# How AI Is Reinventing Computing from Chips to Power The a16z Show · 2026-08-28 · 54 min · https://www.youtube.com/watch?v=Zx1Ec8LWFeM ## Transcript Ben Horowitz We have a whole new technology. That's the most important technology ever. And you need a whole new infrastructure. Martin Casado Normally, when we talk about the infrastructure world, we're talking about the servers, the storage, and the network. Here, it goes all the way down to the copper mines. That's how widespread this thing is going to be. Ben Horowitz It used to be, when you built something, it was an engineering problem. Here, it feels like it really is a resource limitation. So whether it's tokens or not, we're pouring a ton of money into systems, and then those systems are producing a result. Right now, we're bottlenecked on the systems' ability to actually match the resources we're pouring into them. Raghu Raghuram The leading memory vendor said the demand they have today will take them 3 years of capacity to supply. Erik Torenberg If this fund does what we think it will do, how do we see the world in 5 to 10 years? Ben Horowitz America wins in the infrastructure, and that would be awesome. ### Introducing the Machine Age Fund Erik Torenberg Ben, Martin, Raghu, welcome. Ben Horowitz Thank you. Erik Torenberg All right. Thank you. I want to start with a Marc quote to introduce this new fund: “This is the biggest technological revolution of my lifetime. This is clearly bigger than the internet. The comps on this are the microprocessor, the steam engine, and electricity—or maybe the wheel.” Guys, the Machine Age Fund. Please introduce it. Ben, start us off. Ben Horowitz Basically, what’s happened is we have a whole new technology that’s the most important technology ever. Every time there’s a dramatic new way of using all of the things that we love in infrastructure, you need a whole new infrastructure, and never has it been more high-impact than it is on this one. ### Why Founder Interest in Hardware Just 4x'd Not only do we need new chips and new system software, we need new ways of doing power, and we need to replace copper. It’s absolutely everything, so it’s a very exciting time. Particularly for the hardware aspects of this new era, we needed a new approach. Martin Casado Yeah, I would agree. Normally, when we talk about the infrastructure world—at least in computing—we’re talking about the servers, the storage, and the network. Here, it goes all the way down to the copper mines. That’s how widespread this thing is going to be. That’s number 1. Number 2, I think what we’ve seen over the last 3 years is a steady increase in the capabilities of the models, where the model is no longer the bottleneck. In fact, using AI, these models are getting better faster and faster and faster. Now, the bottleneck is everything I call south of the model, and so that’s why we need to work on that. The only thing I’d add very quickly is that we tend to follow founders, and we’ve been watching the number of very strong teams going after complex hardware problems increase over the last couple of years. I don’t know the actual numbers, but I was trying to estimate it over the weekend. I think maybe 5% of the deals from top founders would have been hardware before. Now, I would say north of 20% or 30% right now. The founder community, which tends to be much smarter than the VC community, has identified this as a very active area for innovation, and they’re responding. Ben Horowitz I think 5% is probably generous. Martin Casado Yeah, it’s very low. It’s very low. Raghu Raghuram Yeah. 3%. Erik Torenberg Explain some of the macro conditions that have led to this change—the surge of founders pursuing these ideas. What are they seeing that’s enabled this? Ben Horowitz The obvious thing is that demand for AI is basically infinite, and as a result of that, every part of the supply chain is under duress. Everything, including materials used to make things like memory. It’s also interesting whether there’s something unique about AI. Because demand is infinite and growth is infinite, what you tend to worry about is the margin of companies—how efficient it is. Normally, you worry about growth: Can I just get people to buy this stuff? You don’t have to worry about that here. The question is whether you can do this in a way that’s profitable. ### How Do We Know Demand Isn't a Hype Cycle? A lot of the efficiencies are actually strictly a physical limitation of hardware. Even the business model of the AI wave is putting a lot of stress on the existing systems because they weren’t built for AI. They weren’t built for those workloads. There’s just this global observation that we need to change the core components to get that efficiency, help drive the growth, and drive the value of the businesses. Erik Torenberg How do we know that demand is actually outpacing supply here, rather than this being another hype cycle? Raghu Raghuram There are any number of signals today. First, some of the smartest judges of demand are cutting huge purchase orders. If you look at the hyperscalers, their capital expenditure has been exploding. Next year, it’s supposedly going to reach $1 trillion collectively across the big hyperscalers. This year, it’s about $700 billion. If you think about the hyperscalers’ position in the industry, they see demand from everywhere. They see the frontier labs wanting their compute, the AI-native companies, the enterprise, the U.S. geography, and the international geography. If anybody has visibility, it’s them, and they’ve been jacking up their capital expenditure like it’s never been seen before. That’s a clear, clear sign. Secondly, if you look at the companies that we see on a day-to-day basis, they’re all ripping. All the application companies—the growth is insane. The frontier labs—the growth is insane. It’s been documented. I would say that, on the demand side, the signals have never been clearer that this is not a hype. To top it all off, prices are going up. Martin Casado We’ve never seen prices go up on chips. Ben Horowitz Typically, prices went down. They always went down. Martin Casado Yeah. Raghu Raghuram If you look at the price curve, it