Marc Andreessen's 2026 Outlook: AI Timelines, US vs. China, and The Price of AI

This new wave of AI companies is growing revenue like just like actual customer revenue, actual demand translated through to dollars showing up in bank accounts at like an absolutely unprecedented takeoff rate. We're seeing companies grow much faster. I'm very skeptical that the form and shape of the products that people are using today is what they're going to be using in 5 or 10 years. I think things are going to get much more sophisticated from here. And so I think we probably have a long way to go. These are trillion dollar questions, not answers. But once somebody proves that it's capable, it seems to not be that hard for other people to be able to catch up, even people with far less resources. When a company is confronted with fundamentally open strategic or economic questions, it's often a big problem. Companies like need to answer these questions and if they get the answers wrong, they're really in trouble. Venture, we can bet on multiple strategies at the same time. We are aggressively investing behind every strategy that we've identified that we think has a plausible chance of working. If you want to understand people, there's basically two ways to understand what people are doing and thinking. One is to ask them and then the other is to watch them. And what you often see in many areas of human activity, including politics and many different aspects of society, the answers that you get when you ask people are very different than the answers that you get when you watch them. If you run a survey or a poll of what, for example, American voters think about AI, it's just like they're all in a total panic. It's like, 'Oh my god, this is terrible. This is awful. It's going to kill all the jobs. It's going to ruin everything.' If you watch the revealed preferences, they're all using AI.

First of all, I would say this is the biggest technological revolution of my life. And hopefully I'll see more like this in the next whatever 30 years, but I mean this is the big one. And just in terms of order of magnitude, like this is clearly bigger than the internet. The comps on this are things like the microprocessor and the steam engine and electricity. So this is a really big one. The wheel. The reason this is so big, maybe obvious to folks at this point, but I'll just go through it quickly. If you trace all the way back to the 1930s, there's a great book called Rise of the Machines that kind of goes through this. There was actually a debate among the people who actually invented the computer. It was this debate between whether computer they kind of understood the theory of computation before they actually built the things. And they had this big debate over whether the computer should be basically built in the image of what at the time were called adding machines or calculating machines where you think of essentially cash registers. IBM is actually the successor company to the National Cash Register Company of America. And that was of course the path that the industry took which was building these kind of hyper literal mathematical machines that could execute mathematical operations billions of times per second but of course had no ability to deal with human beings the way humans like to be dealt with and so couldn't understand human speech, human language, and so forth. And that's the computer industry that got built over the last 80 years and that's the computer industry that's built all the wealth and financial returns of the computer industry over the last 80 years across all the generations of computers from mainframes through to smartphones. But they knew at the time, they knew in the 30s actually they understood the basic structure of the human brain. They understood they had a theory of human cognition and actually they had the theory of neural networks. And so they had this theory that the first neural network academic paper was published in 1943 which was over 80 years ago which is extremely amazing. There's an interview you can read or watch on YouTube with these two authors McCulloch and Pitts and you can watch an interview I think with McCulloch on YouTube from like 1946 or something. He was like on TV in the ancient past and it's literally an amazing interview because it's him in his beach house and for some reason he's not wearing a shirt and he's talking about this future in which computers are going to be built on the model of a human brain through neural networks. And that was the path not taken. And basically what happened was the computer industry got built in the image of the adding machine. But the neural network as an idea continued to be explored in academia and advanced research by a rump movement that was originally called cybernetics and then became known as artificial intelligence basically for the last 80 years. And essentially it didn't work. It was basically decade after decade after decade of excessive optimism followed by disappointment. When I was in college in the 80s, there had been a famous AI boom bust cycle in the 80s in venture and in Silicon Valley. It was tiny by modern standards, but at the time was a big deal. And by the time I got to college in '89 in computer science departments, AI was kind of a backwater field and everybody kind of assumed that it was never going to happen. But the scientists kept working on it to their credit and they built up this enormous reservoir of concepts and ideas and then basically we all saw what happened with the ChatGPT moment. All of a sudden it crystallized. It's like oh my god, it turns out it works. And so that's the moment we're in now. And really significantly that was less than three years ago. That was the Christmas of 22. So we're sort of three years in to basically what is effectively an 80-year revolution of actually being able to deliver on all the promise that the people on the alternate path, the human cognition model path, saw from the very beginning. And the great news with this technology is it's already ultra democratized. The best AI in the world is available at ChatGPT or Grok or Gemini or these other products that you can just use and you can see how they work. Same thing for video, you can see with Sora and VO, state-of-the-art. You can see with music, Suno and Udio. And so we're basically seeing that happen and now Silicon Valley is responding with this incredible rush of enthusiasm. And really critically this gets to the magic of Silicon Valley, which is Silicon Valley long since has ceased to be a place where people make silicon. That not long ago moved out of California and ultimately out of the US, although we're trying to bring it back now. But the great virtue of Silicon Valley over the last 80 years of its existence is its ability to recycle talent from previous waves of technology and new waves of technology and then inspire an entire new generation of talent to come join the project. And so Silicon Valley has this recurring pattern of being able to reallocate capital and talent and build enthusiasm and build critical mass and build funding support and build human capital and build enthusiasm for each new wave of technology. So that's what's happening with AI. I think probably the biggest thing I could say is I'm surprised essentially on a daily basis of what I'm seeing. And we're in the fortunate position to get to see it from two angles. One is we track the underlying science and research work very carefully. And so I would say every day I see a new AI research paper that just completely floors me of some new capability or some new discovery or some new development that I would have never anticipated. I'm just like wow, I can't believe this is happening. And then on the other side we see the flow of all the new products and all the new startups. And I would say we're routinely seeing things and again have my jaw on the floor. And so it feels like we've unlocked this giant vista. I do think it's going to come in fits and starts. These things are messy processes. This is an industry that routinely gets out over its skis and overpromises. And so there will certainly be points where it's like, wow, this isn't working as well as people thought, or wow, this turns out to be too expensive and the economics don't work. But against that, I would just say the capabilities are truly magical. And by the way, I think that's the experience that consumers are having when they use it. And I think that's the experience that businesses are having for the most part when they're working on their pilots and looking at adoption. And then it translates to the underlying numbers. We're just seeing this new wave of AI companies growing revenue like actual customer revenue, actual demand translated through to dollars showing up in bank accounts at an absolutely unprecedented takeoff rate. We're seeing companies grow much faster. The key leading AI companies and the companies that have real breakthroughs and have very compelling products are growing revenues faster than any way I've ever seen before. And so from all that it feels like it has to be early. It's hard to imagine that we've topped out in any way. It feels like everything is still developing. Quite frankly, it feels like the products are still super early. I'm very skeptical that the form and shape of the products that people are using today is what they're going to be using in five or 10 years. I think things are going to get much more sophisticated from here. And so I think we probably have a long way to go.

