Ben Horowitz 谈 AI 焦虑、大科技转型与初创公司的未来 | a16z

America’s got to rebuild its entire infrastructure right now. We don’t have enough rare-earth minerals, electricity, or manufacturing capacity. NVIDIA will make enough chips, but then we won’t have enough memory. Almost everything is a bottleneck.

The China graph is like this, and the U.S. graph is like that. How do we make this seem less scary?

The history of technology is that things have always gotten better. Humans are kind of unbelievable in their ability to come up with new things that they need. Now 8 billion people who might have an idea in their head can get it out of their head.

So you’ve been doing this for a long time. I thought maybe I’d start off—and it’s funny, we actually didn’t rehearse this at all, because I thought that way it would be more real, right? More unique.

Let’s talk about your book, where you talked about how hard it is to be a CEO and everything that you went through at Loudcloud and Opsware. That was a giant shift where the market kind of collapsed, the financial market collapsed, and you had to really pivot and change the company.

What do you think a legacy company, or a company from 5 or 10 years ago, should do when there’s this great opportunity but also a great challenge? There are new-age, AI-first companies popping up right now. Hopefully, they’ve got their shit together; they’re off to the races building something new. But what does a 5- or 10-year-old company do when it’s pre-AI? They’ve got to figure out what they do.

Markets hate them.

Yes, yes. The financial markets hate them. I don’t know—maybe riff on that. I’d love to hear your thoughts.

I think the first thing you have to recognize in a huge dislocation like this is that some very basic, axiomatic laws of physics are different. The 2 that are really different with AI, compared to what we’ve been building in technology forever, are these.

First, it used to be very well known that you cannot throw money at the problem. For example, if I had a product and I was 2 years behind, I could not hire 1,000 engineers and catch my competitor. It’s The Mythical Man-Month. 9 women can’t have a baby in a month. Everybody knows that. It never works.

No problem. That’s no longer true. You can throw money at the problem. If you have enough money and some good data, you can buy enough GPUs and solve basically anything in software. So that’s gone.

The second thing that we knew for sure is that, in software, possession is 9/10ths of the law. If you have the customer, you have multiple lock-ins: migration-pain lock-in, data lock-in, and user-interface lock-in. Those are pretty much gone, right?

It’s very easy to replicate the code, and it’s very easy to move the data. Then it’s not even going to be a human talking to your software; it’s going to be an AI. AIs are really flexible in how they use user interfaces, so that moat is gone.

I think that’s the first thing you have to recognize as a CEO: that’s going away. So what is it? Where is your value? What are you delivering? It turns out there are many things that are valuable, but if you’re trying to get good pricing for any of those things, you’re going to be under tremendous pressure. Your price has to be a function of some other, much more distinct value that you provide.

Got it. The other thing that we’ve talked about a lot internally as a firm is that, once upon a time, if you had a good product, you might have 10 years to run with that product—maybe 5 years. Now it might be 5 weeks.

We’ve also talked about this in terms of going public. Companies are staying private a lot longer, which is probably good. If you’re going through an existential crisis, you’d much rather do that as a private company than a public company.

But the reason why the SaaS apocalypse is happening is that there are doubts about terminal value. Everybody who starts a company is doing it because they want to create economic value. They’re capitalists; they’re trying to benefit financially from this equation. But if you wait too long, maybe your company is worth zero. That’s kind of scary. That was always a risk, but it would play out over decades.

Not as fast a risk. I guess, what do you think? If Loudcloud were around today and you were the CEO again—bad, bad example. Sorry to give you that.

I know, I know. Although, actually, Loudcloud would be very well positioned.

Yeah, exactly. You would be very well positioned. But I guess, what is it that a CEO should do potentially differently? Obviously, move faster, cut faster, be more efficient, throw money at the problem—all these things that we’ve talked about.

But if I don’t go public, if I go public and get disrupted, then I have this terrible life of becoming a penny stock. If I just wait, there’s a chance that I get eviscerated. This kind of roadkill-success equation is scary, right? I mean, it’s always scary, but you would have time. And now it feels like you don’t.

