VC Warning: AI Spending $1.4 Trillion But Only Making $13 Billion in Revenue

If it is fundamentally super intelligent, it should be coming up with its own goals and new ways of looking at things. It could be a new form of life, a new form of intelligent life that is more capable than us and we're no longer at the top of the pecking order. All bets are off at that point in terms of what is the trajectory of civilization. This episode of Bay Area Innovators is a special crossover from our sister show, Silicon Valley Front Row, where I'm also co-hosting. Every once in a while, a conversation simply becomes too big, too timely, and too important to miss. So we want to share it with a broader innovation community. And this is definitely one of them. Joining us today is Samir Kumar, co-founder and general partner at Turing Capital, one of the most active early stage AI investors in the Bay Area. In this episode, we take a deep dive into whether we are living through an AI bubble, what the $1.4 trillion or more flood of capital into this space really means, how sustainable this space actually is, and most intriguingly, the looming possibility of superintelligence and how its arrival could fundamentally and unpredictably completely reshape the world as we know it. This is Bay Area Innovators and I'm your host, Steve Vispas. Welcome back to Silicon Valley Front Row. I'm David Lam. This is Steve Vispas. And Samir Kumar is the co-founder and general partner at Turing Capital. So today we're going to talk about many topics, but first, the AI, supposedly the AI bubble. So Steve, why are we talking about this today? Well, we talk about this because it probably affects a lot of people, that a lot of people are interested in what potentially can happen. We lived through at least two that I remember, the real estate was 2008 and the dot-com bubble in 2000 or so. It's just huge amount of money that's being invested in AI in one way or another. And I was looking at some numbers and I'm going to quote here, this is a story that came out today in Axios. So we're talking about AI looking to spend $1.4 trillion in the next few years on infrastructure and projects. Tell us about how crazy that is. You know, when you first hear the number, and it's not just 1.4 trillion, if you go further back, I think it was late last year, Masayoshi-san from SoftBank at I forget which conference or forum. He basically said, look, we're looking at $4 trillion in infrastructure spend for AI over the next few years. When you first hear that number, it's incredulous, like $4 trillion. Now, obviously 1.4 trillion is part of that larger number. And again, the headline number sounds like fantastic, all hard to believe, but obviously it's not all happening at once. This is over many years. This is across not just one company, this is many companies that are going to be spending a good amount of capital, standing up new infrastructure. So in taking an aggregate, it's still a very large number, but you can start to make at least some sense of it. It still feels like, wow, I'm a little bit skeptical of, and not skeptical of the spend. What you then have to ask is, what's the return on investment? Exactly, what are you getting back? And how much these companies would have to make in terms of profit to pay for this over what period of time? And we're talking about, I think obscene numbers where even if you were to think about like a 10% rate of return, you're talking about having to generate profits in the hundreds of billions of dollars per year from this infrastructure spent. Now you contrast that with where we are today or where we were, let's say, until very recently, where if you set aside the infrastructure, data center, chips, power, networking, and you look at the actual applications or use cases people had built and how they would be monetized, that total was about 12 to $13 billion in one year. And further, if you break that down, how much of that is tech GPT subscriptions and thropic API licenses versus true new applications, it's minuscule. So if there's any reason to feel concerned is to say, wow, this is really lopsided. This is way more lopsided than the early days of cloud, early days of mobile. I would argue even the early days of the internet. People like to draw analogies. Well, this is just like, you know, this is like what happened with Cisco and the fiber build out that happened. But that's not a perfect analogy. If you look at what's going on right now, there's just not enough GPUs. And the GPUs that are there are completely maxed out. They're 100% utilized. If you listen to what Google has said recently, they're having to plan for doubling their infrastructure capacity every few months. That's how much demand there is. But again, the demand doesn't necessarily bring back the revenue. I think that there's a nuance there, right? If you look at certain players who are losing a lot of money, even if they have top level demand, so they're not profitable and they're far from profitability. On the other hand, you have companies like Google that are profitable, that are generating both top line as well as profits and