went like this and then went back this. We know that only 5% to 10% of the addressable market has been addressed today. Martin Casado The supply, if you look across the board, is basically all booked out to 2028. It’s so bad that we’ve actually seen multi-day auctions for a few thousand GPUs. The other side of that, of course, is demand. As Raghu said, we’ve seen the fastest-growing companies in the history of the industry. ### Sold Out to 2028: The Unprecedented Supply Crunch Raghu Raghuram Also, the unit of work that AI can do—the value of that unit of work—keeps increasing. But underneath the covers, the number of tokens that are consumed is going up by orders of magnitude, right? If it’s 100 tokens for chat or an agent, it’s thousands of tokens, right? So you’ve got expansion of both sides of demand. One is the unit of work is becoming more and more consumptive of tokens, and then secondly, the number of people therefore that are going to benefit—it’s not just the developers, it’s going to be all knowledge workers, and then all of beyond that. So that’s what we see. Erik Torenberg You said the key components in supply are sold out to 2027, maybe 2028. What does it mean for an entire industry to be sold out that far? I don’t know if this has ever happened before. Do you guys recall? Ben Horowitz Remember in the internet days, when we were doing a massive buildout, the majority of what was actually being put in the ground was speculative and dark. Remember dark fiber? Here, basically every GPU that’s being created is already presold. Martin Casado Yeah, we weren’t quite there. There was a lack of bandwidth in the ’98–’99 timeframe, but there wasn’t that much real demand for it because there just weren’t that many people on the internet. It was a two-sided thing. The companies were all rushing there and theoretically needed more bandwidth, but there weren’t necessarily enough users on the other side to consume it. To really consume a lot of bandwidth, you have to do high-bandwidth things like video, which weren’t really viable for a number of reasons that had nothing to do with how much bandwidth was in the data center. It smelled similar, but it wasn’t this. This is like we’re flat out, and people are reselling GPUs for 4 times what they bought them for. It’s just not the same. Ben Horowitz We’re also out of power and cooling. On top of that, it’s really hard to build because there are these incredible political headwinds going into it. It’s really unprecedented in my career that we’ve had anything like this. Raghu Raghuram I want to give you a quick anecdote. I was talking to the CFO of a large public company that had historically been very resistant to going into the cloud. They had a lot of servers, and they were doing an inventory check. They realized that the memory in their servers had increased so much that it could fund the entire migration to the cloud. I just feel like we’re in a very unusual situation. Ben Horowitz Yeah, that’s right. We’re out of many things: power, cooling, memory, GPUs—you name it, we’re out of it. Raghu Raghuram The flagship conference for the industry is called Hot Chips, and it’s going on at Stanford. The leading memory vendor said the demand they have today would take 3 years of capacity to supply. It’s just today; it’s not even future demand. Erik Torenberg In terms of being out of everything simultaneously, is it because people just underestimated how good the models would be and how useful they would be? They just couldn’t have foreseen the demand? Ben Horowitz Well, I don’t even think it’s that. This stuff came out of nowhere, right? We're only 4 years into this. So even if we had a perfect oracle, once it started working, I don't think— Ben Horowitz We could have built the capacity. Raghu Raghuram We could have built the capacity. There's no way. And we're talking about chip cycles, which tend to be 3 to 4 years. We're talking about breaking ground and building data centers, which is 4 to 5 years. We're talking— Ben Horowitz And connecting, breaking ground, building them, and having a power source. So you either have to build your own power, or usually both: you've got to build your own power and have a power source, which is not easy. Raghu Raghuram The ML industry itself, if it's growing at 20% or 30%, has a great growth rate, right? And it's being connected to an AI software industry that's triple digits as the base, you know. Ben Horowitz So you can see the disconnect, right? It's just a wide gap. Erik Torenberg And so why didn't this fund exist 5 years ago or 7 years ago? Why was it not a great category to invest in, in the same way, Ben? Ben Horowitz Well, I'd like to think we're just in time, but we probably would have been well suited to have it at least a couple of years ago. Raghu Raghuram I will say you could actually point to basically every epoch to an independent company that came up, right? Martin Casado Clearly, in the move from the mainframe to the client-server, we saw a bunch of companies come up. In the move to the internet, we got Cisco and Juniper. Even in the mega data centers, which, by the way, was largely driven by the incumbent cloud providers verticalizing, you saw the rise of Arista. There has been the ability to invest in silicon and hardware, but it's been relatively minor because the change has been relatively minor—one chip company, one switch company—where here, everything is different. And so I agree with Ben: we probably could have started a little bit earlier, but the amount of change is so high now that it's just an obvious thing. And the other thing is, the demand for intelligence is so voracious, with really no end in sight. Every company that's adopted it is growing very fast in its usage, most companies haven't adopted it to a high degree, and consumers are just getting started. So the demand for tokens is probably going to grow close to 1,000% a year, which you cannot grow supply that fast. We're not— Raghu Raghuram The amount of work we're going to have to do across the board to get to the point where we can grow infrastructure at that kind of rate is pretty vast. So I think there's a lot of investing opportunity on the