I think you do have to be honest with yourself about what it is you really have. There are companies that get thrown under the bus correctly and ones that don’t.

If you take a lot of these ideas to their logical conclusion, then nothing is worth anything, because there are no people at companies. And if there are no people, who’s going to buy your shitty software? But it is more subtle, and it just tends to take much longer than we think for some of these things to play out.

The question is, are you getting stronger in the meantime, or are you degenerating? Is what’s happening that nobody’s buying? The money just shifted. The customers are buying other stuff; they’re not buying yours. In that case, you have a huge problem. You probably have to cut deep and pivot.

On the other hand, there are companies that have been slaughtered in the valuation game but are pretty strong. I’m on the board of Navan, right? They’re a travel company. Obviously, according to the SaaS apocalypse, they’re dead. There’s no way you’re doing travel.

But then you look under the covers and realize it’s a little more complicated than that. With travel, you actually need explicit relationships. If I’m providing your travel and you’re any kind of company that’s important at all, you need to travel globally. So now I need a relationship with every single airline in the world, every single hotel in the world, every train—everything. You’ve got to deal with that. You’ve got to connect back to their budgeting systems and all these things.

The second thing is that nobody wants to do this, including OpenAI or Anthropic: sell to the damn travel manager. Nobody has a channel to the travel manager. You can’t even imagine that being a good idea. You want to keep advancing. You want to do the things that Intuit is doing, where you turn yourself into more of an AI company and then hold the customer.

And by the way, the agentic travel experience turns out to be much more complicated than one would think. I don’t know if it stays that way, but that’s the way it is today. So I think it’s very company-dependent. I don’t think it’s all one thing.

But I do think it’s a brave new world, and if you keep looking at it like the old world, when it has completely different laws of physics, you are definitely going to die.

Yeah. Well, maybe let’s talk about venture capital.

There’s a lot of cope going on now, too, so you have to be careful with that.

That’s the thing. There are some things that really are features, and before, it would take a long time to build a feature. You might as well do something else because of comparative advantage. David Ricardo: I could weld my own steel, and I could grow my own food, but I’m just not going to do that because I can do things that produce more economic value for me.

Now it’s just becoming not that hard to create features. But features are not products, and products are not companies. We’ve always had this distinction: feature, product, company. But it’s a little confusing figuring out which one is which right now, because of the ability to create a feature, create a product, and even get all of the data.

You know my favorite saying: the best companies have hostages, not customers. Even getting some of the data out of the hostage company is possible now. It’s a very confusing world in terms of figuring out which one is which, which is maybe a good segue to venture capital land.

When you started this firm in 2009, a big financial crisis—actually, a very, very big financial crisis, the global financial crisis—was going on. The world has changed a lot since then. How much of what’s happening today fits within the mental model of back then, and how much is a brave new world? Maybe riff on that a little bit.

The biggest. It’s really different. Our first fund was $300, and we raised it from all the traditional LPs: endowments, charitable foundations, funds of funds, and so on. We’ve just raised $15 billion for 4 of the 7 funds—not even the whole complex.

We raised it from very, very different kinds of investors. Basically, none of our LP base was international when we started, and we're at about 35% international money now, from all kinds of places. Tech has gotten so much more important. I think we have to think in terms of the world in a way that we just didn't before.

So, for example, why do you raise so much money? Which, by the way, I'm kind of mad at myself because I don't think I articulated it internally well enough, because we could have raised even more money. It's not—don't worry.

Yeah, we had more money on the table. But the way I was thinking about it is, look, America's got to rebuild its entire infrastructure right now because we don't have enough rare-earth minerals, electricity, or manufacturing capacity. We have the wrong chips—they take way too much damn power. They were built for games. We don't have enough of anything to be in this future world, and somebody's got to fund it. Clearly, that's going to take a lot of money.

All that is brand new, and I would say it's fairly overwhelming in a sense, but it's really, really important. We're pretty much out of electricity down in the United States—not 12 months from now, but right now. The China graph is like this, and the U.S. graph is like that.