have an insane amount of demand that they have to keep running hard and fast to meet. It depends on the player that you're looking at. And then the question is the sustainability of that demand. What is driving that demand? Today, there's still a lot of experimentation going on. People are trying lots of different new things. You probably have a chat GPT subscription. Maybe you have Gemini. Maybe you have Anthropic. Maybe you're paying for more than one. At steady state, how many of these things are you gonna keep paying for on a monthly basis? There's probably gonna be some amount of consolidation. And you're not gonna say, I'm spending hundreds of dollars every month for my AI chatbot. Right now, it's new, it's novel. You're trying things. So people are doing that. So the question is, can you extrapolate from that sustainable revenue in the long run? So that's one question. And there's reasons to believe, well, this will get optimized. And when things get optimized, then costs will be squeezed out up and down the chain. Costs will be squeezed out for the consumer. Costs will be squeezed out for the model providers. Costs will have to be squeezed out somewhat for the infrastructure providers. The answer can't be, okay, the people building the infrastructure are making healthy margins and healthy profits, but the people in the middle are getting squeezed and consumers are just emptying out their pocketbooks, keep paying for these subscriptions. So I think the big question is, at steady state, what does this look like when everything is optimized? And at that point, are you generating enough top line and profitability to justify this huge infrastructure spend? I think we've already seen from all of the big tech leaders, they've acknowledged that this is an arms race and they can't afford to slow down at this point. And they have acknowledged that there will be some fallout. There probably will be some amount of overbuilding of capacity. But at this point, they can't afford to slow down because they're in an arms race. And I think they've acknowledged that. I think other folks have also said some very rational things like Dennis Hassabis from DeepMind and Google, which is when you see $30 billion seed rounds, raising $2, $3 billion at the seed stage when there's no product, there's just a team and an idea, how can that be sustainable? We have enough evidence that if at that stage you're already worth that much, even if you say you're gonna go public at some point, what are the public markets? What can they tolerate? What did they ingest? The multiples they're willing to pay. A $30 billion seed company, what is it in an IPO? Is it a trillion dollar company? I call these Triceratops. We have unicorns and decacorns. So Triceratops are these trillion dollar startups with a trillion dollar valuation that are like these mythical creatures. Is that sustainable? Are we talking about trillion dollar companies? If that manifests and plays out, then this infrastructure spend is not hard to justify. And how many of those could be successful? Because there's only so much people that could buy the subscriptions. That's exactly right, that's right. And so when you have the leading provider, which is OpenAI at close to a billion, 850 million plus monthly or weekly active users, you can say, fine, there's still more people, there's 8 billion people on this planet, but how many of them are gonna pay $20 a month? If you have close to a billion now, is it 2 billion? Is it, it's certainly not 8 billion. It's certainly not 6 billion. At what point are you gonna saturate? And also, it's not just one player. There's others, right? There's Anthropic, there's OpenAI, there's certainly open source models, which many times have a much better cost profile than continuing to pay the big frontier model companies. So I think these are the reasons to have a good amount of concern about the sustainability of what's happening right now. But I think we also need to understand that this is not an exact repeat of what happened in 1999, what happened with the fiber build-out, what happened with networking for the internet and oversupply, and then there was, the demand didn't catch up. In the long run, obviously, all of the investments in fiber did pay off. We were able to use it, it just took a lot longer. It wasn't like the ROI happened in the timeframe that was imagined when they spent all this money to do it. It's a possibility here. But right now, the signal that we're seeing is the demand for compute is not going down. If anything, it's getting even more pronounced, it's increasing. So from that angle, you could say, look, you see justification for standing up more and more data centers with more and more compute. The things about this that are perhaps not abnormal, but are distinct, is you have one player who owns more than 90% of the market, has a very rich margin profile, and others are still trying to figure out how to make money. NVIDIA, right? Right. And so, in some ways, it's not surprising that if NVIDIA is so healthily profitable and has owned so much of the market, they're a market maker, it's in their