Yeah. So just start with core business models. You're right. There's basically this industry has two core business models: consumer business model and the enterprise or infrastructure business model. On the consumer side, we live in a very interesting world now where the internet exists and is fully deployed. I'll give you an example. Sometimes people ask us, 'Is AI like the internet revolution?' It's like, well, a little bit, but the thing with the internet was we had to build the internet. We had to actually build the network and ultimately it involved enormous amount of fiber in the ground and enormous numbers of mobile cell towers and shipments of smartphones and tablets and laptops in order to get people on the internet. There was this incredible physical lift to do that. And by the way, people forget how long that took. The internet itself is an invention of the 1960s, 1970s. The consumer internet was a new phenomenon in the early '90s. But we didn't really get broadband to the home until the 2000s. That really didn't start rolling out until after the dot-com crash. And then we didn't get mobile broadband until like 2010. The original iPhone dropped in 2007. It didn't have broadband. It was on a narrowband 2G network. It did not have high-speed data. So it wasn't really until about 15 years ago that we even had mobile broadband. So the internet was this massive lift, but the internet got built and smartphones proliferated. So the point is now you have 5 billion people on planet earth that are on some version of mobile broadband internet, and smartphones all over the world are selling for as little as like 10 bucks. And you have amazing projects like Jio in India that are bringing the remaining population of planet earth that hasn't been online until now online. So we're talking five billion, six billion people. And the reason I go through that is the consumer AI products could basically deploy to all of those people as quickly as they want to adopt. So the internet is the carrier wave for AI to be able to proliferate at light speed into the broad base of the global population. That's a potential rate of proliferation of a new technology that's far faster than has ever been possible before. You couldn't download electricity. You couldn't download indoor plumbing. You couldn't download television. But you can download AI. And this is what we're seeing, which is the AI consumer killer applications are growing at an incredible rate. And then they're monetizing really well. Generally speaking, the monetization is very good, including at higher price points. One of the things I like about watching the AI wave is the AI companies are more creative on pricing than the SaaS companies and the consumer internet companies were. It's becoming routine to have $200 or $300 per month tiers for consumer AI, which I think is very positive because a lot of companies cap their opportunity by capping their pricing too low. I think the AI companies are more willing to push that, which is good. So I think that's reason for considerable rational optimism for the scope of consumer revenues we're going to be talking about here. And then on the enterprise side, the question is basically just what is intelligence worth? If you have the ability to inject more intelligence into your business and do even the most prosaic things like raise your customer service scores, increase upsells, reduce churn, or run marketing campaigns more effectively, all of which AI is directly relevant to, these are direct business payoffs that people are seeing already. And then if you have the opportunity to infuse AI into new products and all of a sudden your car talks to you and everything in the world lights up and starts to get really smart, what's that worth? You observe it and you're like, wow, the leading AI infrastructure companies are growing revenues incredibly quickly. The pull is really tremendous. So again, it feels like this incredible product-market fit. And the core business model is basically tokens by the drink. And by the way, this is the other fun thing: if you look at what's happening with the price of AI, the price of AI is falling much faster than Moore's Law. All of the inputs into AI on a per-unit basis, the costs are collapsing. And as a consequence, there's this hyper-deflation of per-unit cost, and that is driving a more than corresponding level of demand growth with elasticity. So even there, it feels like we're just at the very beginning of figuring out how expensive or cheap this stuff is getting. There's no question tokens by the drink are going to get a lot cheaper from here. That's just going to drive enormous demand. And then everything in the cost structure is going to get optimized. When people talk about the chips or the unit input costs for building AI, the laws of blind demand are going to kick in. In any market that has commodity-like characteristics, the number one cause of a glut is a shortage, and the number one cause of a shortage is a glut. To the extent you have a shortage of GPUs or whatever, if you look at the history of humanity building things in response to demand, if there's a shortage of something that can be physically replicated, it does get replicated. So there's going to be an enormous build-out of all of that. There are hundreds of billions or trillions of dollars going into the ground in all these things. And so the per-unit cost of the AI companies is going to drop like a rock over the course of the next decade. The economic questions are very real, and there are microeconomic questions around all these businesses, but the macro forces here are very strong. Given the underlying value of this technology to both consumers and enterprise users, and given the incredibly aggressive discovery of all the ways people can use this in their lives and businesses, it's really hard for me to see how it both doesn't grow a lot and generate enormous revenue.