Yeah, and the demand for these tokens is straight vertical. But the ability to build that capacity is absolutely not vertical, so we need new era. We invested in a transformer company—not an AI transformer, but an actual power-transformer company—because you need better, easier-to-manufacture, more efficient transformers. The transformer hasn't changed since we invented electricity.

There's an old saying: “The cure for high prices is high prices.” Yeah, but the problem is that there's a lot of latency involved. Right now, there are computers that show up with no RAM. If you go buy a server from Dell, they're like, “Sorry, we don't have any RAM to sell you,” because all of it has been gobbled up. They could build a new factory, or you and I could decide to build a DRAM factory, but that would take us 5 years.

Got to start now. Yeah, you got to start now, but this is actually—if you remember, which you obviously do—1999. It's like, “Well, we have to build more fiber,” right? “We have to build more capacity.” But it's obviously very different because all the GPUs are hot. They're all lit right now, whereas back then most of the fiber was dark.

Yes. But how do you—

Yeah, well, there were bottlenecks when we were building fiber, but the bottlenecks were in different places. The servers weren't capable of putting bits out fast enough to do video, and we didn't have load balancers. We didn't have application servers. We didn't have anything. So, you had all this fiber and all this bandwidth, but you couldn't actually build the applications. Most of the end users weren't on the network, either, so it just didn't work. Then we had the dot-com crash and all these things.

Now we're in a little different place because almost everything is a bottleneck. I do think what's going to happen is we'll probably have enough chips long before we have enough electricity. NVIDIA will make enough chips, but then we won't have enough memory, and we won't have enough electricity. So, we're in that kind of situation now. You really have to study where we are at each point in the supply chain and figure out how to alleviate those bottlenecks.

And, by the way, God bless Elon and the Terafab. That's the idea: he's going to deal with all the bottlenecks himself, which is how he does things and why we need him.

Indeed. I feel like you're an expert in 3 things: hip-hop, AI, and crypto. I don't know anything about hip-hop, but I've heard a lot from you. Let's talk about the other 2, in particular crypto and AI.

I actually just wrote about this. You remember, the origin of crypto was Hashcash. The scariest thing right now, from my perspective, is that everybody with Claude or ChatGPT can go super deep and personalize a phone call or an email. It seems like all communication is going to be completely unusable. I don't know if you agree with me.

I 100% agree.

Because normally I can just delete, delete. I got an email yesterday that said, “Dear Alan at Index Ventures.” It's like, “Well, I'm not Alan. I don't work at Index Ventures. Delete.” I'm very grateful that this person messed up my name, because I can just delete that.

The best way to think about an email inbox is it's a to-do list that has write access for the public. Anybody can get in, and now anybody can personalize. Same thing for phone calls. What do we do?

It seems like there's a lot behind crypto, and that's why I mentioned Hashcash, because it was originally intended to stop spam. Do you think there's overlap between AI and crypto? I know you do, so tell us about that.

Yeah, I do think it starts with the problems that AI causes. One of the first things—I woke up in the middle of the night one day and I was like, “Oh my God, somebody's going to go on a Zoom call. It's going to be AI me, and they're going to tell my finance team to wire $500 million to Nigeria.” That's going to be a problem.

Then we're like, “Okay, everything's a hardware root of trust. Don't believe anything from me unless it's got my cryptographic key on it.” All that kind of thing. I knew these problems were coming, but they're coming so fast now.

I think there are several categories of things. First is just: are you a human or are you a bot? I think everybody is going to really, really want to know that, whether it's social media, a dating app, a Zoom call, anything. You want to know, “Am I talking to an actual human?” Can I prove that I'm a human being? Then can I prove that I'm me? Then can I sign content? How do I know it's true?

There needs to be a distinction between—I get so many AI videos sent to me from my family that they think are not AI videos. They're like, “Did this really happen?” And I'm like, “No, you could actually ask Grok, and it's pretty good at that right now.” But Grok is getting to the point where it can barely figure it out, and I think at some point it won't be able to figure it out. AI will not be able to tell what's AI.

