AI Bubble Tea with VCs: Is AI Overvalued? | DVC's Marina Davidova

Nick, you know like in this time, this is trying to be on time, just not to be late to the party. And by the way, when all these orders are gone, we still have some cookies. Please enjoy. They are real. Yes, please. And finally, the talk that I've been expecting for the whole day because like the whole name of the conference is AI bubble tea, but we are actually trying to figure out whether we are in a bubble or not, what we should do. because like probably today like we have like two different stories. On the one hand, capabilities are up, cost cost down, and we have real usage production, we have real revenue, but on the other hand, there are a lot of uncertainty. half of every venture dollar is going to AI, but at the same time we have fewer companies, we have bigger rounds, we have higher higher prices.

So, the question is not like whether AI is working, but the question is what priced correctly. And that's why I'm super happy to talk to you because you are people who are not just talk about AI, but you are price AI, and yeah yeah you sit on different layers of AI. of venture capital, and I'm Yeah, let's let's talk about what's happening in in the industry. And let's start with like quick round of introduction. Can you please say a little bit about your focus, about your thesis, and what excites you, and what keeps you up at night?

Okay, sounds good. kind of hard to follow the metaphysical talk with this practical discussion of pricing, but very glad to do it. I'm Ann Bordetsky. I'm the founder and managing partner of Alora Capital, which is a new fund focused on the agenda economy. I'm a recovering multi-stage investor, so I spent the last 5 years at NEA, which is a big platform, 25 billion under management. so, I've seen a lot of kind of the multi-stage side of what we're investing in in AI today. yeah, and it look, I think in some ways it feels incredibly sort of like frenetic right now in terms of AI investing. There's so much capital getting raised and so much capital getting deployed, but I'm still not not to hit the punchline early, but like I debate this with my myself all the time and with peers, I don't think we're over-invested in AI.

If you look at the magnitude of opportunity across every industry, across every sector, unlocking sort of wildly affordable intelligence and robotics and scientific discovery, like we're just getting started. The hardest part is just teaching yourself to sort of update the orders of magnitude with which you kind of think about both investment and outcomes in AI.

Hi everyone. My name is Serik Khuliyev and I'm a founder and managing partner of Fello Fund. I started my career as a value investor, hedge fund space. And I sort of transitioned to venture capital later. But it's been more than 10 years ago. And I I mean, I agree with Anne. I don't think we are in the bubble in the sense that you always have to look at the prospective distribution. And as long as these you know, companies powered by AI getting to you know, very formidable and very solid distribution, you you you you have something to to finance. Right? And that is the sort of thing that we look for. We're very early stage, seed, pre-seed. And for us distribution and and you know, product of of course, team is very important, but I feel like these two qualities are sort of the expression of of how team how great the team is.

So, for me, it's it's always a question of whether you have a an advantage in distribution.

Hey everyone, my name's Anu Maheshwari. I I lead venture ecosystem at Nebius. I have a bit of a unique perspective on this on this panel. I used to be a VC, did early stage investing with a fund called Stage 2 Capital, did growth stage investing with a fund called Iconic Growth. Joined Nebius about a year ago. Very much in a go-to-market oriented role, ecosystem role, figuring out how to build connectivity in and around, you know, VC investor-backed AI native companies. How can we at Nebius support them on their compute scaling? I support our corp dev team sometimes on taking a look at, you know, potential investments or M&A as well. So, just sharing not a active VC, but very much surrounded by VCs and, of course, spending time personally on looking at the AI value chain and public markets, etc. For those of you who don't know, Nebius is a purpose-built AI cloud.

We are deeply partnered with Nvidia. So, today we have a full stack cloud for AI. We build and operate data centers. We have Nvidia GPUs that we are maximizing, you know, flops and utilization as best possible for our customers. And we're a managed cloud, so we're providing different tooling both at the, you know, middle kind of developer layer all the way to the agentic layer. Started offering not only dedicated compute capacity, but also offering serverless inference, hosting open-source models. We've made a number of acquisitions that extend our reach into agentic building capabilities, inference optimization. So, really wanting to be a full full stack cloud, not not just a GPU provider.