Yeah, that's right. And that's one really important question and observation. And by the way, that also gets to this other question where there's different theories on it, which is basically big models versus small models. A lot of the data center build is oriented around hosting training and serving the big models for all the obvious reasons. But there's also the small model revolution happening at the same time. If you track the capability of the leading edge models over time, what you find is after 6 or 12 months there's a small model that's just as capable. So there's this chase function happening where the capabilities of the big models are basically being shrunk down and provided at smaller size and therefore smaller cost quite quickly. I'll give you the most recent example that just hit over the last two weeks. There's this Chinese company that produces the model called Kimi, which is spelled K-I-M-I, which is one of the leading open source models out of China. The new version of Kimi is a reasoning model that is at least according to the benchmark so far is basically a replication of the reasoning capabilities of GPT-5. And these new models of GPT-5 were a big advance over GPT-4, and of course GPT-5 costs a tremendous amount of money to develop and to serve. And all of a sudden, here we are 6 months later and you have an open source model called Kimi that is shrunk down to be able to run on either one MacBook or two MacBooks. So if you're a business and you want to have a reasoning model that's GPT-5 capable but you're not going to pay the GPT-5 cost or you want to run it locally, you can do that. And that's just another breakthrough. It's another Tuesday, another huge advance. And then of course, what is OpenAI going to do? Obviously they're going to go to GPT-6. So there's this laddering happening where the entire industry is moving forward. The big models are getting more capable. The small models are chasing them. And the small models provide a completely different way to deploy at very low price points. I think we'll see what happens. There are some very smart people in the industry who think that ultimately everything only runs in the big models because obviously the big models are always going to be the smartest, so you're always going to want the most intelligent thing. The counterargument is there's a huge number of tasks that take place in the economy and in the world that don't require Einstein. Where a 120 IQ person is great. You don't need a 160 IQ PhD in string theory. You just need somebody who's competent and capable. I tend to think the AI industry is going to be structured a lot like the computer industry ended up getting structured, which is you're going to have a small handful of basically the equivalent of supercomputers, which are these giant god models running in these giant data centers. And then my working assumption is you have this cascade down of smaller models all the way to very small models that run on embedded systems, run on individual chips inside every physical item in the world. The smartest models will always be at the top, but the volume of models will actually be the smaller models that proliferate out. That's what happened with microchips, with computers, with operating systems, and with a lot of everything else we built in software. So I tend to think that's what will happen. Just quickly on the chip side, if you look at the entire history of the chip industry, shortages become gluts. Anytime there's a giant profit pool in a new chip category, somebody has a lead for a while and gets the profits appropriate to robust market share, but in time that draws competition. That's happening right now. Nvidia is an absolutely fantastic company, fully deserves the position they're in, fully deserves the profits they're generating, but they're now so valuable, generating so many profits that it's the bat signal of all time to the rest of the chip industry to figure out how to advance the state-of-the-art AI chips. That's already happening. You've got other major companies like AMD coming at them, and really significantly, you've got the hyperscalers building their own chips. A bunch of the big tech companies are building their own chips. And of course the Chinese are building their own chips as well. So it's pretty likely in 5 years that AI chips will be cheap and plentiful at least in comparison to the situation today, which I think will be extremely positive for the economics of the kinds of companies we invest in.

