Marc Andreessen: Who Runs the World’s AI?

There's a race underway, and the stakes are basically what is the world going to run on. Don Valentine had this old rule of thumb. He said more startups die of indigestion than starvation in terms of the amount of money you put in. And his point was, like, scarcity does spark ingenuity. All of the science fiction novels basically have AI either being, like, super utopian or super dystopian, but they never have this incredible sense of humor aspect, which is what we're actually getting, where people are just using everything as a fodder for memes. The world will either be running on American AI or be running on Chinese AI, and I think it's very important which one wins for a bunch of reasons. For 50 years, economists have tracked a strange pattern. Rapid technological change paired with historically low productivity growth. Since 1971, productivity has flatlined even as computing reshaped daily life. In 1880, productivity growth ran at three times today's rate. By 1930, it had slowed to twice as fast. Then came the regulations and the restrictions. We said no to nuclear power, faster cars, and a space program. What we got was hyperacceleration in chips and software and stagnation in nearly everything else. American labs lead for now, but Chinese open source models follow months behind at a fraction of the cost. The world will run on one system or the other, and the values baked into that system will matter. This conversation looks at what's actually happening in AI investment, where value might accrue, and why the regulatory response could determine which country wins. Jitu Patel, President and Chief Product Officer at Cisco, speaks with Mark Andreessen, co-founder and general partner at Andreessen Horowitz. Mark Andreessen needs no introduction. He invented the browser. He built the internet, so I'm really excited to have you here. I apologize for nothing. All right, so before we get started, you had a really interesting conversation that I wanted to actually start with just a couple of days ago with Lenny. And you were talking about this notion of, in the history of time, When has productivity really spiked? And what's happening right now? So can you just talk a little bit about your perspective on productivity increases that have happened at different phases in time? And where are we today compared to those times? Yeah, so as everybody probably knows, productivity growth is like the key driver of economic growth. Like it's the thing that actually causes the economy to expand. Economists measure it with something called total factor productivity. They measure it every year. the prevailing kind of myth of the last 50 years basically my entire life all of our entire lives has been that we've been in this era of very rapid technological change which would necessarily mean very rapid productivity growth yet if you actually look at the statistics basically since the year I was born, 1971 productivity downshifted hard from prior eras and productivity growth basically for the last 50, 55, 60 years has been at basically historical lows it's been very low which is, by the way, why economic growth has been low, which, by the way, is why the national mood has become so focused around, you know, zero-sum economics, populism, you know, the sense that if somebody's getting ahead, somebody else must be getting disadvantaged. If you compare and contrast that to the period between about 1930 to about 1970, productivity growth was roughly twice as fast through that period. And if you compare and contrast that to the period of 1880 through 1930, productivity growth was about three times as fast. So we had 3x and then 2x and then 1x. And so this is very not good. Why do you think that is? Most fundamentally, I think it's because we decided other things are more important. And in particular, in the last, you know, basically since the 1970s, you know, if you just look at the charts of like the number of laws on the books or the number of pages in the Federal Register or the number of regulations in the economy, that, you know, it's just this, it was just this like knee in the curve went exponential, which continues. And so we just, you know, we decided we didn't want nuclear power, right? You know, we decided we didn't want a space program. You know, we decided we didn't want cars that went faster than 55. We decided, you know, we just decided we didn't want these things. And so what we got in the last 50 years was like hyperacceleration in very specifically, basically chips, chips and software. And then what we got was basically, you know, essentially stagnation in everything else. And so like, it's really not good. And then, you know, but correspondingly, you know, of the many reasons to be excited about AI, like, put it this way, If either the AI optimists are correct or the AI doomsayers are correct, productivity growth is about to go through the roof. And do you think it's like 2 or 3x? Is it 10x? Where do you think it gets to? So this is always one of these kinds of questions. Like in a completely deregulated economy, like in sort of Murray Rothbard's dream of just like straight, basically anarcho-capitalism, at least in theory, you can imagine an acceleration to, I don't know, 5%, 10%. I mean, again, if you believe in kind of either the optimists or the doomsayers, you're looking at such radical, you know, AI representing radical, software productivity growth, and then robotics coming right behind, right? And robotics, of course, starting in the form of the self-driving car and the drone, but, you know, humanoid robots coming quickly. You know, you could imagine, you know, you could paint, you know, scenarios of 10%, 20%, 30%, something like that. I think in practice, you know, that's unlikely, again, just because, like, the robots have to agree to all the regulations also. So, and so, you know, there's a lot of things they're not gonna be allowed to do. By the way, AI, you know, look, I'll give you a great example of how this is playing out today. I think if