Yeah, I'm Stewart, managing partner at M12, we're Microsoft's venture fund. I lead three areas for the fund, generative AI frontiers, new AI models, new infrastructure, whatnot, the software side. before this, I was a scientist and an engineer in the semiconductor industry, and lately I've been kind of yanked back into that that side of things. And relevant to the to why, the question of are we in a bubble or not, I actually think in a way we're in a box. we're facing constraints and and many different dimensions, not just the availability of power, the availability of semiconductor chips. We heard from this great comment from Samer earlier about the value of lithography tools driving an entire national GDP. So, part of the investment thesis I'm thinking about is what do we do about certain constraints that actually limit how much or how big AI can be in the near term, how fast it can grow.

These are very concrete not all of them category investments, but very concrete bets we can make where we see on the other side 1,000, 1 million times more token usage than today, and I don't even know if that's that's being, you know, conservative. So, where we want to be investing still is in the software that makes our AI run better, but also in the more disruptive thesis, what is the technology that we're really scaling? Is it the best technology that nature can provide, or is it simply an extension of what we've been doing the last 50 years?

Okay, thank you thank you so much. Be- Before we go like deeply into the bubble thing, let's go back to your experience. And and I'm I would like to ask ask you about your previous experience. You started making AI consumer applications very early, investing in Perplexity, Eleven Labs, soon when everyone else keep telling that application layer is a very thin wrapper. What convinced you that the modes were real?

Yeah, I think we kind of need a new name or a new definition for the app layer because I think the application layer companies that are winning at least in this AI era, they're hardly just an app layer, right? They're actually kind of platform companies and so yes, at the end of the day they're delivering amazing experiences to prosumers, consumers, business owners, enterprises, but if you actually look at sort of what drives the honestly the the value of the product, it is the ability of the teams with engineering expertise, their own research labs, their own unique research kind of and state-of-the-art models and research teams capable of producing that. Like they're more vertically integrated than any app layer companies we've seen before, right? And so they're continuously producing the research or the highly applied sort of productization of AI models, third-party, open-source. I mean, there's so many coming to market all of the time and then productizing that in terms of value and optimizing the cost of serving users at scale.

And so what I find so fascinating about the app layer, and I don't think we're done yet in consumer AI, we will see more incredible companies. There's just few of them in any wave. But they look much more like innovation factories and kind of research and engineering first teams than pure kind of consumer brand companies.

And I know that you found Groq very early when I don't know, custom inference chips were a very long shot. what convinced you to invest in that kind of technology?

Yeah, I mean, I think for us just understanding the landscape was the most important thing. We we did this breakdown and realized, you know, the next phase will likely be inference. I mean, there was no reason for any chip company to compete with Nvidia if you're really thinking about scalable business. And and so, if that's the case, then what are some some of the things that we think could could be a breakthrough, right? When we talk about inference, it's always about speed. And and it's always about this architectural view. You do have these huge energy savings. And so, Groq was was the team that had that vision. Right? And it and it almost we you know, we we we joked that basically when we had a conversation with the team, they they sort of stenciled this the same thesis that we were thinking about. and you know, from that standpoint, we I mean, it would it was easy for us to make that decision.

It's always about people. totally agree. Anu, you have a unique possibility to look at the what company really how they really spend time in production, how they really consume everything in production. What do you think is the biggest gap in how investors think about AI economics?

Interesting thing, I'll I'll I'll share a bit of a preamble before answering that question specifically. And it also relates to the the bubble topic. So, something that that we see at at Nebius, and we're not alone. I think all of the compute cloud companies see this is that there is not enough supply. There is not enough, you know, GPU or accelerator chips capacity relative to the demand. at Nebius, we see a tremendous amount of training workloads. I think initially that's been a large part of, you know, where we have grown. It's companies such as, you know, Carcer that we work with or World Labs or, you know, providing capacity to companies like Microsoft and Meta as well. And we simply don't have enough compute. Every day I have the hard conversation with a VC or a startup to say, you know, give us some time.

