So what's really going on in AI, right? I mean, look, the obvious stuff at the start, obviously, winning in AI, this is something we articulated in 2018. Winning AI, you're going to have this whole eat the work story. You won't just replace cloud software, you actually will replace the ultimate world of work. And that's a much bigger marketplace. No points for that in 2026. We got the points in 2018. And then the other thing is you're seeing a lot of the AI adoption being in spaces where there wasn't software before. So parts of the story are very expansive. The existing markets are getting bigger and AI is also coming to new markets like healthcare and legal, which, let's be frank, were terrible markets for software before this. The weird thing about AI is nobody has a clue what it means. And this is my favorite stupid graph. I think this originally came from the Financial Times and it's basically saying AI could mean anything from the end of scarcity, you know, acute human abundance, no more need for material goods anymore, or it'll be amazing, to on the other hand the actual tech singularity and human extinction are somewhere in between like 2% GDP growth. This, to state the obvious, is not a particularly useful freaking slide. It covers pretty much every eventuality except we're all struck by a meteor and die. So it doesn't really narrow the discussion and obviously as investors you want to narrow the discussion. I have my biases. We'll come to it in a second. So that didn't help. And the funny thing is despite this complete evidence of no one knowing, a lot of this stuff was very knowable. I mean you had in 2012 you had ImageNet proof that deep learning worked. You had in 2017 proof that transformers worked. You had in 2020 the famous scaling law paper which effectively said if you put money in in the form of compute you get better results out and the more money you put in the better results you get out. That was a real clue if you were paying attention that smart people are going to spend a lot of money and make these models much better. Obviously in 2022 to make it even easier you had ChatGPT and a little reminder, you know, sometimes things are unpredictable but amazing. That was not what the team at OpenAI expected but it exploded and suddenly we're all paying attention. And then to make it even easier in 2024, you have the Situational Awareness paper where Leopold the trader says, 'Dudes, these are the stocks you should buy to make a ton of money in the next two years. You almost can't miss because the people I know at OpenAI are going to spend a lot of money, and this is the stuff they're going to buy. You can't make it any easier than that.' And still, we all struggle. And here we are in 2026, and we're spending 600-and-something-odd billion a year in capex. So that's kind of big picture what's going on. And it was knowable.
Right now, anywhere you cut it, the AI is in invest mode, which is what you say when you're spending more money than you're taking in. Right? And zoom out a million miles. There's a lot of ways to measure this, and some ways are unfair, some ways are not quite right. But the big animal, I'm a big believer in investing. In the words of John Maynard Keynes, I'd rather be roughly right than exactly wrong. This is roughly right. We are spending in AI capex just among the hyperscalers $688 billion this year. On the other hand, and that's what it takes to make AI, think of that as all the money, I'm very simplistic, all the money going in to just make AI happen, and then what's coming out the other side. What's coming out the other side is about $110 billion in actual revenues, most of it, $89 billion coming in estimates on a GAAP basis from the two foundation models, and then round up everything else, 20 or $30 billion, however you want to cut it. And you know, we could be wrong, plus or minus $10 billion, but who the hell cares. The bottom line is we're spending half a trillion dollars more than we're taking in. This is meaningful. This is a meaningful fact. And anything you do in terms of investing has to bear that in mind. And that's the power. I mean, that's everyone buying into the scaling laws because that shows that putting money in gets good results out. And buying into the ChatGPT moment that shows people want to use this. It's as simple as that. American capitalism is saying if we can make money here, we're in. And it's going to go on for a long ass time. Right? If you look here again, simple big animal pictures, right? You look at the sum of OpenAI and Anthropic's revenues estimates. Estimates may vary. You can add in everything else plus or minus. You can see by early 2032, plus or minus a trillion dollars, real money. The stunning thing is it takes until 2031 or 2032 before the revenue from those two companies surpasses the total capex. And again, I want to acknowledge it's not a perfect measure. Some of that capex, Meta will tell you it's being used for other stuff. Yeah, right. Some of that stuff is useful for Microsoft. Yeah, maybe. But the bottom line is most of that capex is going to support the AI buildout. And the AI buildout is the foundation models to a rounding error. So it's going to take probably five or six years before we're not in quote-unquote invest mode, right? And what that means is you just got to acknowledge this fact: when things are in invest mode, when you're spending more money than you're taking in, at any point you could have a hiccup. It would not surprise me, even if the long-term story is amazing, that sometime between now and then someone wakes up and goes, 'Oh my god, what are we doing? We're spending half a trillion dollars. We're only making 200 million bucks. Maybe we should slow down.' Right? So I look at this and I go, you know, I think the journey makes sense. Who the hell knows how it shapes out? And then underneath, a number I just look at is that: is there enough money for these guys to make a trillion dollars? Well, yes, there is. There's a lot of money out there, it turns out. Um, but you have to take an appreciable slug of the total knowledge worker wage bill. I mean, we're talking about an appreciable slug of the largest pile of money out there, right? If you just looked at US knowledge workers, if you said all this money was spent in the US, you'd probably be talking about something like 15, 17% of all knowledge worker dollars, right? And that probably means, to make it real, for software developers, you're probably talking well north of 25%. In other words, for every $200,000 software developer, they're spending $50,000 on tokens. And you all know where you are right now. That's an appreciable shift. So the point here is to make these kind of numbers, a lot has to happen, and these foundation models have to eat a significant portion of what currently is knowledge worker dollars.