Yeah. Well, that's the other thing. You have these disruptive startups. And actually, we were not really big investors in chips because it's kind of a big company thing, but it's a little bit of historical happenstance that AI is running on GPUs. GPU stands for graphical processing unit. For people who haven't tracked this, there were basically two kinds of chips that made the personal computer happen: the CPU, central processing unit, which classically was the Intel x86 chip, the brain of the computer, and then the GPU, which was the second chip in every PC that does all the graphics, 3D graphics for gaming or CAD/CAM or Photoshop or anything that involves lots of visuals. The canonical architecture for a personal computer was a CPU and a GPU. Same thing for smartphones. Over time, these have merged, so a lot of CPUs now have GPU capability built in, and a lot of GPUs now have CPU capability built in. So it's gotten fuzzy. But the fact that that was the classic breakdown meant that while Intel had a monopoly for a long time on CPUs, there was this other market of GPUs which Nvidia basically fought the GPU wars for 30 years and came out the winner as the best company in the space. But it was a hyper-competitive market for graphics processors. It was actually not that high margin and not that big. And then it turned out that there were two other forms of computation that were incredibly valuable that happened to be massively parallel in how they operate, which happened to be very good fits for the GPU architecture. Those two highly lucrative additional applications were cryptocurrency starting about 15 years ago and then AI starting about four years ago. Nvidia very cleverly set itself up with an architecture that works very well for this, but it's also a twist of fate that if AI is the killer app, it turns out that the GPU architecture is the best legacy architecture devoted to it. If you were designing AI chips from scratch today, you wouldn't build a full GPU. You would build dedicated AI chips that were much more specifically adapted to AI and would be much more economically efficient. And to your point, there are startups that are actually building entirely new kinds of chips oriented specifically for AI. We'll have to see what happens. It's hard to build a new chip company from scratch. It's possible that one or more of those startups makes it on their own, and some of them are doing very well. It's also possible that they get bought by big companies that have the ability to scale them. So we'll see exactly how that unfolds. And of course, the Koreans are going to play here for sure. The Japanese are going to play. And the Chinese in a major way as well. They have their own native chip ecosystem that they're building up. So there are going to be many choices of AI chips in the future. It's going to be a giant battle that we observe very carefully and make sure our companies are able to take full advantage of.