We're adding more data centers. We're adding more supply. And this, I believe, is indicative that we're not quite in what may end up being a peak of a bubble just because the the true demand is so high. There's a lot of capital going to new models. There's a lot of capital going to, of course, frontier models, but then all the new ones that are spinning out. There's a lot of domain-specific models, et cetera. And if we that the VCs that that we often talk to believe that these new models are going to monetize the way that Anthropic or OpenAI successfully monetize, then we have a long way to go. Similarly, for app layer companies, they're, you know, just getting to a lot of consumer growth and the token usage is expanding, right? So, that is also going to increase. So, all that to say that it feels still early, strangely, on that on that bubble topic.

And and then, you know, what might investors be missing? I think for a while investors were not paying as much attention to what the unit economics will look like for companies both at the model layer and the app layer. Specifically, at the app layer, most companies start building on top of frontier models. And today, the the token costs are, you know, ex-exponential. They're explosive. If you're successful as a app layer company, that means you're spending a lot on tokens. You might get rate limited. And now VCs are telling myself and my team that at the board meeting they're thinking about, okay, we we have to care about margins. We have to care about potentially alternatives to purely closed models. So, that's an area where I think folks are missing paying more attention now, but maybe you know, three to six months ago were not really baking in what the unit economics would look like.

And then on the model side, I think it's just uncertainty on whether all of these neo labs and spin-outs from the large frontier labs will successfully monetize. And so, all of the underwriting is based on, "Hey, this is an amazing team. They're best at building research benchmarks. There's some early, you know, demonstration or proof, but it's unclear if they were monetized. And if they can't monetize, then that's that's a challenge cuz a lot of capital was sunk into them."

Okay. Thank you. Thank you. Michael, you close to both realities, to hyperscale infrastructure and enterprise adoption. And going back to your introduction, what is actually limiting AI scaling right now? And what can be changed to have like the next AI adoption in enterprise?

Yeah, I think there's two, you know, there's a physical limitation like I knew it was just kind of talking about. Every single CSP, every basically everybody who is reselling compute capacity to some degree is doing a triaging or a ranking or a series of painful conversations of like, "How can we service the demand in front of us?" Some of it is contractual over long periods of time. Some of it is opportunistic. So, that type of market and just just everybody has to appreciate like these changes in the market come within weeks or months sometimes. And just like the Grok phenomenon in AI hardware, that is something of just the last six months, right? Before that, we would not be talking about how great it is to invest in chips. like today I predict the demand will be there. There's other ways to shift customers over to more efficient or lower cost VMs or other, you know, ways to get the tokens that those customers don't care about saving money on yet.

So that's a that's something that's going to unfold a series of like oh, we actually can use a cheaper open source model. We can hill climb we can build an agent on a lower cost model and hill climb to the eval that's meaningful. During this period of time when we're just kind of turning on to the productivity of agents, we're going to have this kind of crazy overbuying, right? And yet after, you know, after that kind of adjustment, the demand will still be there. There's not that much difference in those efficiencies to, I think, free up the capacity. So what we are trying to do as investors is one, there's some technologies or even even investing in end-to-end neo clouds that offer more efficient hardware and software together that could simply run the workloads better for a particular application. You're seeing some app companies and just being sensitive to what Ant said, an app company isn't the same thing as it used to be.

They're not sitting there and just taking like their judgment that they can't grow. They're going to go out on their own and procure capacity, sometimes buy data centers. So in that environment, again as investors, we just want to find the fastest way customers get the compute. We're getting agnostic about the technology itself. We're we're not necessarily thinking there's a particular model a model architecture a method of retrieval a KB cache technology. You got to put that aside to just see like what gets the tokens to the customers the fastest because we know it'll sell. That simple.

All right. I think we have a consensus is we are not in the AI bubble.