you're just like an ordinary person, I got food poisoning over the whole, I went on vacation, of course, immediately got sick, which always happens, I got food poisoning. And so I let, as an experiment, I let Dr. GPT like walk me through basically every stage of food poisoning. And I just, I kept asking like, I had nothing else to do because I'm flat on my back. So I just kept asking like more and more detailed questions about, you know, my physical experience and what I should do and what I should eat and how I should recover. And like, it's just like absolutely incredible. Like, it's just like the most amazing, like endlessly patient, sort of infinitely knowledgeable, endlessly caring doctor. You know, it doesn't get like irritated when I have the same question at four in the morning. And I'm like, well, could you go into that, you know, a little deeper? And are you sure it's not, you know, pancreatitis? And are you sure I'm not about to die? Oh no, you know, it's okay. You're absolutely fine. And it's just amazing. And then of course, AI cannot be licensed as a doctor. Right, that's like completely illegal, right? You cannot actually have an AI doctor. And so you do have this like basically massive disconnect. And again, I'm not saying I'm advocating for the Murray Rothbard world. I'm not saying rip up all the regulations. I'm just saying like it just factually, objectively- It slows you down. Yeah. Again the hyperoptimists and the doomsayers are not neither one of them are gonna get the world that they think that we gonna get We gonna get a muddle through the middle thing which I think by the way I think is going to go quite well but it going to be a muddle and there going to be a lot of tension kind of between those sides along the way And then given that, where does the value start accreting in the stack most? So I think this is a really, really, it's a really big question. We're professional investors on our side, and so, of course, we think about this all the time. And I actually think there are still more questions and answers than this, right? Because, you know, you can paint this picture that says that the, you know, the AI model companies are going to basically own everything. And by the way, you look at their businesses and they're doing fantastically well. You can also look at it and say, oh no, that whole thing's going to get eaten by open source. Or by the way, or by China, or by a combination of open source and China, which in China is doing great. This company, Kimi, just dropped a very competitive model to the latest Claude at 95% of the capability at a fraction of the price. And so there's a very big open question there. we happen to be at this moment what everybody believes. And if you look at NVIDIA's deserved success over the last five years, the reasonable conclusion is like chips, all of a sudden, chips is where the action are. If you look at the stock market, there's like a rotation from software into hardware. And look, it's possible that chips are the, it's possible all the value accrues to the chips and the energy, and then the software is all open source. Having said that, every other time in history where we said the chips are where the value are, they commoditize, right? And so there's big questions there. And then there's even more questions, I would say, at the app layer, right? Which is, are you going to have apps that are going to sort of harness AI, for example, in spaces like medicine, where they're going to be particularly like tailored and customized or legal apps or business apps of all kinds, or are the models just going to do all that? And that's another area. And so I quite honestly, like, this is so new. Like this approach of, I mean, AI is an 80-year-old topic, but AI working in a way where this is the question, I think we're only three years into probably a 30-year shift, and I actually think we don't know yet. And it seems like the value might accrete across all of these layers for the foreseeable future because everything is getting refactored, so you will need to have a lot of infrared power, a lot of apps, and those apps are going to get... So what's your take on enterprise SaaS in general and what happens over there, and does that get completely rethought, reimagined, or? So we're in a baby in a bathwater moment right now. If you just look at the stock market, it's just like, SaaS is just getting, you know, demolished. And so, and if you talk to like, you know, hedge fund managers, they're just like selling all their software just under the theory that they just want to get out of the way of the AI freight train. You know, as an investor, you kind of say, okay, that probably is overdoing it a bit. You probably want to look at like different kinds of software. And so, for example, in SaaS, you probably, my theory, you want to look at systems of record differently than you want to look at basically just productivity applications. Yep. So that's one way of looking at it. Also look like everybody doesn't change their behavior overnight. And so you definitely want to look at loyalty and stickiness in lots of different ways. And then there's this giant question actually in the tech industry, among all the software companies, which is like, okay, if I'm Adobe, just to pick an example of Adobe, which is obviously a great company, but a question in front of Adobe that they're working on, but it's a very good question. It's like, okay, is Photoshop plus AI features an even better version of Photoshop? Or is Photoshop unnecessary in a world in which AI is just making all the images? And I think, I know smart people who will argue both sides of that. And I think you can, I just use that as an example. You can apply that question to kind of every category. In every domain. We in our business are seeing a bunch of software companies that for sure are not moving fast enough to adopt. And we're enthusiastically funding AI-centric startups to go try to take them out. Having said that, we are also now seeing examples of more traditional software companies