way. And, by the way, the other thing is, all the architectures of the hardware systems were built for a whole different era of computing. So more than just needing more capacity, we need capacity to build. There are lots of opportunities to build different kinds of infrastructure. Martin Casado Yeah, they're all reaching the physics limits for what they were designed for, right? Like what Ben was talking about with copper, and so on and so forth. You could go across every one of these categories and find, “Okay, this is the limit of this type of technology.” So now you've got to get some technical breakthroughs to get to the next one. Erik Torenberg Yeah. I want to dive deeper on the demand side for a second. As we've moved from chatbots to reasoning to agents to multi-agents, each step has multiplied the number of tokens a single task takes up by increasingly larger orders of magnitude. Raghu Raghuram Yeah, nobody likes to use AI more than AI. ### Tokens, Scaling & Why There's No Natural Regulator Erik Torenberg Why does that keep happening instead of leveling off? Do you see that happening indefinitely, just continuing to— Martin Casado Well, for sure, right now, if you look at the way we're achieving scaling, the way we're doing it is through a lot of inference, so through a lot of tokens. If you think about what RL is, it's a lot of inference. If you think about chain of thought, it's a lot of inference. If you think about long-running agents, of course it's a lot of inference. That's just basically been one of the approaches that we've been using to scale. I think if you want to step back and say, “What is the macro trend here?” it used to be that when you built something, it was an engineering problem. You'd throw a bunch of engineers at it, and that didn't scale. There was a natural law of engineering physics, which is what The Mythical Man-Month came from. Here, it feels like it really is a resource limitation. Whether it's tokens or not, we're pouring a ton of money into systems, and then those systems are producing a result. Right now, we're bottlenecked on those systems' ability to actually match the resources we're pouring into them. I think tokens are probably where we are on the scaling curve right now, but we don't have a natural regulator like engineering, as we did before. So I think we should expect this to continue, and we have to build a supply to support it. Raghu Raghuram Yeah. Like, the simple way to think about it is that any problem you have can be solved with enough infrastructure— Martin Casado Basically. Raghu Raghuram GPUs, power, and money. Until we run out of problems, we're not going to run out of demand. That's the challenge. Martin Casado AI's answer to getting better and better is to use more AI, right? Inference is one basic building block that it keeps using over and over and over again, and so that's why these tokens multiply at each step. Raghu Raghuram Yeah, even the autocatalytic effect—the idea of using AI to create more AI, like creating a GPU kernel, of course, is just using more AI as part of the process. One way that we think about it is that, in the past, money would come in, you would have an engineering problem, we knew that it took 2 years, it normally failed, and there was a natural governor. Then you got the product on the other end. Here, there's nothing between the money going in and the hardware creating intelligence. So now we're just limited by our ability to create supply. It's a very, very different— Martin Casado Yeah, that's right. Raghu Raghuram So that's the cycle. As long as you have the money, the GPUs, and the data, for the foreseeable future you'll be able to scale these things. Erik Torenberg It's fascinating because, over the last decade, it feels like there were so many people saying the pervasive sentiment was that there was too much money going to startups. We were overfunding these startups. There was too much money in venture capital. ### Agents as a New Kind of Employee: The GrokBot Moment Say more, Ben, about what that means, because there used to be this skepticism that the more money you put into the industry, the bigger the outcomes would be. And now we were saying at the off-site that, to some degree, the market is as big as we collectively contribute to it. Ben Horowitz Yeah. This is—look, the one thing we all knew in the startup world is that if I have a 2-year lead on you and you try to catch me by hiring 1,000 engineers, you're going to wreck your company. That never works. It's The Mythical Man-Month. Nine women can't have a baby in a month. That never works. Now that works. But it's not hiring 100,000 engineers. It's taking $3 billion and lighting up a magnificent cluster, and then all of a sudden, Grok can come out of nowhere and you're like, “Oh, all of a sudden it's real.” Or Kimi, or what have you. These leads—you can throw money at the problem, and you can throw money at almost any problem, and that works. So that is just completely different from anything we've ever lived through. We're all psychologically adjusting to this. Erik Torenberg The ChatGPT app has a billion weekly active users. There are about 30 million developers who are using a relatively big portion of compute demand. How do we think about compute demand needs now and in the future in light of what people are actually doing with AI? Raghu Raghuram And now you have Grok bot, which is kind of what happened with coding, kind of happening with all use of computers via bot. So we're in a whole other wave of demand, and most certainly there's going to be more to come. It does seem quite unlimited at the moment. We haven't even gotten into embodied AI or robots, which are going to be another source of demand. Erik Torenberg Martin's an expert, but my understanding is that the bot uses computer use, which is just like— Martin Casado A human being sitting inside the computer, typing away. Erik Torenberg I literally used it over the weekend to update my credit card with a bunch of services that I'd been lazy to do, and to cancel a bunch of subscriptions. This is not coding or whatever. This is true computer use. Raghu Raghuram All of a sudden, you're