's not just a point solution like the old SaaS was 5-8 years ago when you needed someone to manage that software. Now it works independently. And in fact, in our set, this is where we see significant revenue growth. I have
already cited statistics that the median startup, when it gets to YC, has zero revenue. Previously, by the end of the recruitment, it was about 8 thousand MRR. It's now around 20k MRR, and to a large extent that's. a huge jump. That's a pretty significant jump for the median, right? The average value is even higher. And this involves doing the work from start to finish, for example, insurance broker services, conducting clinical appointments, full workflow, I don't know, medical billing, etc. They are the ones that are growing rapidly now, and I think there is another factor here. I think we talked about this a few episodes ago.
We are now almost a year into agent programming after Opus 4.5, and we are seeing these workflows fully blossom. And the results are: people just need the job done, and they are willing to buy software that simply does the job. I think when people hear about such rapid revenue growth, the simplest explanation is that it's just AI hype, and companies are spending money on AI products just because it's trendy. To be fair, maybe that's partly true, but I think the optimistic scenario is what we talked about in the episode about two years ago, when agents were just starting to emerge, we said that products would just become more valuable. If they automate all the work, they will become more valuable than some
accounting systems that only track work but don't perform it, so companies will pay more for such products. I definitely see this in the companies I work with where they come to the customer and the value proposition is so high that large enterprises are willing to write big checks very, very early on. Yes, a company where I see such effects is Juice Box, an AI-based recruiting tool. They've been showing incredible growth over the last few years, but they started out, I would say, essentially as an LLM-based people search. They worked like this: you enter a description of the person you want to hire, and the system perfectly matches profiles of people you might want to contact. But then you still had to contact them yourself, and
recently they launched an agent product that is really taking off, because the agent doesn't just look for people, he contacts them himself, schedules interviews, and performs a bunch of other tasks. Yes, and they see that this will lead, if we take the calculation per account, I think to a doubling or tripling of income from one client, because clients want more and more such agents. I don't think it's fair to say that this automates the recruiter's work. It just changes her. Recruiters didn't really want to do this monotonous mailing to 500 people anyway. What makes a recruiter's job more qualified and interesting is, for example,
matching the corporate culture. It will be very difficult for an AI to conduct a phone interview to assess a person's cultural fit and understand this human aspect. So I think they see that recruiters themselves are actually very happy to use agents because it frees them up to do work that they find unique and interesting. Another shocking statistic is that companies that accelerate their recruitment start to take off rapidly. One of the things we've seen this year that we've never seen before is companies going from zero to seven-figure revenue during the hiring period. And this is over a period of 3 months. And this is shocking.
Previously, it would have taken the company about 18 months or more to reach this level. And they do it in such a short time. Part of this is because they solve real problems, and because of this agent programming, they create much more mature products. These founders, who are completely immersed in AI, are running, I don't know, 20 sessions of agent coders to get to that level of product maturity. And there is another category of companies that has also been growing at a very fast pace recently-those that sell data or reinforced learning environments for labs. This can be an interesting topic to talk about, as
many of these companies are quite secretive. They don't have the incentive to talk about their successes compared to most companies that like to brag, so I think people may not realize how big this category has become. When YC invested in Scale in 2016, it was a tiny niche business. It wasn't even a separate category. At first, it was actually only Scale that did this, then Mercury started, and then a few more companies, but in the last couple of years it has become a big category. We recently analyzed the data, and in the past 2 years, YC has funded over a dozen companies, each earning over $10 million per year, selling data or reinforcement learning (RL) environments to labs. And in many cases, hundreds of millions of dollars. In many cases,
hundreds, hundreds of millions of dollars. And these are companies that are only a few years old. This is a pretty quick path to income. Honestly. Yes, this is just crazy. Do you want to talk about any of them, Gary? Well, I mean the big ones, I think, after Corey and Data Curve, both very, very great. There are probably too many to name, and honestly, they might not even want to be mentioned, because when something works, you almost don't want people to know about it. I think it's quite natural to understand. Data is one of the pillars of the law of scaling. And, you know, there's a lot of talk about computational power, but without
data, how are you going to make these models significantly better? RL environments are an interesting thing. There is a lot there. There is a lot of purely individual customization going on for specific use cases. For example, you can have RL environments for finance, and someone can delve into that very, very far. It's a bit of expertise, a bunch of computer science, and some systems work. But RL seems to be one of the big drivers of how…people are trying to maximize performance, maybe a little more than they should, but it costs money, and it seems like it's going to take a long time, given the approach of the big modeling companies.