So possible, maybe even likely, it won't happen without bumps. So given that that's the framework, the next question is where do you invest? And again, as you'll see, I like simple pictures. This is on the, uh, whatever it is, your left. Had to think for a second there. That silly little kind of pile of blue rectangles that's incredibly ugly because these are just internal slides. Um, that's actually the kind of model that actually Jensen talked about in all his recent talks on how he thinks about the stack. So I'm going to go with the guy who's made $5 trillion as an approach. And you know, you've got energy, chips, infra, models, apps, right? And as a software investor, my mental model is the bottom three categories are all about making AI. It's about the stuff it takes to build $680 billion of capex. And as investors, when you look into that, 80% of it plus or minus is not really venture-backable stuff. Half of it is chips. Now, we're going to have an amazing venture win in Cerebras in the next few weeks here, which is great. But obviously, the vast bulk of the dollars have gone to Nvidia. At the energy level, obviously, it's all the obvious kind of energy providers. And by the way, that's a big rectangular block because it's hard to get energy, but the dollars are actually relatively small. Chips are the big spend. And then infrastructure, you know, ranging from some stuff that's pretty, you know, low-tech, buildings and stuff like that, to some parts where you are dealing with things. I mean, I lump, rightly or wrongly, things like labeling companies and your RL training gyms into that category of stuff that the foundation model companies buy to make AI, right? So there's been some interesting venture bets there. There's a little bit of software there. It's not a whole ton, but there obviously has been a huge spend. I mean, you've gone from $200 billion in capex four years ago to $600 billion in capex, which means someone is pocketing $400 billion a year, mainly Jensen. But there you go, right? There's been plenty for everyone, including the memory vendors. So that's the making-AI bet. It's not the software space, but just to acknowledge that right now it's the biggest space. And then next level up, you have the models themselves. You know, the foundation models. Obviously, you have the two dominant venture-backed survivors at this point, OpenAI and Anthropic, and then some of the others that fell away. And we'll talk in a few minutes about do you need to be, what are the opportunities in foundation models from here. But that's kind of, they are, you know, the graphic doesn't do them justice, someone else's graphic, but the truth is they are the apex predator of this distribution. They are the guys who have taken the value, the delta value, and probably will continue to do so under those forecasts. And then on the right, you have the using AI, in other words. This is all the rest of us. This is how soft... this is the question of how software uses AI to deliver value to corporate customers. And as I'll say in a second, the million-dollar question is how that $1 trillion gets spent between the model companies and then the using-AI software companies, many of whom are here today trying to figure out what they should be doing. And that's kind of the big picture framing. So think of it: the $600 billion below the line in making AI, and the $110 billion that has to become a trillion dollars above the line to be divided between the foundation model companies and everyone else. And that's kind of the big picture. And this is kind of some, I hate the open questions, hot takes, whatever, I don't like. But if I'm really honest, these are the questions. I mean, one of the reasons I had this deck fairly ready is because internally, you know, as we do a quarterly review on what's going on, and as one of my partners said, hm, if people are seriously talking about is software dead and we are software investors, we probably should pay a little attention here, people, and come up with our perspective on that on behalf of our LPs, right? So these are the questions you're wrestling with: are the foundation model companies going to take it all? Are the old pre-GPT software companies, Ed, do you have to become full stack? Is the enterprise ROI there? You know, where are the weakest points? Where are the strongest points? All these are really a list of questions that we were running through. And to the extent I don't cover any of these answers, you're obviously more than welcome to ask, right?