creating half a billion knowledge workers, except they're all sitting inside the computer doing what? Ben Horowitz I do think that Marc Andreessen is right. The right analogy here is the steam engine or electricity, in the following way. We've introduced this new thing that you can put to work, and there are some very obvious applications now, but there's probably 30 or 40 years of throwing computers at problems. Anything with a clear reward signal, and we're just starting: we've got language and code, that's it, and we're just starting with computer use. What else are we looking at? Science, materials, biology. Of course, creativity is a massive use. We're at the very, very early part of a very long journey, and we've removed this key bottleneck, which is traditional software engineering. Of course, bottlenecks will move and there'll be more complexity elsewhere, but I think we're at a very early point in a very long run of throwing computers at problems. So let's expect this compute need to persist for decades. Erik Torenberg But because we mentioned it, Marty, talk about Clawdbot, because we were talking at the offsite about what struck you about it. Obviously, we're involved in every possible way you could be involved, but what did you find so interesting about it? Martin Casado I think we've, as an industry, gone through multiple realizations of how AI enters our lives. Very early on, we were like, "Okay, well, you add AI to a product," and it's whatever—it's like a search bar. Then you chat with it and it chats back, because that's the traditional way to do it. Then OpenClaw showed up earlier in the year. With OpenClaw, I said, "Okay, well, maybe it just being like Google but better isn't the full embodiment of it. How about we'll have it be a standalone thing, but it'll be an extension of you? It'll share your keys, it'll know your passwords, and it'll just do stuff that you would do." So it's kind of an extension of you, but it's more like a human. What I think Clawdbot got really right is: No, how about it's actually an employee? Now you have this thing that's an entity, and it doesn't have special access to your keys or whatever. It has its own computer and its own browser, and because these are the smartest models in the world, it can do whatever an employee can do. ### What's Actually Bottlenecked Right Now It's interesting because now, if I want something done, my first thought is, "Well, can it do it for me?" Often the answer is yes, even if it's something you wouldn't expect. The obvious ones are things like managing my calendar or booking a meeting, but there are non-obvious ones as well. For example, I'll have it read through my email and do triage. I don't tell it how to do that, but it knows to check with me before actually doing the triage. These things are sophisticated enough that you can give them a relatively high-level task, and they'll do sophisticated things as a result. Erik Torenberg Ben, I know you're thinking a lot about this and how this works in the organization. You think a lot about culture, of course. What are your thoughts here? Ben Horowitz I think if you just look at us, it's like having a new kind of employee, and there are going to be a lot of them. Just like with our regular human employees, we spent many, many, many years figuring out how to work with them, and now we've got these other kinds of employees. There's a learning curve with them. They can burn a lot of tokens, spend a lot of money, and get nothing productive done. They can forget stuff. They can make stuff up. They can have good behavior, and they can have bad behavior. They can create security problems. There are all those aspects to it, but they can also be super-duper productive. I think figuring out how to integrate them and have them work nicely with the people they're working with—the actual humans—is all something that we're learning how to do. I don't want to sit up here and say I've cracked the code. The whole firm is completely automated now, and I'm going to slowly get rid of all the humans because I can. That's not at all where we are. We're much more like, "Okay, how do we make all our humans superhuman without wrecking the place because the bots got out of control?" We've tried a couple of different ways to get agents into the system, if you will, and eventually it was Martin's insight: Just treat them as people and get it done. That's what we're doing, and that's turned out to be the most durable way of getting this thing going inside an organization. ### What "AI-Designed" Infrastructure Actually Looks Like Erik Torenberg I want to go back to the supply side and go deeper into the bottlenecks. We were talking about how, in terms of the data centers, the chip architecture, system software, and the facilities themselves, none of them were designed with AI in mind. What would it look like for them to be designed with AI? What's the mental model for thinking about what that could mean? Raghu Raghuram If you start with the statement that the original model of infrastructure for any of these models has to change, you can go category by category and see how it breaks. Then you start unlocking the bottlenecks in each one of these things. Eventually, you have to get to a system where, if you look at what an inference engine does, it takes up a lot of memory and generates new tokens along with the compute. You can think about how to optimize all of this. What does the memory need to be? What does the compute need to be? How do they need to talk to each other? How much power does each of them need? If they need all of this power, how do you cool each of them? Then how do you put collections of these things together? That is the exercise that's underway in the industry right now with a lot of the founders. They're breaking down the problem into its fundamental components and saying, "What is the exact nature of the compute that's getting done?" It's going to be matrix multiplications. How do I optimize my compute around that kind of a scenario? They all need to progressively generate these tokens. What is the best way of hierarchically arranging this memory? How is the power consumed? Then you've