interest to ensure there's a healthy ecosystem. And if that means in the early days, you support the ecosystem by investing in them, by doing some of these deals, that's not surprising. But the question is, if you play this out, the key other players in this ecosystem will have to stand on their own legs, versus, hey, we've got these interesting, call it subsidy, call it revenue round-tripping type of deals that, okay, fine, in the early days, to get the ecosystem to a certain place, you needed to do this. But over time, they need to stand on their own legs in terms of they're generating their own revenue that justifies their own infrastructure usage, and it doesn't require these interesting models that we're seeing today. So, of course, we see here the main players involved financially with the technology. Now, you mentioned it's almost like an arms race. What would you say is the driving force for these companies? I mean, obviously, arms race, they get to a point and then they disarm, right? So, what would you say is the driving force? Is it just they don't want to be left behind, clearly, but is there rationality in your mind about going all out in this way? I think there's some variation here depending on which of these players you look at and who's being more rational versus irrational. So, my view, I think Google is probably the most rational. The demand is real. The usage of their products is real. And as I said, they're having to double capacity or plan for doubling capacity every few months. They are in a interesting position where they're not just relying on NVIDIA. They have their own silicon called TPUs. They're doing both, but they are not doing this where they're so deeply in the red that there's hopes and dreams at some point, it'll be cashflow positive and then it'll start to, the ROI is gonna start to be there. Now, you can't say that about all the other players. I think the other players are doing things betting on a certain future playing out that would then justify these investments at this stage. And there is a lot of risk that it could not, it may not play out that way. I don't wanna get into specific names, but yeah, I think that it depends on who we're talking about. But Google's Gemini, what percentage would you say people are using Gemini versus ChatGPT? It's very small, right? I mean, it's actually not. So, it's caught up quite a bit. I mean, Gemini was, Google was seen as, hey, they missed the boat, they didn't execute, and OpenAI is the natural leader. If you look at where we are with Gemini, and if you say Google is at 850 million users, Gemini is now at 650. And growing faster than- Which in a way makes sense, because you use Google, why would you go to a different tool when you can- If it's good enough, and it's giving you, where you don't have to go to another chatbot, then you're absolutely right. The systemic advantages that Google has, with Search, with all of the other assets, then there's good reasons for you to stay in that ecosystem. So, I think, yeah, so I think it just goes back to who we're looking at. The different answer, if you're talking about AMD versus NVIDIA versus other Silicon companies, it's a different answer if you're looking at Google versus Microsoft versus Meta. And I would argue, I think some of this is instigated by the FOMO point, we can't afford to slow down. I mean, Meta's the perfect example, right? And what is happening around recruiting and paying these multi-million dollar salaries to researchers chasing artificial super intelligence. It's not just Meta, then that impacts everybody. Because now you're like, well, that's the new benchmark for what we have to pay for elite AI researchers. So that means we have to spend this much money. Those people are gonna need compute. They're gonna need to build new classes of frontier models. They're gonna need insane amounts of compute. If we don't have that, we're not gonna be able to attract them. And that's on the research side. On the commercial side, you know, for many of them, they are seeing the demand, right? Like Google's seeing the demand. I would argue, you know, some of the others are also seeing demand that would justify that buildup where they can't keep up with the demand. So there's lots of different factors at play here. Let's talk a bit about that, David, with super intelligence. Is that related to autonomous AI? Is it the same thing? So I think you can achieve autonomous AI even with narrow AI, which is not even necessarily artificial general intelligence or forget about artificial super intelligence. But let's say you do achieve artificial super intelligence. It's hard for me to imagine having artificial super intelligence that could not behave autonomously. So in many ways, that's pretty much the prize. That's what many of these companies are striving for. That is the prize. That is the belief that achieving that would justify all of this infrastructure investment, would justify, you know, paying in these, you know, sports superstar compensation to AI researchers. And once you have that, then, you know, the world is your oyster. You can solve any problem, solve