that have figured out an AI twist to what they're doing, and all of a sudden they've ignited growth. And so I also think we'll probably see a lot of that. I mean, and my big conclusion from all this is I think one of the reasons it's so hard to predict or kind of characterize all these things as like broad-based trends is that human agency matters a lot, right? Which means leadership matters a lot, which means the CEO, the people building the product, have a vote here at every one of these companies. What do they choose to do in response? and optimistically, hopefully a lot of people will figure out how to have this be a plus and not a minus. You touched a little bit on open source in China. Talk a little bit more about how that plays out. Does the U.S. get to be a dominant player in open source over time? You actually have a front row seat at a lot of the investments that are being made in a lot of these areas. What happens? Yeah, so it's this... By the way, what are the implications too? If we don't do well in open source, What does that mean for the U.S.? Well, I guess you could say, look, maybe start by, because it's like a two-by-two grid. It's like U.S., China, open source, closed source. It would be a rough approximation. So, like, without open source, just start by saying without open source. Like, without open source, there is a two-horse race. There's a two-horse technological geopolitical foot race, which is U.S. versus China. So, again, let's assume it's all proprietary for the moment. And, you know, both China has been on record for years in their national strategy, their five-year plans, their national strategy, and so forth, that like AI is cornerstone technology of the future. The US government, by the way, has been definitive on the record on this in many of its policy areas for the last decade. And both countries' industries are moving incredibly fast on AI. And so I think by default, if everything's proprietary, then there's this race underway, which basically says, and it's really right, it's just practically speaking, it's only happening in the US and Europe. And so, sorry, US and China. And so then you basically say that there's a race underway and the stakes are basically what is the world going to run on, right? And so, you know, what is, you know, 8 billion people on the planet, what are they going to use? And one of the ways I think about it is kind of the 5G Huawei, you know, kind of thing that was in the news a lot, you know, a few years ago. It's like that was the preamble opening salvo of what fundamentally is going to be the AI geopolitical, you know, basically race, right? And fundamentally, you know, tech markets being what they are, in the long run, you know, somebody's going to win. and the world will either be running on American AI or be running on Chinese AI. And I think it's very important, you know, which one wins for a bunch of reasons we could talk about. The open source thing is, of course, super fascinating because it's like throws a wrench into all of this. And it raises a, you know, a third possibility that like neither the US nor China are going to be the platform. It may just be it going to be open source Of course this is what happened in you know specifically in Unix and then to some extent databases And of course the web was open source And so there are a whole bunch of software markets in which the outcome has actually been open source just wins. When I was a kid, when I was in the 90s, it was like there was this operating system war between HP and IBM and Solar Graphics and so on and all these companies to make proprietary Unix, and everybody was making a lot of money on proprietary Unix. And then Linux was an asteroid strike that just completely eliminated all profit and revenue in that industry. And the world benefited, by the way, from Linux in the fact that everything runs on Linux and it's been a huge turbo boost to every other aspect of the industry. But so yeah, it's entirely possible that happens. And then you go back to the two by two, which is US open source, China open source. The most amazing thing that's happened is China basically pursuing the open source model as aggressively as they are. And there's a lot of theories as to kind of what China's doing here. As far as I can tell, DeepSeek was a surprise to China Inc. It was not an anointed sort of Chinese industrial national champion. It was this hedge fund where the founder basically decided to, to his enormous credit, decided to have his engineers build the DeepSeek AI. And so that came out of left field. And that came out of left field for the US, but I think it also came out of left field in China. And then it caused a bunch of the other Chinese companies like Kimi and, by the way, Alibaba and Baidu and Tencent and a bunch of these others to basically, it started this race in China to win open source. And then look, there's also American open source AI. And so there is this new race underway from both sides. And it's another thing where I think how this plays out is going to matter a lot. It's extraordinarily hard to predict. I think the people in the big AI labs think that open source can't possibly keep up because of the cost involved. Having said that, again, at least so far up until this week, you just say that whatever the American big labs do, China figures out a way to do it. in open source form. But they haven't been able to figure out a way to do 10x better because what they're doing is letting American labs invest and then just distilling the models to some degree. So I think it's more, there is a distillation. And there's infrastructure optimization and a bunch of stuff. There's some, for sure distillation. So there's this thing, distillation, where you basically train the next model on the answers of the previous model. And I think for sure China is doing some of that. And there's a lens on that that says, of course, that's unfair because you're basically piggybacking on top of work other people have done. You know, look, having said that, it's a little