But, I want to ask you to name the single most mispriced, I would say, company or product. Can you, May? It should be something. It should be

After today, right? So, we we had in the public markets, we had three or four weeks or or really just an entire quarter or year-to-date of the entire AI value chain just trading up, right? And And today I think based on the jobs report that came out yesterday and people suspecting that rates might be going up, there was a large sell-off in in that market. So, I think in the public markets, a lot of things in this value chain are on sale. If If you believe us that, you know, if the demand will be there, that there's not enough compute, there's not enough memory, as agentic workflows increase, you know, we'll we'll need more memory. a lot a lot of the different companies, right? Memory, compute, data center build-outs, et cetera. they they might be on on sale like today for for example, right? So, that's that's that's more in the public markets.

And name maybe a shout great raised growth recently.

Yeah. Can't Can't say more than that.

Anyone else?

Okay, I'm not a public markets investor. So, it's dangerous for me to answer this question because I try not to pay attention to it every single day and nor should you when you're doing long-term like investing. Although, long-term in AI is probably like 2 years cuz things move so fast. But, I feel like the market is really bullish on Meta and I don't get it. Like I know I know there's they're investing so much in their talent, there's so much in their engineering, they're trying to go agent native internally. And when I see announcements from them,

lot of money on tokens.

They are. They are. There was just an announcement today that they're trying to figure out how to do kind of a Google type of financing where you basically do maybe like a second IPO to raise more for more compute to unlock these constraints. but then, and I say this as an app layer investor, when I see that their announcements essentially like amount to and now we're offering a subscription product for businesses to buy an agent, like an AI assistant from Meta, I'm like, "Womp womp." Is that like I'm sorry, like all of this compute and model development cannot be Oh, here's a totally like I don't know, expected and probably like 10th best in its category assistant that you should pay to subscribe to when all the others already exist. So, maybe there's something I don't know, but so far I'm like bearish on Meta to figure this out.

Well, we all expecting IPOs and we'll see how it goes. All these labs are moving up the stack and trying to compete with startups on the APIs. Well, how to distinguish, like how to divide like between a feature and a real company? What what's your what's your angle on that?

Oh, feature versus company. I mean, look, the the end of the day, it's like can you deliver real value to your users or to your business customers? And I think increasingly part of delivering that value and being more than a feature, hopefully a platform over time, is you have to have a really thoughtful way in which you either control distribution or control supply of intelligence, supply of AI, ideally both, because we're moving toward a world where the big tech and the big token companies, right? This is Google, OpenAI, Anthropic, etc., they're becoming a SpaceX, too, so vertically integrated, right? All the way down to the data center and compute and all the way up to distribution of the application layer. So, if that's where big tech is moving, if you're launching a startup today, you have to own something that is incredibly valuable, and that could be specific data that you have access to that nobody else has access to for a data network effect.

It could be a user workflow. It could be the particular way that you're sort of customizing and optimizing for a particular industry, right? Or a very expert use case. Like you have to put a lot of different things together to build real value for customers when the off-the-shelf models are frankly very very good and getting better. So, that's what I look for.

Yeah, I agree with that. I mean, I think Suno is a great example of that. I mean, there's there's 11 Labs obviously that launched or tried to launch the application and and it's I mean, the feedback has has basically been terrible and I think the the main reason is because I I just don't think that Labs get how hard it is to build a consumer business and how hard it is to to establish certain cadence. You have to talk to the customers. You have to learn. You have to There's There's a lot of data that that goes through this application layer. And it and I think Michael, to your point, I mean, I think what will happen is the application layer will will start to move down and and start to kind of secure this, you know, compute whether or or it's hardware, but it's Yeah, it's going to be interesting.

I I certainly am bullish on the application layer.