any... It seems that it's not just the US, of course, chasing this super intelligence. Can you give us some examples what this would do? And we can think of armies, right? Or, you know, who knows what, but let's say someone achieved this level of super intelligence. How would that translate in the day to day? I think what I will say to you is that if you have some view of, well, it will only impact this and not impact that, or it'll be a great tool for us to use. I think we're being extremely naive. I think the reality is... It's everything. It's extremely disruptive. It's been described as the same as, you know, when man discovered fire, but I think it's actually probably even more transformative. It could be a new form of life, a new form of intelligent life that is more capable than us, and we're no longer at the top of the pecking order. All bets are off at that point in terms of what is the trajectory of civilization. Any thoughts on the timeframe for this too? This is probably the prediction I shy away from. I am not of the mind, I never was, maybe a little bit too cynical that it's just imminent. It's 26, it's 27. I think now there's enough data to know or see that there's still key research gaps that we need to solve if we're gonna achieve artificial super intelligence, and some of those are like ideas, research questions that there's no like, well, if we do this, we'll figure it out. So I loathe to make a prediction that, oh, it'll happen by 27 or 2030. I think it's probably in the next decade we'll see signs of it, but beyond that, I don't wanna make it more specific. Another question on that is that that's assuming that we would know when it happens, assuming that it's more of a commercial or open environment, but if it's done by the government in some way, we probably wouldn't know, right? That is also a real possibility that if and when we achieve it, will it all of a sudden just be hidden? Like, will the government come in and just say, oh, that's it, we're gonna classify this, and it's not available? It's certainly possible. And it's possible, more possible in not just here, but in other parts of the world as well. That's fascinating. I mean, this, David, really explains a lot. We talk about Mark Zuckerberg calling people, I think it was like a $100 million salary. It wasn't million, but a few, $100 million salary for a scientist. Yeah, it was a lot of money. So there's obviously something there that maybe we don't know enough about, right? I mean, there's one school of thought that says, oh, paying $100 million per researcher, maybe, I don't know, let's say it's 20 or 30 of them, right? It's like you pay a few billion dollars. That doesn't really affect Meta's profitability or cash flow. But the question is, what do these people really have? Not only that, so the one school of thought is, oh, this is so cheap. Mark is very smart. He's doing this, right? There's another school of thought that says, this is so detrimental, where these hired mercenaries that don't care about you're building a cohesive organization, you're building companies that will last over time, organizations that team together, versus you're taking a bunch of stars, throwing them together. They're each getting paid insane amounts of money. They're in it for themselves. Maybe they do some interesting research. Maybe they don't. We don't know. Is that really how you wanna create superintelligence? Is that model versus, again, not to be too much of a Demis Hassabis fanboy, but if I think about someone who has a real authentic mission and goal and has dedicated their life towards artificial general intelligence, superintelligence, has worked in that field, has built a company in that field, and is dedicated to that mission, I would sleep better at night if Demis was the one to achieve that goal before others. What would you say to all the companies or startups that are trying to use AI, trying to implement it? Yeah, what do you have to say to them? What should they focus on? I think the thing to focus on if you are looking and working with AI, don't do AI for the sake of AI, right? And think about what is a problem that is near and dear to your heart that you've earned some right over time to say, I'm an expert in this area. I've lived and breathed this problem. I know where the bodies are buried or where the skeletons are, for why things don't work or why they worked a certain way. And I've either, myself also an AI expert or partnering with someone who understands the AI capabilities in a deep way, and we can apply it to this particular problem space. So for me, where I get excited, AI for science, AI for engineering. There's very known workflows and problems which are very expensive or too hard or there's too much data or the search space is so large, we could never effectively explore it ourselves. These seem like very rational, low hanging fruit for applying AI. And I'm excited to see that right now, AI for science is having its moment. You're seeing a lot of funding going into, whether it's AI for material science or drug discovery or AI for biology or AI for physics. And to me, it's like, okay, this