And if you step back, this is the biggest question of them all. How does enterprise want to consume $1 trillion worth of AI over the next five or six years? Right? That's the money question, right? Do they want to buy it all direct from Claude and Anthropic, in which case you should all go home? I should go home. We're going to need a very, very small conference room in five years, and it's going to be a pretty sad moment. Or is there an opportunity for a whole bunch of companies to probably buy AI from the foundation models or from the open-source vendors, add value on top of that, be closer to a specific set of customers, build value for those customers and deliver an economically differentiated product and build businesses like that? Right? That is the question, right? And it's one version of... it's the modern version of is software dead, because when people say is software dead, what they don't mean is software dead because OpenAI and Anthropic are software and they are most definitely not dead. Sometimes they mean is old-school, what I call plain-vanilla SaaS dead, and we'll come to that in a second. And then sometimes they mean this question, which is how does economic value get allocated between the foundation models and modern software companies, right? In other words, companies founded post-2022. I put Sierra there because they just raised at $15 billion. You know, obviously you could have any, Harvey, Lora, many of ours, many I'm sure of you in the audience, right? And you know, worth pointing out that no one knows is the answer. And most VCs are betting both ways. They're chucking a bunch of money in at a modest $300 or $600 billion pre at Anthropic, and at the same time they're doing a whole bunch of investments that only make sense if in fact there is room separate from that. Right? So you have to come up with a theory of the case here as to what kind of technology, what does it take to survive in that marketplace, and what kind of companies, and what does it, what technically does it mean, what kind of business models does it take. And this is what we've literally, as I said, been wrestling with right now in the last few weeks just talking about, because obviously we've been investing post-GPT in a bunch of these companies. But when you step back and say, do you have a kind of clear sense in your head of what works and doesn't work, right? And I, I edited some. They told me to cut out slides. So we have about six or eight slides. You know, we, we meet every Friday and talk about kind of markets. The last couple of Fridays it was, you know, you have this whole harness discussion. What's one vector we, we kind of come at it from? Sometimes we come at it from a tech perspective like this, and then in the next few slides we come at it from a business model perspective. And you know, I think partly it depends on each individual partner thinks a different way, right? But this is some of our infra team talking about, you know, what is a harness? Simply put, what is the software stack on top of a raw model that turns it into a dependable software system, right? Because I don't know if you saw that software convergence article. It is true, many software apps are starting to look the same. It's often a chat interface, a bunch of stuff, interaction with a model, and then back it comes, right? But then the harness is the stuff that makes it all happen. And you know, list of things here that you got to do. You got to decide what context the model sees. In other words, what information you put in, allow the actions to take place, you know, select the outputs, log what happened, and then escalate and manage models depending on what you need. And you can squint one way and say some of this is trivial and Claude and OpenAI will have very basic level harnesses that can do all this. And you can squint another way and say in fact each of these things in and of themselves is a separate product within an overall harness that someone building an app like a Sierra might use. And opinions even within our group differ pretty markedly. Right? I don't love the word harness. I think it's a weenie word. I actually think a better way of thinking it is like we lived for 15 years in the LAMP stack, which was, you know, how you build a SaaS application. This or something like it is going to be how you build every software application in the age of LLM. And just because every SaaS app had the same LAMP stack doesn't mean they were all the same app. Obviously they weren't. They were vastly different. And I think in the same way, just because many of these apps will be roughly constructed the same way with the same kind of framework, because you're trying to do the same thing from a technological perspective, access and reason on a model and then distill it into something useful for a business user, they will all sound the same, but how they're instantiated, my gut is, will vary very differently. If you're a legal firm, if you're a customer interaction firm, if you're an app. I mean, the odd thing is if you're a software coding app, because even software coding itself is a, is now an app, right? So my belief is that there are, in fact, that even though they all look and sound