Name: Don't Burst my Bubble
Uploaded: 2025-12-09
Duration: 14 min 48 s
Description: The rush to get in on the ground floor of a revolution has led AI and SaaS startups to list monumental revenue and valuation numbers. How real are they, really? What here is actually part of the AI bubble? What are the red flags VCs are looking for? What metrics can we trust when we're looking at AI startups? And when the AI bubble bursts, what will be washed away, and what will survive?

The rush to get in on the ground floor of a revolution has led AI and SaaS startups to list monumental revenue and valuation numbers. How real are they, really? What here is actually part of the AI bubble? What are the red flags VCs are looking for? What metrics can we trust when we're looking at AI startups? And when the AI bubble bursts, what will be washed away, and what will survive?

Don't Burst my Bubble

Conferences

Hello, hello. Can you all hear me now? Hi. Thank you. Hi Cathy. Hi Steve and hello to all of you. Um, I have a question for everybody. Show of hands, who believes there's an AI bubble? Come on, don't be shy. Wow, fewer than I was gonna say fewer than I expected. You're surprised, Cathy. Why? Well, I think the headlines are everywhere. We're in an AI bubble. I actually tend to agree with some of the people who kept their hands down. I think. It's hard to say that AI is in a bubble across the board. Certainly for some companies and some um pockets of AI valuations have far outstripped any sort of fundamental, but we have to remember. We don't know what the demand for these AI tools look like. There's two big differences. The growth curves that we're seeing now far outstrip the growth curves of companies we've ever seen before, and I'd love to hear Steve's perspective on this. But number 2 is the t, the total addressable market. I don't think we have a good sense around how big some of these companies can ultimately become. And in hindsight it's so easy to say, oh, you know, XYZ paid so much for that company, but then when you're right, when when they're right, people tend to forget about that, that it was spicy at the time. Steve, were you surprised? Uh, no, I mean, I, I actually agree. I think it is a bubble, but it bubbles are good for innovation. Um, I think bubbles popping are bad, right? And so bubbles bursting and all of that. So, um, you need a bubble, you need a bubble and, uh, technology and startups and venture capital to not only attract the world's best talent to work on a certain set of problems, but you also need the capital to fund them. And you don't have that that that focused concentration of talent and capital without what we would call a bubble. So bubble is a finance term that finance people use, but what we would say it's a new technology wave, and every 567 years there's a new technology wave that captures everyone's attention, and you'll see it. All the big tech companies lose people to start that they're starting companies and as an early stage investor we love that. We love to see the top talent. That's a good signal, right? Uh, that's not a counter signal. That's a good signal. That's a signal that says this is a real technology wave when the best and brightest PhDs, engineers, founders leave cushy jobs at Google and Meta and Uber, and they want to go start companies and and so you know uh uh the AI wave or the AI bubble depending upon where you come from in tech or finance is a good thing, you know. I think bubbles bursting are obviously bad because that means that you know the the the the blast radius on that includes everyone exactly the thing that I'm wondering about, but we actually, so by the way, first if we keep markets our perceptions, right, markets are a set of perceptions that are dynamic. So as long as we keep talking about a bubble and fortune. Other magazines and publications keep talking about the bubble. It's good. Markets will release pressure and take air out of that bubble. They'll take money off the table, and this ebb and flow is actually a healthy thing. That's what we want to see, right? Just like Bill Gurley doesn't like IPOs that pop, we don't like bubbles that pop, right? And so, so the idea here is that. Well I think what you're trying to get at is whether or not there is foundationally core economic value being created for all of this capital and all of this talent movement, and what I would say today as an early stage investor is that absolutely there is a ton of commercialization now happening with AI. So where, so where do you get that? Um, you see that in enterprise, you're starting to see it in consumer, you're starting to see it, um, already, I would say actually most, most pronounced way would be in, uh, developer tools. So look at codegen, look at how people create apps, everything from. Uh, uh, text-based code generators, uh, looking at generative media, looking people creating advertising and videos through generative video models and generative image models, uh, look at, um, we have a company Play AI that got acquired by Meta superintelligence. Their their speech models are already inside of Instagram being used by Meta. And so this is, you know, they acquired them 6 months ago. So that's a real case with an incumbent. Cathy, where are you seeing value definitely be created beyond a shadow of a doubt. The key things that I look for across any type of company, not just AI company, is number one. Is this a feature or is this actually a workflow? Let's get real. AI is no longer a differentiator. Actually I think we're entering into the era where it's no longer AI for X, AI for sales, AI for customer support. In the future it's just going to be a customer support workflow tool and every company will be powered by AI. The key difference is companies that are able to actually embed. And fully overtake an existing workflow while having significant and building significant switching costs. The second thing I look for is really is distribution built in, right? What does that mean? So does the tool Is a tool integrated in usability? Is it integrated in all the other solutions that it needs to be in order for the workflow to be fully functional? People don't want to learn yet another tool. They want to work where they're already working and have almost super powers to let them work faster. And the third thing that I look for is. Do these companies get stronger over time, right? People call this the compounding durability. So with every new user, does the solution get better? Does it get cheaper? Does it get faster? And when you have all three, that's something that I'm really looking for. And where I'm seeing that today, by the way, I like to answer your question directly, is at the application layer and particularly around vertical AI. I was going to say we get the application layer gets a lot of love. Who makes it in the end? Application layer companies, model companies, both? How does this all shake out? Well, you know, with any sort of new a true technology way that's very differentiated, the whole stack changes, right, not just 11 layer horizontal layer in it. And so right now you're seeing it all the way from chips all the way up through application layer, um, but we're still in the early phase where we're still building out we're designing and building out that entire new stack. So what that means is, is that we have chips now and we have more chip