Yeah. So a couple things. One is you want to start these discussions by just saying there's vigorous debate in the US and around the world about how much we are in a new cold war with China and exactly how hostile we should view them. It's very tempting, and I think it's a very good case made that we're in a new cold war that in a lot of ways is like the US versus USSR in the 20th century. The counter argument would be it is more complicated than that because the US and the USSR were never really intertwined from a trade standpoint. A big part of that quite frankly was the USSR never really made anything that anybody else needed other than weapons. The USSR's primary exports were literally wheat and oil. Whereas of course China exports a tremendous number of physical things, including a huge part of the entire supply chain of parts that go into everything that American manufacturers make. By the time an American company brings a toy to market, or a car, or a computer, or a smartphone, it's got a lot of componentry in it that was made in China. So there is a much tighter interlinkage between the American and Chinese economies than between the American and Soviet economies. Adam Smith might say that's good news for peace, that both countries need each other. The other part of that argument is that the Chinese governance model is based on high employment. If China ended up with 25 or 50% unemployment, that would cause civil unrest, which is the one thing the CCP doesn't want. So the corresponding part of the trade pressure is China needs the American export market. The American consumer is a third of the global economy, a third of global consumer demand. So China needs the US export market or it has high unemployment and unrest. So we have this complicated intertwined relationship.

The mood in DC basically for the last 10 years on a bipartisan basis has been that we need to take China more seriously as a geopolitical foe. Under that school of thought, there's the military dimension, the risk of war in the South China Sea or around Taiwan, which has everybody in Washington on high alert. There's also the economic question around de-industrialization of the US and dependence on China. And then there's the AI question, which is both economic and geopolitical. AI is essentially only being built in the US and China. The rest of the world either can't build it or doesn't want to. So it's US versus China. AI is going to proliferate all over the world, and the question is whether it will be American AI or Chinese AI. Across party lines in DC, that's how they look at it. The Chinese are in the game for sure. DeepSeek was the starting gun in the software race. Now there are four main Chinese AI companies: DeepSeek, which is from a hedge fund; Qwen from Alibaba; Kimi from a startup called Moonshot; and Tencent and ByteDance are also doing a lot of work in AI. So there are three to six primary AI companies and a tremendous number of startups. They're in the race on software, and they're working to catch up on chips. The common understanding is that the reason we haven't seen the new version of DeepSeek yet is that the Chinese government instructed them to build it only on Chinese chips to motivate the chip ecosystem. The main chip company there is Huawei. Then there's everything to follow in robotic form. There's a global technological, economic, and robotics competition kicking off. China starts out ahead on robotics because the entire supply chain of electromechanical things moved from the US to China 30 years ago and never came back. That's the DC lens on it. The big supernova moment this year was the DeepSeek release. It was surprising on a number of fronts: how good it was, that it was released as open source from China, and that it came from a hedge fund. It was a surprise even to the Chinese government. This is encouraging because it means maybe smart kids can just build this stuff. The success of DeepSeek kicked off a trend in China of releasing open source models. The cynics in DC would say they're dumping, trying to commoditize the industry. There's probably something to that, but I think that's too cynical. They're really in the race. We've talked about the policy fights in DC over the last two years. There was a big push to restrict or ban a lot of AI. It's easy for a country that is the only game in town to have those conversations, but it's different if you're in a foot race with China. The policy landscape in DC has improved dramatically as a consequence of the awareness that this is a two-horse race, not a one-horse race.