got to connect it together. What are the ways of connecting it on the same chip, but also across chips and across data centers? How much power does each of these data transmissions take? You have to progressively break it all down and rebuild it from these fundamental building blocks. That's what we see underway, and that's where we see the opportunity. Martin Casado Let me give you an interesting mental model to think about how the landscape has changed. Today, to build a frontier model costs, let's say, $3 billion to $5 billion, and that's to train it. The inference has to pay back at least that, of course, in order for any of this stuff to be viable. Let's say 2 times that, so now inference has to make $10 billion. If you can save 20% in efficiency on that, that's $2 billion, and you can easily build an ASIC for $2 billion. We've actually gotten to this interesting point in the industry where it makes sense to build an ASIC per model, just because of the amount of capital investment in that model. Unlike traditional software, which has a lot of state and is very dynamic, these models are fixed. The model weights are fixed. ### Rack Power, Liquid Cooling & the Data Center Redesign We don't know if the world goes to per-model ASICs, but it gives you a great mental model of how you would evolve the architecture to be far more bespoke for these massive capital investments we're making. I don't think in the history of the industry we've ever created a digital artifact with something like $5 billion going directly into that artifact. This is going to put the greatest demands on hardware that we've ever seen. Erik Torenberg To that end, rack power requirements are moving from roughly 5 to 10 kilowatts to 100 to 50 kilowatts. Compute density is climbing something like 70x. Cooling is moving from air to liquid as a requirement. What are the investment opportunities as a result of this? Ben Horowitz First of all, when you get to that level of power per rack, AC power doesn't work anymore. That's a pretty wild thing. Now you're into DC power, which, by the way, also requires its own cooling and is super dangerous. It's kind of ironic because Edison promoted DC power by claiming how dangerous AC power was and demonstrating it by electrocuting animals and things. Martin Casado The horse. Yeah. But he was right about his own kind of power, which is extremely powerful—that's the good news. Starting with power, that's going to be very, very different. With cooling, yes, we're going from air cooling to liquid cooling. I think we're already at liquid cooling for any state-of-the-art data center. That's already kind of a done thing. But it gets into the fact that, given the political environment and so forth, liquid cooling isn't enough. It's got to be eco-friendly liquid cooling. DC power isn't enough. It's got to be power that contributes to the power of society, not takes away from it. So you have data centers that have been behaving badly—a small percentage, actually, probably 10%—wasting a lot of water. Not as much as pistachios or almonds and so forth, as people demonstrate on the internet, but they could be a lot more efficient with that. And then there are ones that are parasites of power and don't contribute power back. I think all that's going to end. It's going to have to end just because we've gone through a one-way door on that. That requires a level of engineering that many haven't invested in yet, so that's coming. And then, if racks are that dense, there are other things—the way the floors are designed have to support that kind of weight. That kind of thing is actually for real. And then I think that you just need a lot of everything. Also, by the way, things are really loud, so you have to build the data center with thicker walls, or you're going to disturb the peace in the neighborhood, which is not going to be acceptable. I don't think any state is going to allow that. Raghu Raghuram And so a lot of the ways people have architected and designed the buildings themselves are already completely obsolete. Once we get to Feynman, a much smaller percentage of the data centers that we have today will work. In fact, everybody talks about memory prices, but one of the fastest areas where prices are increasing is reinforced concrete, for heaven's sake. Martin Casado The other thing that happens when these data centers are sending 800 volts to the rack is that it becomes so dangerous, number 1. Secondly, we don't have enough electrical contractors who have the expertise to deal with 800 volts inside the data center, because this is high voltage. Only 2% of electrical engineers or electricians in the US have been certified on DC power, so that gives you an idea. Now Meta has a whole program to train people up, which is great. It's like a new Job Corps where they train people for free to do this job. But it's funny: AI is taking all the jobs, and AI is going to create a lot of new electricians. Ben Horowitz Yeah. I think we're doing something in this space, too. The big guys that own the big cloud data centers all are furiously experimenting with robots, right, to do the work of assembling or putting servers into the data center, et cetera. And so you'll see that increasing as a result of the evolution in AI. Martin Casado By the way, to be clear on the actual fund that we're raising, our focus is on computer science infrastructure. So anything a model runs on—that's computer science, right? Think chips, network interconnect, storage, all the way down probably to the electricity. Erik Torenberg Yeah, and say more about the robotics arm in terms of what we'll be doing versus maybe American Dynamism, or how to think about that. Martin Casado Yeah, for sure. Again, we think that any platform that AI will run on—one of the great breakthroughs that AI does is it allows computers to interact with the physical world, right? It can see, it can hear, it can talk, right? And this means new platforms, right? ### 44 Gigawatts by 2028: Why Building Faster Is So Hard The simplest way people say “edge device,” but that doesn't really mean anything, right? I mean, it could be a mobile device, it could be a CDN, it could be a