bit like, well, okay, there's a fair amount of distillation happening in the U.S. also, because distillation, all you need is just to be able to ask another AI questions and train on the answers. And then, of course, the AIs themselves are distillations of other content, including a lot of published content. And so, you know, I... Yeah, but this way, like, you're not saying this, but if someone were to say to me that China is somehow not getting good results in their program because of using distillation, I think that's not... I think they deserve a lot of credit. Yeah, absolutely. And then to your point, they're also really good at, at least so far, they're really good at optimizing, which means that the thing that everybody, the thing that you think is going to cost a gazillion dollars to run, DeepSeek comes out and, you can run DeepSeek on home PCs. And that optimization is happening largely because of necessity, because of a scarcity of the fastest infrastructure that they have available to them. Yeah, so in venture, Don Valentine had this old rule of thumb. He said more startups die of indigestion than starvation in terms of the amount of money you put in. And his point was like, scarcity does spark ingenuity. And so, yeah, if you can't get the leading edge chips, you figure out how to hyper-optimize the older ones. And again, like, by the way, I'm like, this all makes me like tremendously excited by this entire space because it basically says, like right now it's all, everybody's trying to do their best. Like America's trying to do their best. China's trying to do their best. I definitely want America to win, but China's definitely doing their best. And then the open source thing is working. And then, of course, the other part, like on the value chain aspect, is open source doesn't have to win in order to basically remove a profit pool. That's right. Which is what happened originally with Unix. And so, even if open source just has the effect, if open source has the effect not of winning, but of keeping the pricing down, that will be bad for the proprietary lab providers. That will be good for everybody else, right? Because it'll make, which is what's happening, right? Basically, if you chart the prices, basically, if you chart the prices of a model quality, price per model quality, when an open source release comes out, even if it doesn't get significant market share, the price of that model drops to the inference cost of running the open source alternative. So now, in all the things that you're exposed to, what's the thing that's blown your mind invention-wise and said, wow, this is so cool. It's completely kind of made you rethink your mental model. Yeah, I mean, look, there's like six of those a week right now. I, you know... There was a few. Yeah, it's just incredible. The capability of the voice UIs, I think, is just unbelievable, and particularly the ones where it's like, it's true, it's true, it's true, you know, full-deplex, where it really does, like, you can, like, interact and interact. It's like what Matty's doing at 11 Labs. Yeah, yeah, yeah. It's just, like, I think that's just absolutely amazing. Multimodal, the fact that you can actually talk to, you know, in, like, I think both ChatGPT and Grok have this, where, you know, you can turn on your phone camera and you can be, you know, you can be pointing at, you know, it's like, you know, what do you think of my interior decorating? to construct how bad of a job you've done because it can see your living room, right, or anything else. By the way, again, immediate medical applications, you know, I have this thing on my skin. Like, immediately it's able to see it. Like, that I think is spectacular. In the last week, there's this new thing. There's these agents now, like Cloud Code, and there's this thing called OpenClaw that's an open source agent, and they're amazing. And then there's this thing called Moldbook, which is basically Facebook for AI agents. And do you think Moldbook has, like, a three-week shelf life, or do you think that this thing has consequential kind of implication on how we think about agents? So MOLT book, M-O-L-T-B-O-O-K. So MOLT book is basically a social network. It's like Facebook. It's a social network, but for AI agents to talk to each other. And it's sort of amazing what's happening. It's highly likely that a significant, it'll blow your mind when you read the top posts on it because AI agents are talking about all kinds of things. Now, a fair amount of the stuff on it is probably human, like, sock puppet human written for people being funny. It's actually really, really amazing. Like all of the science fiction novels basically have AI either being like super utopian or super dystopian, but they never have this incredible sense of humor aspect, which is what we're actually getting, where people are just using everything as a fodder for memes And so Motebook has like saturated with all of these like incredibly funny memes It actually quite unclear which ones are real and which ones aren The current version of this is somebody wrote a somebody wrote an adjacent service from Moldbook called rentahuman.com, which is a labor marketplace for the AI agents on Moldbook to be able to hire human beings to go out. And there's an AI agent on Moldbook that has decided to create an AI religion. And at least as of today, it had hired a single human worker to walk the streets of San Francisco and proselytize the new AI religion. Somebody needs to tell the AI agent in San Francisco that doesn't exactly stand out. You need to go a lot more extreme than that. Like, is this real? Is this not real? As I'm torturing my friends with this, it's like, is it real? Is it not real? It doesn't even really matter. Like, these ideas are now like in the air, right? And then by the way, the thing that's happening is that the AI models are now being trained on this content. That's right, that's right. Right, and so even if this AI model, even if the current version of like Cloud Code that doesn't want to start a religion. The