The only way is you're going to have to have something—some cryptographically strong indication, a signed piece of content that says, “Okay, yeah, I made this,” or, “This is really a video of me—Marco Rubio—giving a speech. This isn't something that somebody faked.” There needs to be a source of that truth.

Who are you going to trust for the truth? Are you going to trust Google? Are you going to trust Meta? Are you going to trust the U.S. government? I think you want to trust the mathematical, game-theoretic properties of the blockchain. I think that's going to be a very, very important part of the infrastructure.

Then you get into fraud. How do you know somebody's a citizen to get them money? Everybody's talking about, “Well, let's do UBI.” Well, great. But when we did the stimulus program, we found out that the government is very bad at getting money to people. Depending on the numbers you read, somewhere around $450 billion got stolen. What you really need is everybody to have an address where you can send them money. I think that's a crypto problem.

Finally, how does an AI become an economic actor? How do I make money as an AI? How does somebody send me money? Can I be a merchant, a credit-card merchant, if I'm not a human? I don't think so. I think that's actually kind of hard, and it's probably not the right infrastructure anyway. You need a bearer instrument on the internet. You need internet money for these AIs to be economic actors, and I think that's very likely to be crypto.

So, I think there are many opportunities in crypto that have been generated by AI. It feels like this old Yogi Berra saying: “It's so crowded, nobody goes here anymore.” We're entering that era, because number 1 is: are you a real person? But the problem is that co-work is so good right now that or you know, open claw. I just say, “You are a real person. You were a real person.” But now your addresses are being used by a machine.

Yeah, right. CAPTCHAs don't make any sense. CAPTCHA is an acronym. What is a CAPTCHA, right? It feels like the solution lies in economics somehow, and game theory.

Yes, yeah, and that, too. Are you going to just have to—well, maybe I think half-caste is kind of a relevant idea again.

Yeah, no, totally. So maybe why don’t we talk about where you think venture capital is going? I mentioned this because Mark got some crap for saying, “All the jobs will go away except for one job: venture capital,” which was seen as a self-serving comment. But in his defense, I will say it’s partially because it’s a nondeterministic problem.

Yeah, right?

It’s like, all right, you’re betting on an entrepreneur first and foremost, and you want to know that this entrepreneur, as I like to say, can materialize labor, capital, and customers. You can’t just run an algorithm on it. I mean, maybe you can, but there’s just not a lot of data out there. It’s very, very hard to do. So that’s the logic by which—and also, just personal relationships in general will probably survive AI.

Yeah.

But if there’s a venture capitalist, then that kind of assumes there’s an entrepreneur job now.

Yes, yes, that is true. It takes two to tango.

But, yeah, if you’re very bad, you just raise money and never allocate it, I guess. But I guess what do you think the world of venture capital looks like today? We’ve obviously done a lot of things internally as a firm to try to embrace AI very, very fully, but 5 or 10 years from now, given what’s potentially going to happen to white-collar work?

Yeah, I think it’s really tricky. You kind of go back to the last transition like this, which was the transition to the Industrial Revolution. The venture capitalists in the railroads, the automobiles, and so forth ended up becoming JPMorgan Chase, Goldman Sachs, and so forth. They ended up becoming banks, and some of the reason for that was just how fast that materialized.

I think in the ’30s, 20% of American workers worked for the auto industry, which is spectacular compared to what it is today. Things in the Industrial Revolution started out very much like we are today in venture capital, where there were, whatever, 300 auto companies and so forth. Then it consolidated very hard into, in the U.S., the Big Three and so forth, and the venture capitalists went upstream with the companies.

I think that’s one scenario where, okay, there are going to be a small number of very gigantic companies, and they’re going to own everything and so forth. There’s another kind of future where it’s like, okay, they got really big, and then we’ve finally hit the asymptote on this intelligence idea. They’re as smart as they’re going to be or whatever.

We’re either going to nationalize the big labs and say, like, they’re utilities. They’re electricity-plus-plus. Like, F-U if you’re going to think you’re going to collect all the money, and then everybody’s just going to build on this utility set of things. That’s a very different venture capital world.

So I would say—as I’ll quote Yogi Berra—the problem with predictions is they’re very hard, especially about the future. I think this future is particularly hard because it’s so dynamic.