laptop, but it also could be an embodied device that goes around. And so, again, as infrastructure-focused investors, we don't do heavily regulated industries or more verticalized industries, but any sort of computer science platform that's going to push AI further out, we're quite interested in. Erik Torenberg Yeah. Going back to the data centers, by 2028, new data centers are going to need something like 44 gigawatts of additional power, against maybe 25 gigawatts of expected grid additions. Ben Horowitz Hold on, hold on. We use that word “gigawatt.” Raghu Raghuram No, it's like we'll have 100 gigawatts. Erik Torenberg Martin, what's a gigawatt? Martin Casado I mean, how big is it? It's multiple football fields. I mean, it's massive. It's 50,000 people. Ben Horowitz What do you mean? What is it? Martin Casado The equivalent is like 50,000 houses. Ben Horowitz A city? Martin Casado 50,000 homes. I grew up in Flagstaff, Arizona, which is a town of 40,000 to 60,000 people, depending on the universities. We have less than a gigawatt of power consumption. I mean, this is— Erik Torenberg So you can basically light up and air-condition your entire town for a gigawatt. Raghu Raghuram Yeah. I mean, this is— Ben Horowitz We're just throwing them around. Raghu Raghuram No, but by the way, everybody talks about the gigawatt. There are very few gigawatt data centers that are actually up. We've got a long way to go, right? But then why can't utilities and hyperscalers just build faster? Oh, there are so many things there. Well, first of all, right now, you need humans to build them. There's just the regular construction, but much more than that, you need permits. You need access to power that you can plug into. So you're doing a combination: you've got to get access to power, which is a massive kind of regulatory bidding struggle. There are very limited amounts and things you can tap into in terms of natural gas, power grids, what have you. But then you also have to build your own power, and guess what? We've got shortages of transformers and turbines and everything that goes into that. So you've got to get all that stuff. This is not a software problem. It's not just like a bunch of engineers can work weekends and that type of stuff. That doesn't work anyway, but there are real bottlenecks in this, and these lead times are not that easy to compress. And look, we have the best minds in the world trying to figure out how to compress them, but it's not easy. It's not easy, and the demand is not slowing down. We're already behind. The demand is growing 10x a year right now, and the supply just can't grow that fast. Ben Horowitz By the way, it is so bad that right now, if new companies are going for GPUs, it's often in Mexico or Australia or another country, just because it is so difficult in the United States. Yeah, we're creating huge job and long-term economic opportunity in other countries by banning data centers here. I think the right answer would be to set a standard where a data center contributes back to the community: the power gets better, there's no noise, there's no water issue, and it's adding jobs. That ought to be the standard. And then everybody ought to be held to that standard. ### Why "Machine Age" Is the Right Name And by the way, there are data centers that do that now. That's not a futuristic dream or something. Their energy rates have gone down every year, and the reason is they provide their own power. They give power to the state during the day, and then at night they borrow power from the state, when the state doesn't need it, because the way power plants work is you're always generating peak capacity. Since a data center has steady capacity during day and night and a city goes way up in the day and way down at night, that's a symbiotic relationship. Erik Torenberg Zooming out, we've been batting around the name for a little bit. Why do we think Machine Age is a compelling term for what we're doing here? Martin Casado Well, listen, let me take it. The first one is, I think Ben's absolutely right: artificial intelligence was the wrong word. We shouldn't have called it that. It's machine intelligence. Erik Torenberg Say more about that. Why is that? Martin Casado Because it's not how humans think, necessarily, right? I mean, it is a cache of how humans thought; it's a collection of human thoughts. But to date, we don't know how to take an AI with no knowledge, put it out in the world, and have it reconstruct language, right? That's not what we've done. We've built something that can learn off of everything we've already learned and then use that in a productive way. And listen, AI is a general term that goes back 70 years in computer science, formally, that applies to many different things. And of course, it's got a lot of baggage, either from science fiction or from Nick Bostrom, who wrote about it, or whatever. And so the first one is just an acknowledgment: this really is machine intelligence. And then you want to emphasize the machine part of it. I mean, there's kind of this deep irony—and this is from the “software is eating the world” people—that you've really come to a place where you pour money into something and then you're limited by the actual machines below it. And so I think it is a kind of nod to how the hardware component is so significant in this wave, and we want to acknowledge that. Ben Horowitz Yeah. I think that's what's going to create the next breakthroughs in intelligence: the quality of the machines underneath. And so that's basically a reason for the name. Raghu Raghuram It's also a cool name. Erik Torenberg That sounds good. Martin Casado Futuristic. ### Won't Incumbents Like Nvidia Take Everything? Erik Torenberg Yeah. Given how much has been spent on AI infrastructure to date and how capex-intensive these businesses can be, are we past the point where new companies can break in at material levels? Why not incumbents like NVIDIA, CoreWeave, et cetera, just take the lion's share of these markets? Ben Horowitz They're all doing well. There's no question about it. But to our discussion earlier, when you need fundamentally new innovations to keep the growth continuing, or