companies, right? So it's starting to flush out. Then we have neo clouds competing with hyper scales, which is really interesting and it creates, you know, you'd think that there would be massive amounts of price pressure right if there's that many new entrants and AI compute. But guess what, you know, the neo clouds, they have 60-70% gross margins. The hyper scalers have 50 to 70% gross margins depending upon which one. So look, if the margins are still high, that means that in aggregate they have pricing power on all of the application layer companies. And so I think what you're going to continue to see is that we're we're innovating and investing largely from the bottom up because it's a new wave and AI compute is so critical to this wave. I mean, it's a super wave, right? It's, I don't want to call it a tsunami because it has negative connotation, but it's a super wave because it really affects every industry, every global market. Um, the application layer as an early stage investor, we're looking probably about half at specialized frontier models and AI infra, specialized specifically, not generalist. We can talk about that. And the application layer we're also investing, but we know that anything in the application layer, it's art and science, and it's a lot of hard work and luck. So the risk is much higher, I think, on the application layer. If you get it right, it's incredible. Everyone will know your brand, everyone will be using your products, right, every industry, but there's a lot of risk there because the substrate. The underlying layers are still being worked out, so we're like in robotics, we're at sort of the GPT 3.5 phase of robotics where everyone's using these basic VLMs and VLAs and motion planning. And that's all going to get 10, 100 times better. So a whole bunch of robotics startups that we're building on top of these primitive models are going to have a lot of like, you know, heartbreak when the models improve and they built for something sort of in the past. And so, uh, and this happened to Application layer and consumer device startups two years ago in the LLM based era, right? And so I think that like if you look at what's happening right now, even at an early stage fund. We are looking at how do we invest in companies accelerating and modernizing the AI infrastack so that GPU is a service, GPU marketplaces, chips, frontier models that are specialized that perform a very deterministic high quality outcome, no hallucinations. You don't have to worry about your regulated. industry product getting in trouble with bad information output which makes a big healthcare fintech. Think about those industries where it makes a big difference, Cathy. I would love to chime in. You know, I, I'm spending most of my time on the application layer right now. I would agree with Steve that it's tricky, right? Things are changing so quickly, but I do think there's a pocket. Within the application layer where it's very, very interesting right now. I, I tend to think about it as a 2x2, right? Um, on one axis you have either PLG or kind of core enterprise. And the other axis is is it a feature or a workflow when you go to the quadrant that is enterprise it's a workflow that's where we see real stickiness happening, right, because data has a center of gravity. Think about selling into a large company, Fortune 500, being able to collect all their data, being able to understand their business processes and convert that into your product. That is something that is very, very difficult to rip out once you get there. And when you think about prior technological waves, people always say, oh yeah, the first wave of companies never make it, right? It's like the MySpace fallacy. But if you look at the enterprise, OK, what was the first cloud HR company? It was Workday. What was the first cloud CRM? It was Salesforce first mover advantage is actually real, yeah, and they're still around. The, um, we're going, we have to go to the audience for questions. We have mic runners. If you could give me your name and company. Any questions for Kathy and Steve? I have more. Can I, can I please go ahead. So, um, I would actually say, uh, I agree with Cathy on this that it's a mixed bag, right? There's first movers, second movers, right, um, that win. uh, I don't know if there's a rule or, um, on, on one or the other. I think it's, it's mixed, but I think when I look at enterprise, I look at that not as application layer only, but my framework on that is that. Whether it's enterprise for, you know, productivity software of some sort of business productivity. Some function or generative media, it's actually a vertical alignment, right? So the models that you use, you may post train your own models, you may use third party models, you may pre-train your own models, who knows, but there's a mix of that happening and then you have to build your own developer tooling and your infra in a particular way and then your application that faces the customer and the customer uses. Um, is also customized. So I look at things like enterprise, generative media, robotics as vertical columns within this, this technical stack. Now you have quite a spicy prediction since we're getting towards the last 90 seconds here. You have quite a spicy prediction about the future of robotics, Steve, don't you? Well, I think I said it earlier, which is, um, robotics is robotics is difficult. So what's the um Who, who here has ridden a Waymo or a cruise or anything like that? OK, so, so good. Majority of the crowd, good crowd. So that is the number one modern state of the art robot that everyone is using today, right? It's a Tesla FSD. It's a Waymo. It's a Zoos or a cruise cruise, rest in peace. Zukes is out there. Zeu is out there. Neuro is out there, uh, um, uh, and they'll be powering Uber's, uh, Uber's own branded self-drive. Fleet later next year. So we're all going to be using robots on a daily basis. Our kids are going to be riding in robots on a daily basis. That area is super exciting, but think about all of the startups building humanoids. Think about all the armed robotics for industrial. You're either faced with a huge enterprise sales cycle that's within industrial. I can't think of anything more painful for a startup than that. Maybe second to that is to try to prove that a humanoid won't fall down, it won't screw up, won't have bugs, and it's going to move around in your office or household or on the street, even. So there are long development cycles to get society ready, governments ready, your home ready, your family ready, your friends ready, so. Think about that, and that is a deep tech problem that is going to, you're going to see a lot of consumer robotics companies fall by the wayside, get sold and shut down because of one, what we've talked about before, the underlying models need to improve to help them, and then the deployment and adoption cycle is going to be really long. We're at time, Cathy, last word, one sentence prediction for 2026. AI it's gonna be even more difficult to go into the enterprise. We think that model is getting better, better accuracy. People want to use more AI, but I think for the reasons Steve mentioned, it's actually gonna be harder because people are gonna be more focused on trust and visibility, and we haven't really solved that problem yet. We'll leave it there. Thank you so much, Kathy Steve.