It's a little bit complicated. Two years ago, I was very worried about ruinous federal legislation on AI. The good news is that the risk of that is very low now. There's very little mood in DC on either side of the aisle to do anything that would prevent us from beating China. On the federal side, things are much better. That has translated attention to the states. Under our system of federalism, states get to pass their own laws. A lot of well-meaning people are trying to figure out what to do at the state level, and there's opportunism because AI is a hot topic. We're tracking on the order of 200 bills across the 50 states, not just blue states but also red states. I've spent a lot of time complaining about what Democratic politicians are threatening to do, but there are also quite a few local Republican officials with misinformed views trying to put out bad bills. It's a little weird that this is happening because the federal government does have regulation of interstate commerce, and AI is obviously national in scope. The federal government should be the regulator, not the states. There was an attempt to add a moratorium on state-level AI regulation as part of the negotiation for the 'one big beautiful bill,' but that deal blew up at the last minute. The moratorium was probably too much of a stretch. We're having very active discussions in DC about the next turn on that. The administration is very supportive of the federal government being in charge. Most Congress people on both sides of the aisle get this. We just have to figure out a way to land this. Some of the state-level bills are wild. Colorado passed a very draconian regulation bill last year against furious objections from the local startup ecosystem, and they're now trying to reverse it. The really draconian one we fought hard was in California, SB 1047, which was modeled after the EU AI Act. The EU passed the AI Act two years ago, and it has killed AI development in Europe to a large extent. Even big American companies like Apple and Meta are not launching leading-edge AI capabilities in Europe. It's a classic European thing where they think if they can't be the leaders in innovation, they can be the leaders in regulation. They pass this ruinous self-harm, and then a few years later they realize what they've done. The darkest people I know about Europe are the European entrepreneurs who moved to the US. They're furious about what's happening. It's so bad that there's a process at the EU to try to unwind the GDPR. Mario Draghi's report on European competitiveness outlined how Europe was holding itself back through overregulation. In the middle of all that, California inexplicably decided to copycat the EU AI Act. It would have completely killed AI development in California. Our governor vetoed it at the last minute. It would have assigned downstream liability to open source developers. If you develop and release an AI model, and five years later it gets built into a nuclear power plant and there's a meltdown, the liability would be assigned back to that open source developer. This would completely kill open source, startups doing open source, and academic research. That's the level of playing with fire that state-level politicians have become enamored with. The good news is the feds understand this. I suspect this will get resolved because it doesn't make sense to let the states operate suicidally. We call this our little tech agenda. We're extremely focused on freedom for starters to innovate. We operate in a completely bipartisan fashion. The other question we get is why us. There's a collective action problem. Every venture firm and tech company should be weighing in, but most simply don't. Ben and I concluded that the stakes were too high. If we're going to be the industry leader, we have to take responsibility for our own destiny. That's the cost of doing business for being the leader in the field right now.

That is a fantastic question. It's one of the trillion-dollar questions. Usage-based pricing is actually fairly amazing from a startup and venture standpoint. The big tech companies with incredible R&D capabilities are building these big AI models. They were already in the cloud war, so they proliferated their magic new technology through their cloud business. This has made the most magical new technology available to everybody on the planet to just click and use for relatively small amounts of money on a usage basis. Usage is great for startups because they can start easily with no fixed cost. They can tap into OpenAI, Anthropic, Google, or Microsoft for intelligence tokens by the drink. It's a marvelous thing. That model is working, and those companies are happy and growing fast. But that doesn't mean the optimal pricing model for all applications should be tokens by the drink. We spend a lot of time with our companies on pricing. A core principle is you don't want to price by cost if you can avoid it; you want to price by value. Some AI startups are pricing by the drink, but many are exploring other models, like pricing as a percentage of the value of what would otherwise be a human job, or by marginal productivity. We see a lot of experimentation on pricing models, which is super healthy. High prices are often good for the customer because they allow the vendor to make the product better faster. I'm encouraged by the degree to which AI entrepreneurs are willing to run these experiments.

I think this is still very open. The closed source models keep getting better. The people working at the big labs say progress is continuing at a very rapid pace. There's periodic concern that capabilities are topping out, but the people at the big labs have tons of new ideas. They're optimistic. The open source models also continue to get better. Every month there's another big release. The other nice benefit of open source is that it's easy to learn from. If you're a computer science professor or student, or a normal engineer trying to learn, the existence of state-of-the-art open source models is amazing for education. This is leading to a rapid expansion of knowledge about how to build AI, compared to a world where it was all bottled up in two or three big companies. AI researchers are at an enormous premium, getting paid more than professional athletes. But shortages create glut. Some of the best AI people in the world are 22, 23, 24 years old. They've come up to speed over the last four or five years, and there will be a lot more in the future. The long-term answer may well be both. There will be a large business of the smartest thing regardless of cost, and a giant volume market of smaller models everywhere.

Big companies are definitely playing hard: Google, Meta, Amazon, Microsoft. Then there are the new incumbents like Anthropic and OpenAI. In the last two years, we've seen the birth of brand new companies that are almost instant incumbents, like xAI. Mistral is the great outlier to my Europe thing, doing very well as the French national AI champion. There are a bunch of these now doing quite well and becoming new incumbents. There are also tons of foundation model startups. We funded Ilya from OpenAI to do a new foundation model company, Miriam from OpenAI, and Fei-Fei Li from Stanford to do a world foundation model company. There are new swings, all early but very promising. On top of that, there's a giant explosion of AI application companies. These are startups that take the technology and field it in a specific domain like law, medicine, education, or creativity. It's amazing how sophisticated things are getting very quickly. Take an application company like Cursor. They take core AI capability purchased by the drink from Anthropic or OpenAI and build an AI coding system on top. The critique has been that these are 'GPT wrappers' that won't preserve value because the AI isn't theirs. But what's happening is the opposite. The leading AI application companies are discovering they're not just using a single AI model; they're orchestrating multiple models and adding their own proprietary data and workflows. They're building deep moats through user experience, domain expertise, and data network effects. So I think application companies can create tremendous value.