Large labs are reportedly spending about a billion dollars on this. This is not a well -known fact, but you can build a real business around it. RL environments are trending right now, creating things with long-term goals, and I think this is starting to appear in robotics as well The labs also want to solve the problem of how AI works in the physical world. So they need a lot of real-world environments. So things with egocentric data, telecontrol, and robotics are starting to stand out in the big data category, where labs are making eight- to nine-figure deals with these companies. We had a number of companies in the pipeline that are working on this and have been able to generate revenue in this area. These are companies like in the current
summer 26th recruitment, such as Praxis Robotics. There is one I work with that has a network of locations around the world where industrial production is conducted. They collect data from this. There is another company that Brad has worked with called Deep Reach, which also has data from local entrepreneurs around the world. And Human Archive in the winter set on the 26th. Yes, a hundred such companies have appeared recently. Exactly. I think if I were to make predictions, one of the things, going back to the fact that all accounting systems need to be integrated with AI, is that they might have to start training their own models. And this is where things like River AI
or Tinker get really interesting. "Out of the box" you can just sit in Cloud Code or even Open Cloud to train your own models, and that's very interesting. It does the data cleaning and all that stuff itself. But it hasn't been a big factor today, but I can see how it could become a much bigger factor. I mean, when you have your own data and you can train, open- weight models are almost at the cutting edge. If you can specifically train these models to do something even better than the leading solutions, that would be very, very powerful. I think in robotics it will actually matter even more. This is just a hypothesis, it has not yet been proven, but the fundamental models in robotics, in my opinion, have completely different characteristics
compared to LLM. The essence of LLM is that you model reality as a language. And for robotics, you're modeling reality in a physical 3D space, which has many more degrees of freedom. And, perhaps, to make robots work in a certain niche, say, robots that perform operations in data centers. I have a company called Boost Robotic that builds robots for data centers, for example for working with cables. It is possible for these robots to perform such work. It is better to get a model that is pre-trained and trained on special data that works in this particular environment, because the difficulty for robotics is that they have to work in real time,
respond very quickly to stimuli, and have a plan of action that is different from LLM. In LLM, there is a feature where you can start a process and walk away, but in robotics, that won't work, because, say, you're connecting a cable to a data center , and someone accidentally pushes the robot, and something can connect to the wrong connector. Yes, as far as I understand, all YC companies that use physical intelligence models for robotics are retraining Pi models. I don't think any of them are using Pi models without modification; While this is a great start, you need to further train it for a specific case, such as working with cables in a data center.
Ugh. You work with Ultra, right? Yes. They start with the Pi model, but they have thousands of hours of video of them putting things into boxes, which makes it very effective at that. I've heard the argument that you can take cloud code, use transcripts to identify the best developers, and use that to train the next programming model to be even better. If you own TikTok, you have all the data about what people are watching, what they are clicking on, and what they are hooked on, and you can use that to create much more engaging videos. So, you know, it's already happening. I think this trend will continue to develop at a rather impressive pace. We've noticed, and
you've probably all noticed, that the strongest and coolest founders we've seen lately are people in their 40s, 50s, and even older. That is, we are witnessing a certain renaissance of experienced founders. Many people say they want to create a YC for solo founders, but it turns out that YC is just YC for solo founders. Diana, you have some statistics that surprised you. One of the stunning numbers: a year ago we were only accepting about 5% of companies with solo founders, and now there are over 18-19% of them, this is the biggest jump we have ever seen. Almost a fifth of the recruitment, and it looks like this trend will continue. You know, before you had to have, well, so many different skills. You had to be a great entrepreneur, you
had to be able to explain, and, you know, we would say they had to be good speakers, right? That is, you need someone who can be, well, a bright personality. Someone who can actually come and convince a person of something. And if you combine that with someone who is a world- class technologist, that's the perfect combination. So classically for this you needed co-founders. You know, not that they were mandatory, but it significantly, very significantly increased your chances. I feel like in many ways things are changing just like that. It turns out that knowing what prompt to write and what to create is much more complex
and valuable than simply knowing how to write code for a CTO. I think what's happening is that we've actually always had, um, extremely successful single founders. I think people don't realize this applies to YC. There are different definitions of this, but by all indications, Apoorva from Instacart, Brian Armstrong from Queen Bees. At least when the recruitment started, they were single founders. Yes, Parker Conrad came to YC as a solo founder, and then I interviewed Lakshminarayan, who eventually became his CTO The bar for having an idea, being able to sell it, and implementing it on your own was simply very, very high. Ahem.