the same and in terms of that the descriptives are the same, I, I think there will be very much differences there. It's a subject of discussion, and I think you, I'm wrestling, and this is literally real time. We have another week talking about it. You know, what's your pro... what's your rule of thumb for simply put stuff that Claude and OpenAI can roll over versus stuff where you really have defensibility, and what is, what are your technical moats, and then on the next slide, what are your business moats? And that's really, yeah, the, the right-now question, right? And I'll move on to the next slide just to talk. This is an internal slide. Um, you know, you can imagine lots of different... I mean, I, I don't love it because there's both business and this is actually, it's an internal slide just as our things, which I don't love. I like to kind of re-circulate to a wider group, but there's a lot of different business moats and some technical moats I want to talk about here. On the business moat side, there's obvious stuff like you can think of businesses that are sufficiently different from what Claude or OpenAI is doing that you can say, yeah, yeah, I get it, they're not going to be worried by these guys, they're not going to be competed away. And then you can think of other markets where it's much tighter. And if I'd had time I would have reordered the slide a little bit, but let's take the ones that are easily... If you're doing kind of a combination of software and sensors, you've got something that's obviously manifestly different. You know, you've got maybe visual sensors, you've got touch sensors, you got something out there that says OpenAI ain't going to start shipping sensors, right? That's just not a thing. It's just not worth their while. So a lot of these deeply vertical, real-world-based apps that are a combination of sensors, then reasoning, vision models to allow business outcomes, I think they're eminently defensible. I don't think it's an issue. Another category is marketplaces. We just invested in Parform, which helps recruiters find the best candidates. And from a business model perspective, if you're creating an incentive for all sides to want to be on your platform and use your product, where you have those kind of effects, then you will find that that's again not something that a single person can take away with a Claude code app. So do you have a situation where everyone on your platform, whether you know the classic network effect is the more users you have on the platform, the more value it is to everyone else. Marketplaces are typically like that, right? And I'm going to skip FDEs to capture context and I'm going to skip data flywheel because they're the closest to the Claude side. And then network effects and proprietary data. If you have some proprietary, visibly non-public LLM data that you build your business on, and many of our companies have that, you're again eminently defensible. You don't have to talk about why because they just, because the model simply can't do what you do. And then obviously the one at the end, full-stack businesses. That's where instead of you as a software company selling software to a business to pursue whatever they want to do, you actually become the business yourself. You integrate forward. And Range, a company of ours, they're actually doing wealth management using LLMs. But you know, I can tell you OpenAI is not going to get into the business of wealth management. They're clearly, by the way, going to get into the business of hiring wealth managers for literally every one of their engineers. But that's a different point, right? This is just something they're not going to do. The two I skipped are the trickiest ones because they're, I think of as where you're still dealing with core software companies where you don't have something weird that makes you different, but you're making a much more subtle argument, which is that, you know, either a data flywheel from usage, this whole context graph and getting the, where you can build up value over time. You start probably with a simple app and then as you learn how your users do it, you develop a very much a specialist and better and differentiated app. Or in the case of this FDE concept, where to make the software work you need people on the ground, and to some extent that could be differentiated, though obviously you're seeing a lot of even the, I mean, you're seeing the foundation model companies in the one sense admit they need to do that by partnering with PE, and on the other hand making it clear that they will do that, right? So this I would say is a work-in-progress slide. In fact, I see things now and I'm like, didn't want to do it that way, right? But you get the point. There are purely software architectural things that are the hardest to distinguish, but I think are really there. And then there's a bunch of really obvious business model or physical instantiation things that together to me say there are plenty of software opportunities that aren't just, oh my god, the model companies are going to roll over you. Right now they're starting with a trillion dollars, they'll be just fine, but there's probably going to be room for everyone.