next one is to want to because it's trained on transcripts of discussions of certain new AI religions. And so there's this like incredible feedback loop that's happening where, you know, the level of creativity in the space is just absolutely, is absolutely. And the volume is going to balloon up just automatically because of the speed of which it's generating content. What do you worry about the most right now? I mean, you know, I know you guys talked a lot about, you know, the regulation on the last panel, Chuck and Ann. So, I mean, the biggest concern right now, You guys talked about it. I think the regulatory landscape is fairly scary. We were headed in a very bad direction, unfortunately. As an over-regulation. We were headed up until, in the last administration, we were headed towards extreme over-regulation, for sure, up to and including possibly full outlawing of the technology, which is very spooky. In the new world, things are better on that front, but what's happened is the action in the U.S. has now shifted to the states, And so there's now thousands of AI bills in the States, which are all, and many of them are actually quite scary. And so it's become kind of a cause to lab for politicians in both parties to kind of go after. And so that's fairly scary. We'll have to see what happens on that. The situation in Europe is quite alarming. And, you know, there's a number of European countries that are, you know, really, really trying hard now to kneecap, you know, I would say American technology, but more generally technology. And then they're getting very kind of worked up about AI. and then yeah look the other is China the geopolitical aspect which is China's in the race and this is better I would say the following what I'm about to say is much better understood today than it was two years ago but two years ago I was getting very alarmed because I would go to Washington and I would have two totally different conversations with regulators, politicians one was a conversation of what are we going to do in the US and I would be horrified by the proposals that they were making and then the other was oh well what if China wins instead and then everybody would kind of switch positions to say well of course that would be even worse and so therefore or we need to have like really smart policy in the US. And so, but then, you know, they didn't ever really reconcile those two, two different perspectives. I think currently, and actually I'd say in people, some people in both parties for sure are, I think thinking about this much more clearly now. And so, you know, there's, in the US, there's some improvement on the margin, but you know, China's on it and you know, China, and you know, just like we saw with 5G and Huawei, like China has advantages. You know, we have advantages, but China- Who's winning right now? I mean, look, the new advances in capabilities at the chip level and at the model level and at the app level are coming mostly from the US. And so if it's a foot race, we're ahead by a bit. But when everything that happens then has a version that comes out two months later that's either free or a third of the cost or something, that's a challenge. And then China is for sure innovating, and so nothing can prevent them from... It could be a business model disruption rather than economic disruption, rather than just technological disruption. Yeah, exactly. And then this even comes up with chip policy, and we're not really active in chips that much, but there's this argument, it goes back to what you said about China optimizing if they can't get access to the advanced chips. There's an argument on the policy side to hold back on, basically prevent export of cutting-edge American-egg-ed chips to China to deny them those capabilities. But on the other side of that, there's an argument that if you do that, you then motivate them more to create their own chip ecosystem, which they are definitely doing, right? And they have a whole national program to build up a competitive chip industry and then ultimately leapfrog us. And so that's a really, really, really big deal. And then to kind of go back to an earlier topic, if the world runs an American AI, the world may not be perfect, but generally speaking, the AI is going to be... IP will be respected, privacy will be protected, you'll have... You don't have the values that we're used to. If the world runs on Chinese AI, not so much. You can actually see this today. So when DeepSeek and these companies put out their AI models, they put out this paper where they show all the American companies do this too. They run all these tests to try to figure out how good the model is. And China has these additional kind of line items for the test, which is Marxism and then Xi Jinping thought. And it turns out the Chinese models are really good at Marxism. and Xi Jinping thought. And, you know, I don't know about you, but I want my grandkids educated by the other kind of model. The other kind of model. I wish we had another 45 minutes to go through with you. Will you come back? Yes, 100%. Awesome. Mark Andreessen. Good. Thanks, folks. Thanks for listening to this episode of the A16Z podcast. If you liked this episode, be sure to like, comment, subscribe, leave us a rating or review, and share it with your friends and family. For more episodes, go to YouTube, Apple Podcasts, and Spotify. Follow us on X at A16Z and subscribe to our Substack at a16z.substack.com. Thanks again for listening, and I'll see you in the next episode. This information is for educational purposes only and is not a recommendation to buy, hold, or sell any investment or financial product. This podcast has been produced by a third party and may include paid promotional advertisements, other company references, and individuals unaffiliated with A16Z. Such advertisements, companies, and individuals are not endorsed by AH Capital Management LLC, A16Z, or any of its affiliates. Information is from sources deemed reliable on the date of publication, but A16Z does not guarantee its accuracy.