How does the electricity shortage play into it? Does it make the big companies all-powerful because they suck up all the electricity and nobody else can get it? Nobody else can get any GPUs? Or does that push all the computing out to the edge, and then the models just get really good and small? Everybody’s like, “Well, I’ve got enough in my phone. What they’re going to charge me for their mega-GPU farm is just outrageous, and I’m just going to do that.”

So there are many ways it could go, and I don’t know. I guess I don’t know, but I could see venture capital being much bigger and much more exciting because everybody in the world is an entrepreneur. Or I could see it being more like what happened in the Industrial Revolution, where new companies are just harder.

Yeah, well, that’s kind of a good follow-up, or a good parallel question: How do we make this seem less scary? I don’t know if you saw Bernie Sanders. It’s a lot of change, you know; it is scary.

Well, but yes and no. I mean, 98% of Americans were farmers in 1789. I’m pretty sure they’re not farmers right now.

You made this interesting point: If you go to a third- or fourth-world country, if there is such a thing, everybody’s an entrepreneur. 100%. The guy says, “I sell bananas by buying them here and selling them there.” Everybody’s an entrepreneur. There were no organized companies.

The cool thing is that now 8 billion people who might have an idea in their head can get it out of their head. Maybe it’s a bad idea—probably is a bad idea—but there’s no longer a gate for them. There’s no capital gate. There’s no idea gate; it’s just, boom.

It’s not just for code. I can write music. I can make a movie. This is super exciting. If you’re trying to make this not look dystopian, I don’t know if you saw Bernie Sanders interviewing Claude. This is literally “old man yells at cloud.”

Yeah, yeah, yeah.

Like, metaphor, no metaphor, right? It’s just like he’s yelling at the cloud. That’s the dystopian view, and it’s wrong. I feel very passionate that that’s wrong, but we need a better narrative.

From a macro standpoint, I would say the history of technology is that things have always gotten better. Would you like to live in the world before electricity? Probably not. It doesn’t sound that appealing. You can if you want, but nobody seems to opt into that.

I think we’re very much in a period like that, but the transition is always scary because it’s a different world. Everybody was a farmer. Everybody was a farmer in 1750. I think it was 93% or 94% of America that was farmers, and then almost all those jobs are gone.

The jobs that we think are jobs are jobs that they would have thought were ridiculous. If you were a farmer, you would think what I do is the dumbest thing in the world—or a product marketing manager, any of this stuff. It’s like, that’s not a job. You’re not making any food. You’re not building a house. How could that be a job?

I do think it’s very hard to see to the other side of that, but I think it’s very, very likely to be way, way, way better for everybody, just like electricity ended up being way better for everybody.

And to me, the most salient wrong idea was from John Maynard Keynes. He wrote a paper that wasn’t that famous—the great economist of the Depression—where he said, look, things are going to be so abundant and everybody’s needs are going to be met. Everybody’s going to have a house, or shelter, and everybody’s going to have enough food to eat. Once you have your needs met, you’re going to work way less—15 hours a week maximum—because your needs are met.

But what he didn’t realize was that we’re not just going to need 1 car. We’re going to need a car for every person. We’re going to need computers and television sets and this and that and the other, and awesome vacations, and food that takes a chef 10 hours to prep, and all this kind of thing, which did not exist then.

There were no foodies and tasting menus and all that we have now. But that’s all a need. That want goes to a need very fast. Humans are kind of unbelievable in their ability to come up with new things that they need, and then you have to make those and so forth.

I think it’s going to be—I think in 15 years, the truth is everybody in America, and probably around the world, is going to live better than the very best life, in terms of luxury access to information and so forth, that anybody did in 1980. That’s the world that we’re almost certainly going to get to. So you shouldn’t be so mad about it. But it is disconcerting.

All right. Well, on that—

Especially if you’re trying to teach little kids, they’re like, “What should I do?” I don’t know. That’s a hard one.

Well, on that note, Horowitz at Andreessen Horowitz, thank you very much. We really appreciate it.

All right. Thank you.