And then the other version of is software dead that you have is what happened to every software company pre-2022, right? Did AI result in a mass extinction event? Because this is what our investors ask, because you've got companies that have been, you know, around a long time, or what happened. And this is, you know, a rough take, which is there was some, we did a lot of pre-AI, pre-GPT AI investing, and about 10% of our portfolio was literally DOA when ChatGPT comes out because they'd solved an AI problem that was just so much easier to solve with ChatGPT that just wasn't fixable. You literally go, 'Oh, we made the thing for $30 million and now they make the thing for $5 per million tokens. You're done.' Right? So you have a casualty rate, right? And then the remaining 90%, what you see is roughly thirds, roughly thirds, where you kind of go, there's a third of it that's very GenAI insulated. They do something that's sufficiently different, often with data effects, often using different kinds of AI like predictive AI, that you go, this business won't be impacted by this, it's just orthogonal to it, you don't need to talk about it. There's some businesses actually where it's additive because you have this core product today and then you can add other stuff on top of it. I think one of our companies, DroneDeploy, does a really nice drone management and flight coordination product. Now with AI, you can enhance the value of that product by doing, you know, safety reviews. You can do progress reviews. AI on top of your product increases the value and there's no way that AI is going to replace your value. So that feels good. And then there's about a third where you're absolutely threatened, where you typically what we call plain vanilla SaaS. You automate a workflow and then you either got to move and survive or you're going to be dead, right? You're either going to be the quicker or the dead, as they say in a crisis, right? And there's no doubt that some of that, and you could argue those percentages might vary. We might have been grading ourselves on a bit of a, we might have been squinting a little optimistically on the last category, but that's roughly how I think about it. And you know, that's going to be the dynamic for all of you in terms of your pre-22, you know, you got to figure out where you are in that world of how your companies are being impacted and where you're facing. So that was the other version of is software dead. It's the is the new software dead killed by foundation models and is the old software models dead, and yes, some are, some aren't. The other question that's obvious is do you need your own model? Right? I mean, if you think back, and this is a list that DD, a guy at Menlo who's a very smart investor, did a long list of 65 new labs, and we took it and we sorted it into buckets because your venture people are simple people. You just want to put stuff in buckets like the other stuff. And there's around 10 or 15 different categories of non-pure foundation model that may be off because one way to compete with the foundation models is to have your own model. And the obvious category is, you know, LLM replacement companies. You got Thinking Machines, Safe Superintelligence, all the people just going directly at next generation. Then it's worth pointing out that voice, video, and image all had foundation model options that weren't kind of pure the LLM. Anthropic, which doesn't do any multimodal stuff at least until recently, OpenAI did but deemphasized it. So for example, ElevenLabs, they have a speech model and they've built a wonderful business on top of that model. So there are other AI models that aren't as compute and capital intensive as the core foundation model where you get that model rolling and then you can build a business on top. And there's a bunch of other categories here. There's kind of next-gen coding specialist models. There's chem, you know, models modeling chemistry and physical science properties. There's a bunch of different, you know, we think of this in terms of what bets can you take? There's a bunch of different bets you can take on different model types that we've looked at and tried to figure out. Do you need to do that? You know, going back to the big picture here, you know, if you step back and think, you know, what's going on with this huge change in venture investing, you know, what do you need to do, right? I mean, this is us literally trying to figure out. There's no doubt the game has changed. There's no doubt that some bets that I've been in this business 30 years, I find myself thinking again about things I last thought about 30 years ago like semiconductors, FPGAs. There's that whole trend of these new deep tech products that were, you know, typically not made in San Francisco, were made down in Milpitas and, you know, Santa Clara, right? Those kind of bets. There's the foundation model bets themselves, which are an order of magnitude, actually two orders of magnitude more capital intense than anything venture has done before. But then I still think there are going to be over the next 10, 15 years, just as Microsoft dominated the client-server and PC age and was the operating system, but there were hundreds of client-server apps companies that achieved successful and great outcomes in the public markets. Just as AWS dominated the cloud compute space and there was a whole bunch of great outcomes in the public markets of SaaS companies on top of that. I think the same thing's going to happen here. There's no doubt that a two-way or a three-way dominance of the foundation model category will be between OpenAI and Anthropic and Gemini with maybe some other players. They will dominate that. They probably will have, you know, oligopolist profits, but just as the last two times, I'm pretty sure that there will be a whole bunch of software opportunities on top of that. And because of that, I think software isn't dead. It's gotten harder, there's no doubt, and it's tricky, but it isn't dead. And that's why I'm here to find great deals. So with that, I will stop talking. And I would love to get questions because please don't leave me hanging here for 17 minutes and 32 seconds. I love it. Good.