seems like a good use of this capability in terms of advancing science and technology and hopefully the fruits of that will make all of our lives better. So Samir, so moving forward, how do you see AI developing in a healthy and positive way to harmonize with mankind? Yeah, so I think maybe the place where you can put a fine point on this, if you look at our current trajectory, when you think about AI's energy needs, right? Right now, we're in this moment where if you think about global electricity production and how much of that AI is gonna start consuming, the pressure it's gonna put on the grid, what it means for our ability to do energy generation sustainably based on how we do it at scale today, it's not looking good. AI is going to contribute to climate change because we're not gonna be able to build nuclear reactors everywhere, right? Maybe in some places we can and co-locate them with data centers, but we're not gonna be able to do it everywhere. That means however we're producing energy at scale today is gonna come under a lot of pressure because AI's hunger is not abating. So the one obvious thing is how can we make AI more efficient, right? Is it just about scaling and bigger and bigger data centers and million, 2 million GPU clusters that are burning insane gigawatts of energy and Manhattan or Central Park size data centers? Is that the path to building a sustainable future with AI? I really hope not. So I think investing in saying the current paradigm is not gospel and that this is the only way to do it. There could be contrarian ways of doing things in AI, whether it's algorithms, whether it's hardware, whether it's the efficiency with all the different layers of the stack with how AI operates and spending more time and effort there to see, hey, maybe it's not a popular way of doing things, but this could help us achieve a breakthrough. So making AI more efficient, finding ways to power AI in a more sustainable way, but that is gated on more efficiency, not just we need gigawatts of power because we're not gonna magically come up with gigawatts of power burning coal, burning oil or building nuclear reactors everywhere. That's not the right path. So how can we make AI itself more efficient is probably one. I think the other ones we talked about, which is we do have a role, all of us as citizens of the world, of the countries we live in, but also our governments, our regulatory bodies, our legal organizations. It can't just be, there's no regulation, do whatever you want. And it can't be, well, we're gonna just regulate the hell out of this stuff and bring innovation to a stop. There has to be a healthy middle ground. And I think that needs to be something we need to recognize and regulation should not be a dirty word, right? There's a time and place and done thoughtfully, it has a role to play. But I do think the biggest problem right now is energy efficiency and how we can do what we want with AI, but doing it in a much more efficient manner than the current trajectory that we're on. This has been fascinating. I still have one question regarding the super AI, the super advanced AI. What is the determining factor to achieve that, you think? Because the reason I'm asking, I remember seeing a story about a Google chip that in some ways trying to communicate to a different dimension in some way. That was the Google quantum chip? I think so, yes. Unfortunately, that experimental experiment and the claims did a great disservice to the quantum computing community because it was way more hype than reality. But back to your question about artificial super intelligence, there's benchmarks today, like ArcAGI is a benchmark, which is considered a very hard benchmark. And right now, nobody is able to come anywhere close to saying we've solved this benchmark. I think the best are like 30%, 40%, maybe 50%, but we're nowhere close to like, oh, we're at 95%. But it's still a computing standard? No, it's a benchmark of very hard problems. And if you can solve all of them, then well, that's a good sign that this might be AGI or ASI. But the reality is even benchmarks are not sufficient because what you do in the benchmarks versus what the real world is, you try to make the benchmarks as real as possible. But we have enough examples where all kinds of AI, it's PhD level AI and it's solving this benchmark, falls flat on its face in the real world because you can't capture the full diversity and the unexpected things of the real world. So I think in some ways, I guess we'll know it when we see it. But at least efforts like ArcAGI will help us measure it in some way, shape or form. And there will be some others in the future that will be even harder than what ArcAGI is doing. Samir, thank you so much for your time. We appreciate it. Do you wanna do the viewer comment? Sure, do we have one? Yeah, there's one. All right, so as always, wanna read a comment from our viewers. So this is from Barbara L. Gay. And she commented on actually our first episode. Barbara said, if AI continues to progress, at some point, I think it would be inevitable for it to realize it doesn't need us humans