And this is actually quite real. Yes, I think that's exactly what's happening. In that case, these three are simply incredibly exceptional people. And they are simply the same people who are capable of doing this. And now you can really start working. So I think you don't have to be that exceptional, at least in one of these aspects, the development part, to get started. But overall, having co-founders is still valuable. This is still an indicator that if your co-founders are elite, it means you are probably elite too, and that significantly increases your chances of success. And in each of these cases, they still brought in co-founders. I think in each of these cases, you just start, it goes, and they add co-founders at some point.
Perhaps the distribution of shares is different or the dynamics are slightly different from the classic "you two are in a room and you have exactly 50/50." I do n't even want to get into it, but yeah, I don't have statistics, but I definitely see a trend where people add co-founders later, after the business has gotten off the ground. I think this will be a trend. I think we'll see more solo founders at startups, but at least among those who succeed, I expect they 'll still add co-founders as the Gary, do you also want to talk about the trend of more experienced people starting companies? It seems that people who have already "went through fire and water" do much better. I consider Peter Steinberger to be a kind of canonical example. He
seems to be in his early 40s, was a development manager, worked in startups before, and then, you know, got really into AI in a unique way at an early stage. He's tried a lot of things and now he knows what to build, and I think that's really encouraging. Essentially, if you've been "in the thick of it," you know where the dragons hide. You have a certain flair, and such people are especially effective right now. Yes, there are so many classic barriers that used to be unshakable. For example, you definitely need a co-founder or a circle of cool investors, but now this is becoming less and less relevant. In fact, the main thing is to know what exactly to build. This is
the most important thing now: you need to understand the product. If you've lived a little, worked somewhere, and have your own opinion, then you have no excuses now. What's your excuse? You've been such a vocal critic on the internet for so long. Why aren't you building anything yet? Just open the code and get to work, you know? Prove your point with actions, not words. I also wonder if managing coder agents is in some ways similar to managing people. Oh yes. So people like Peter, you, or Forrest Jurney, or Toby at Shopify, who have dedicated their entire careers to managing teams of engineers, are actually very good at
it and can build huge teams of agent coders, managing them perhaps even more effectively than a very smart 19- year-old who doesn't have that many years of experience. We can be a little less rude to our agents. Try to understand what their problem is. You need to capture their emotions 99.9% less. Yes. So, yeah, it's pretty useful. I wonder what specific advice you have for someone who wants to start and wants to build a company right now, in this day and age?
I mean, just start prompting. To be honest, today's launch of GPT-6 was simply incredible. I mean the point where your agents become, you know, noticeably smarter. They, you know, solve a bunch of things that were annoying you, like these mistakes that you didn't have time to look into deeply yourself. You used to be able to say, " Listen, can you go back to the list of things that you can deal with? " Look at, you know, all our recent chats, and, you know, anything that looks like you have n't dealt with, try to deal with it now, and he does that every time. It's a strange moment in history when you wake up in the morning, plug in a new model, and things
that a month ago would have made you think, " Why isn't this working?" They just start working. And you know, to think that this could be what we're going to be doing for the next 18, 24, 36 months-who knows, I don't know when it's going to end, but I guess that's what it is- programming in the age of AGI. Well, that's all for today, but if you haven't figured it out yet, we're all really excited about what's happening right now, and you should be too. So we look forward to seeing