I think it comes down to is the company solving a pain point or problem consistently for the end user? Today with just how much AI has become intelligent, people talk about outcomes as being the really the end end result. So, it's less about like a specific task or specific workflow. It's what is the end outcome. And some of the examples in specific domains where this has worked really well, at least as of today, you know, companies like Legora and Harvey in in legal AI. Yes, general purpose frontier models can do a lot probably in the legal realm, but what makes Harvey and Legora so successful, at least, you know, hearing the law firms that use them talk about it is they're very context-specific. They go very, very deep, and they solve the outcomes for you know, a trial attorney or a securities attorney in a way that is contextual, and it's consistently driving value and and ROI as opposed to you know, if someone was making just one specific feature or one narrow part of the legal legal workflow.

And I think you can actually can monetize a feature in the infrastructure layer if you're really, really good at relating to devs. It depends on what size of a company you want to grow. If you want to grow like a trillion-dollar scale company, it's probably not going to be a feature. If you really, really solve a problem for all those other app developers, and it spreads fast, and you're super trusted, you actually can build a great business that stands on on its own.

Thank you. I think we have time for a couple of questions from audience.

Hey guys, I think I'm going to ask the question that everyone here is probably wanted to know. What do you guys think about the upcoming SpaceX IPO?

I heard it's going to happen.

It's hard to bet against Elon Musk companies.

Yeah.

The

We have Steve here who is like the first investor in SpaceX. We can ask him.

But yeah, like as a venture capitalist, we are very optimistic about like a lot of things, and probably this as well.

He's making chess moves every day. I love it.

Oh.

I I don't know if Marina knows, but I I just applied for like an allocation in the IPO. So just like with our personal money.

Thank you to all of you for your very subtle conversation. I like the question about SpaceX IPO. I have a little bit relevant question but not exactly. SpaceX initially not just SpaceX AI initially started building their own agenty AI technologies. Then they shifted more to burn metal things and started making deals with others. Roman Turchynin in the morning he talked about how they methodically building layers after layers to make things available first to the big customers when they need it. When they will get out they will have like neighbors will have enough technology to give other services. I wonder what you guys think based on all of the amazing experience you guys have. How soon we're going to see agenty AI stuff coming from that side? From SpaceX SpaceX AI. Really wonder. Given the same narrative. I'm not talking about betting against or anything. No no God forbid.

I just wonder what you guys think. Because you have perspectives that most of us don't have the luxury to see it.

Okay, I'll I'll take a stab at this. I mean I think in terms of agenty AI from SpaceX it's already there, right? I mean if you think about the Coursera acquisition and the team and the code base and the technology that they're bringing over with that. I don't know where they are in terms of specific model development. But they're going to be in the game because they to the earlier discussion have access to the one really constrained resource well the only two constrained resources which is exceptional engineers and an exceptional amount of compute, right? And if you have those two ingredients and you can go after any market in the world because it's Elon like I assume they're going to do very, very well. What I find interesting about the company, we have not seen so much in the past, but they might be the company that proves this out is essentially the internal network effect.

Like, the idea of a compounding company, right? Where the way that they're building agents for specific robotics applications, or maybe it's something related to SpaceX itself, is going going to allow them to hopefully differentiate or unlock certain capabilities that nobody else will have. And then, how do they create all those feedback loops cuz they're doing so many types of engineering at once within this company. That will be exciting to see.

Yeah, for me, I think it's it's more of a question of understanding markets. so, if you were to imagine robot-to-robot transactions, you know, e-commerce obviously is is it's great application. you know, powering these agents within these robots, within cars, where cars are talking to each other, this is where you're going to have a real opportunity. Otherwise, I I just Yeah, I mean, because it's the question of bandwidth, it's it's I you know, I always think about this new market that that's going to be here, where robots will not really need us in a sense. and for those for those robots, obviously, these these agents will be crucial.

Looks like we've just proved with our panel that we are really optimistic as venture capitalists, and we don't believe in the AI bubble, but I actually agree because, like, for me, the difference between breakthrough and the bubble is just time. Who survives long enough to see the results. Thank you so much for pricing this uncertainty and believing in the future and being optimistic. Thank you.