They actually, as these products get more sophisticated, they actually end up using many different kinds of models that are custom-tailored to the specific aspects of how these products work. They may start out using one model, but they end up using a dozen models, and then in the fullness of time, it might be 50 or 100 different models for different aspects of the product. And then B, they end up building a lot of their own models. A lot of these leading-edge application companies are actually backward-integrating and building their own AI models because they have the deepest understanding of their domain and they're able to build the model that's best suited to that. And then by the way, also AI open source, they're also able to pick up and run open source models. If they don't like the economics of buying intelligence by the drink from a cloud service provider, they can pick up one of these open source models and implement it instead, which these companies are also doing. The best of the best of the AI application companies are actually full-fledged deep technology companies actually building their own AI.

Well, some of them. I mean, I will let them announce whatever they're doing whenever it's appropriate, but some of them are now also doing big model development. And again, this is also part of the learning just in the last two years. Here's a big learning from the last two years, which is very interesting. Two years ago or three years ago, you would have said, 'Wow, OpenAI is way out ahead, and it's probably going to be impossible for anybody to catch up.' Then it's like, okay, well, Anthropic caught up, but they came out of OpenAI, so they had all the secrets. So okay, they caught up, but surely nobody can catch up after them. Then very quickly after that, there were a raft of other companies that caught up very fast, and xAI is maybe the best example of that. xAI, Elon's company, basically caught up to state-of-the-art OpenAI/Anthropic level in less than 12 months from a standing start. That kind of argues against any kind of permanent lead by any one incumbent that's just going to be able to lock the entire market down. And then as we've discussed, the China part is all new in the last year. The DeepSeek moment was in January or February of this year, less than 12 months ago. Now you've got four Chinese companies that have effectively caught up. These are trillion-dollar questions, not answers. It's one of these things where once somebody proves that it's capable, it seems to not be that hard for other people to catch up, even people with far less resources. I don't know what that does. Maybe it makes you slightly more skeptical in the long-run economics of the big players. On the other hand, maybe it makes you more bullish about the startup ecosystem. It certainly should make you more bullish about startup application companies being able to do interesting things, which is why we're so excited about that. It should make you a bit more excited about China. On the other hand, the Chinese competition putting pressure on the American system to not screw itself up is very positive. So it should probably make you a little bit more bullish on the US. These are live dynamics, and I think we still need more time to pass before we know the exact answer. Sometimes I freak people out when I say these are open questions. When a company is confronted with fundamentally open strategic or economic questions, it's often a big problem because a company needs to have a strategy, and the strategy needs to be very specific. A company has to make very specific concrete choices about where it deploys investment dollars and personnel, and the strategy has to be logical and coherent or the company collapses into chaos. Companies need to answer these questions, and if they get the answers wrong, they're really in trouble. Venture has its issues, but a huge advantage we have is we can bet on multiple strategies at the same time. We are doing this. We are betting on big models and small models, pre-trained models and open source models, foundation models and applications, consumer and enterprise. The portfolio approach means we are aggressively investing behind every strategy that we've identified that we think has a plausible chance, even when that's contradictory to another strategy that we're investing in. The world's messy, and probably a bunch of things are going to work. There's not going to be clean yes or no answers to a bunch of this. A lot of the answers are just going to be 'and' answers. If one of these strategies doesn't work, we're not trying to hedge per se, but we're going to have representation in the portfolio of the alternate strategy, and we're going to have multiple ways to win. That's the goal. That's the theory of why we are taking the approach in the space that we're taking. That's why I have a big smile on my face when I say that there are these big open questions, because I think that actually works to our advantage.

Disagree and commit. We agree. I mean, Ben and I, we're an old married couple, so we argue constantly, but we've been...