the pace of improvement continuing—whether it's tokens per second per dollar, tokens per watt, tokens per rack, or power—you take any metric. If you want to have a 10x on those metrics, you've got to have new innovation. And new innovation traditionally comes from brilliant founders thinking about solving the problem from first principles in a different way. That's what's needed here for the next jump in innovation. I mean, this is the law of markets, right? Let's assume that the existing silicon incumbents are multitrillion-dollar companies in terms of market cap, which is absolutely the case. Martin Casado Even 5% of that is a massive private company. You could say, “Well, NVIDIA could do that.” They could, but why would they if they're focused on things that are in the 90%, which is also driving the same amount of growth? You always ask these questions. We asked these questions during the cloud days: Why wouldn't Amazon do this? You asked these questions during the Microsoft days: Why wouldn't Microsoft do this? There's a very natural law of markets: Once you get to a certain scale, there's tremendous opportunity for innovation at the margins. Ben Horowitz Yeah, there's a funny quote from our partner Alex Rampell. He had this startup called TrialPay and was trying to sell its services to Meta—then Facebook. Dan Rose, who was the head of corporate development at the time, said, “Alex, that's great. It sounds like you can collect a lot of silver bricks, but I have so many gold bricks I can't even pick them all up. The last thing I'm doing is looking at a silver brick.” I think NVIDIA is in that position. Martin Casado 100%. We were talking about this as it relates to the model providers: If you're in the sweet spot of what OpenAI and Anthropic can do—one of their main areas of interest—that might be a tough place to be. But anything outside of those, maybe, 3 to 5 areas might be attractive. As markets expand, they fragment, right? It happens all the time. Remember, in the early days of Ford, in 1913, there was the River Rouge plant. Literally, it was like water, coal, and rubber trees went in, and out came cars. Ben Horowitz By the way, he bought a whole— Martin Casado Rubber plantation in the Amazon jungle, right? Ben Horowitz And there's a great book called Fordlandia. He wanted to own the complete vertical, so he created this city called Fordlandia in the Amazon jungle. It was all Americanized—bandstands, ice cream, and all this kind of stuff. It actually worked for a while, until he made people show up to things on time, and they were like, “Screw this. Get out of here.” Erik Torenberg Yeah, the use cases are multiplying. There's no way, if you're the biggest company, you can get to the biggest use cases, but there are so many use cases—and, as Martin was saying, they're all very valuable—that it's just very hard to get to them in a great way. Martin Casado Yeah, even inference used to be one simple architecture, and it no longer is. It's so complex now that it's inevitable you can optimize things in a different way. Ben Horowitz By the way, here's a very interesting thing: People often don't understand that the margins kind of fell out of the standard way of doing the technology with software. It wasn't really a technology problem. Once you got the business working, you tended to have pretty good margins, because that's just how software works—certainly when you shipped it, but even as a service. That's not necessarily the case with AI. We may actually be entering an era where optimization in the hardware is absolutely meaningful to the upside of the business in a way that we haven't seen in the past. So there's a lot of opportunity here. Let's get deeper into the types of companies we'll be investing in. Maybe we could start by illustrating the subsectors, or we could talk about a few investments that we've made. I know there are some that haven't been announced yet. Raghu, do you want to take us down? Raghu Raghuram Yeah, the subsectors, as we've been talking about, span every one of these categories. The obvious ones are computer chips, but these days it's not enough to build a chip; you need to build a full system. Therefore, what goes into the system? There's potentially memory innovation, networking innovation, power chips, and so on. Each of these categories is one where you can see public-company-style companies emerging, and those are all things we're looking into. Once you put it all together, there's a layer of software around it to automate all of these things, manage these fleets, and so on. That's another important area. These things keep building on each other, but every one of these categories is important. Erik Torenberg Talk about what's different about these kinds of companies from the usual company. I mean, one thing you can tell from the companies we announced is that their first rounds have been massive—hundreds of millions. Is it a different kind of founder, or what else is different as we think about the practice of building and investing in these kinds of businesses relative to our traditional software? Ben Horowitz Well, I think the big thing is that a lot of money goes in before they get to a product. That's just the nature of it. It's true of big models, too, but I would say that's a little more of a known path, whereas this has a little more risk and a little more money than some of the other things we've done. Raghu Raghuram A lot of the chip founders are here from the past. You know, the guys who know how to make memory, they're not young. That part is different, too, but it's kind of exciting. Martin Casado Yeah. The other thing about these founders is that they've all got to be systems founders. What I mean by that is you can't just be a researcher or a great computer scientist. You've got to be able to architect and design the chip or the system, whatever it is. Then you've got to think about how this thing is actually going to get manufactured: Who's going to be supplying this? There are a whole bunch of downstream things which, normally, if you're building software, you don't have to think about. Really, the best founders—and, of course, Jensen is the Michael