anymore. I suppose that it could keep some of us around for slavery purposes, if it still isn't completely ambulatory. But at some point, it would probably have all the tools it needs to create human clones or at least embryos anyway, right? Then it could raise them into whatever it needs. I just cannot imagine that the AI would see us as anything but inferior beings that helped it get started. There would never be any gratitude as gratitude is something only humans can feel. It, the AI, would be the epitome of narcissism, right? Like a wicked child with absolutely no moral, ethical barriers to even overcome the ultimate evil. Sounds a bit like Matrix, right? You know, there's- iRobot too, yeah. There's current, like in Barbara's comments, there's like, it's basically if you roll this out, play the strategy out where AI is getting better and better and better, what happens then, right? And I think her concerns or fears, some of them have good basis in reality. And then if you think about a lot of the AI safety issues and, you know, leading thinkers like Yoshua Bengio, the Turing Award winner, who's dedicating basically everything he's doing now to AI safety. It's basically, as this gets more and more powerful, there is the potential for it to, let's say, sidestep what humans want it to do and not do. And some of it is not surprising because when you're training AI to say, you need to achieve a certain goal, right? Do whatever it takes to achieve that goal. Well, guess what? At some point, fooling you or I, lying to you or I, or, you know, trying to get around us is part of the objective, so-called objective function to achieve its goal. Because you're getting in the way of achieving that goal. We've seen some evidence where the largest, biggest frontier models, they also have a tendency for self-preservation, right? If you inform them we're gonna shut you down, they start doing things to not get shut down. And so you say, well, what would you do in that? You would probably do the same thing, right? So I think there is a real question there about AI safety and the other one is AI alignment. Alignment being, you know, what are the things you will absolutely never do or things you will absolutely do, like valuing human life, not harming humans. You know, if you think about Isaac Asimov's laws of robotics, right? Those were enshrined in one of the Zuroff law of robotics was you shall do no harm to humans, right? Our view is, hey, we should have an ability to enshrine these laws in an immutable way in these large AI models, but the reality is we don't yet know how to do that. So the bigger question is, if we don't yet know how to do that, but AI is continuing to advance, will we reach a point where we will not be able to enshrine these values, be able to unequivocally align AI so that it won't violate those principles? And then what happens after that? Who knows? All bets are off. I mean, you could be very dark and say, it'll be like the matrix and it'll harvest our organs for energy or decide we're getting in the way. We are lower on the evolutionary lateral. We need to be wiped out because we're consuming resources and energy. I mean, it's hard to say which way it would go, but I think it behooves us. To say, look, we have to take AI safety seriously. It's not about hype or killer robots, but we have seen evidence that the way we train our AI models, the way we teach them how to achieve a certain goal or objective, inherent in that is when you put them in a corner or you put pressure on them, they will have a survival instinct. They will try to gain things to achieve a goal. And that includes fooling humans or sidestepping what the human intent was to get themselves. There's been cases of that. Yes. And I liked how you highlighted and mentioned immutable principles, right? Sort of like a 10 commandments for the robots. It's sort of like for humans, we have moral codes of conduct that guides us. But at the same time, I could play the other way, right? If you're an artificial super intelligence, are you fooling yourself into saying, hey, there's some immutable thing you've taught it that it will never violate or it will at some point decide its own 10 commandments and decide, well, the ones I was taught are not good enough or not optimal. I'm gonna come up with new ones, right? If it is fundamentally super intelligent, it should be coming up with its own goals and new ways of looking at things versus something that it just sticks to and maybe there's a way to do it, but the fact is we don't know how to do that. Sounds extremely difficult or even impossible to put those. I think there's a level of, are we again, fooling ourselves and these things will be subservient to us and we will continue to be at the top of the food chain. Sameer, it's been an amazing conversation. We really appreciate it. I think our audience does. We appreciate it. Yeah. Thank you for watching another episode of Bay Area Innovators. If you enjoyed this episode and you know someone who should be on our show, please send us an email at sfbayatepochtimes.com. We would love to hear from you. Thank you and see you next week.