Jordan of this—think about the entire ecosystem from the get-go, before they start designing the chip, because of the nature of the bottlenecks and all these things that have to come together. That's a big characteristic that's different. Raghu Raghuram There are 2 environmental factors that are important, too. The first one is that the labs are so desperate that they will engage with startups. We actually have quite a bit of signal early on because labs are inking deals with companies before they actually have hardware available. That's a big shift from 5 years ago, right? You just didn't go and sell your janky hardware thing to Google or whatever. That's a shift. ### The Founder Profile: Why Hardware Needs Experience The second one is that capital availability has loosened up a lot. I think there's general consensus that it is time to reshape this stuff, and so there's a lot of capital available in follow-on rounds. Of course, you want to be investing in areas where capital is available, so the atmospherics are also just different. Erik Torenberg Patrick Collison remarked a few years ago, “Hey, it feels like there are fewer young founders today, in the way that Zuck, in college, was building the next Facebook, or Gates was doing the same with Microsoft. And, of course, Michael Dell was building Dell. There are still some young founders building iconic companies, but it does seem, to your point, that there are more older founders building these companies—or fewer 20-year-olds.” I'm curious whether that resonates and why. Ben Horowitz Well, I think it's Raghu's point: If you're building something that has a very complicated supply chain, has to manufacture things, and is technically complicated, some experience helps. If you look at Elon or Travis Kalanick, their companies, when they were young, were software companies. It wasn't until they got a lot of experience—even those guys, the best guys, needed some experience in building companies, building technology, and so forth—that they graduated to much more complicated, or I would say elaborate, domains. There's just much more there; there are many more moving parts in these things. Look, when you're learning how to build a company, it's hard enough if you completely understand the product. If you don't completely understand the product and have to learn it while you build the company, that's such a steep learning curve for a brand-new entrepreneur. So I think what we're seeing is Michael, on the one hand, who is a very young guy, brilliant, but what he built was kind of a pure-software AI thing. Martin Casado Yeah. And on the other end, you have an Elon or a Travis, who's got enough experience. I think Michael could probably do that 10 years from now, but today that would have been hard. It's important to remember that it's been defocused by the entire industry and academia for the last 20 years. There just hasn't been the same opportunity. Martin Casado Like it’s been there, but it’s never been a growth area. The growth areas have been software, networking, things like that. I also think we have a paucity of people coming out of universities or gaining experience at large companies that have done this. There just aren’t that many. Ben Horowitz You don’t go intern and build a chip. But a lot of that’s changing now. Listen, we’re going to create a whole generation of founders who come from these new companies, who will know how to do this, and they’ll be hired much more junior. I would say, actually, one of the greatest legacies of Elon toward this is, of course, that he’s created these great companies. But the number of entrepreneurs who have come out of SpaceX and are changing the entire industrial complex may be an even greater legacy than the companies themselves. I think we’re going to see the same thing for computer science and hardware. Raghu Raghuram Yeah, as a matter of fact, all of our investments were started by 2 founders in their 20s. But if you go to one of their offices, you see the experienced people as well. So it’s an ideal combination here. Erik Torenberg Yeah. Martin Casado Yeah. Yeah. It doesn’t necessarily have to be the founder with experience, but that founder better be able to tap into that experience with them. Ben Horowitz Yeah. Yeah. Well, they have to be able to work with them, and they have to be good, and all these kinds of things. It’s complicated. Erik Torenberg Speaking of experience, this is a big new fund we’re launching, and there are no new GPs. We’re sort of collecting it because you guys have a lot of experience, and the rest of the group has experience in this field, which has been kind of latent and dormant. Speaker 1 Yeah. Well, it’s kind of funny. I think we almost had to be warned against it, almost just because our backgrounds are from hardware. I think the reason that we needed a reminder is because all of us have spent so much of our careers in software and hardware. We’re kind of drawn to that. So, listen, we’ve clearly invested in hardware over the years, right? We’re in SpaceX, we’re in Anduril—all of these are very early checks. We’re in Astranis, we’re in Waymo, so even early on, we did a number of those investments. But this is so much in our DNA, and I don’t think this necessarily needed to increase the team’s competencies; it was just a bonus. Martin Casado If this fund does what we think it will do, Ben Horowitz How do we see the world changing or looking in 5 to 10 years? Well, hopefully, America wins in the infrastructure game. We have lots of super eco-friendly, efficient data centers out there, and lots and lots of chips, an abundance of memory, and an abundance of power. That would be awesome. I think it goes back to this: We really think America is a special place, and we’re important not only to everybody here but to anybody in the world who wants to make a contribution and do something bigger than themselves. It’s the best place to come with nothing and do something profound. We’d like to keep that going, and I think that doesn’t continue if we lose our lead in technology. I think we’ll be in another era, and it’ll be another country, and maybe they have a different set of values around that.