Marc Andreessen 的2026年展望：AI时间线、中美之争与AI的价格

This new wave of AI companies is growing revenue—actual customer revenue, actual demand translated into dollars showing up in bank accounts—at an absolutely unprecedented takeoff rate. We’re seeing companies grow much faster. I’m very skeptical that the form and shape of the products 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.

I think we probably have a long way to go. These are trillion-dollar questions, not answers. But once somebody proves that something is capable, it doesn’t seem to be that hard for other people to catch up, even people with far fewer resources.

When a company is confronted with fundamentally open strategic or economic questions, it’s often a big problem. Companies need to answer these questions, and if they get the answers wrong, they’re really in trouble. In venture, we can bet on multiple strategies at the same time. We are aggressively investing behind every strategy we’ve identified that we think has a plausible chance of working.

If you want to understand people, there are basically 2 ways to understand what people are doing and thinking. One is to ask them, and the other is to watch them. What you often see in many areas of human activity, including politics and many different aspects of society, is that the answers you get when you ask people are very different from the answers you get when you watch them.

If you run a survey or a poll of what, for example, American voters think about AI, 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.

A lot of folks have sent questions ahead of time, and what I’ve done is curated them into a few different sections. In an AMA this morning with Marc, we thought we’d cover 4 big topics: AI and what’s happening in the markets, policy and regulation, all things a16z, and then we’ve got a fun catchall that we’re calling “Sandbox of Things,” if we get to it.

Starting with the biggest question: We’re sitting in the middle of the AI revolution. Marc, what inning do you think we’re in, and what are you most excited about?

First of all, I would say this is the biggest technological revolution of my life. Hopefully I’ll see more like this in the next 30 years, but this is the big one. In terms of order of magnitude, this is clearly bigger than the internet. The comparisons for this are things like the microprocessor, the steam engine, electricity, and the wheel.

The reason this is so big may be 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 was actually a debate among the people who invented the computer. They understood the theory of computation before they built the things, and they had a big debate over whether the computer should be built in the image of what at the time were called adding machines or calculating machines—essentially, cash registers.

IBM is actually the successor company to the National Cash Register Company of America. That was the path the industry took: building these 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. They couldn’t understand human speech, human language, and so forth. That’s the computer industry that was built over the last 80 years, and that’s the computer industry that 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 smartphones.

But they knew at the time—they understood the basic structure of the human brain. They had a theory of human cognition and, actually, a theory of neural networks. The first academic paper on neural networks was published in 1943, over 80 years ago, which is extremely amazing.

There’s an interview you can read or watch on YouTube with these 2 authors, McCulloch and Pitts. You can watch an interview with McCulloch from around 1946. He was on TV in the ancient past, and it’s an amazing interview because it’s him in his beach house, and for some reason he isn’t wearing a shirt. He’s talking about a future in which computers are going to be built on the model of the human brain through neural networks.

That was the path not taken. The computer industry was built in the image of the adding machine, but neural networks basically didn’t happen. The 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 for essentially the last 80 years. It didn’t work. It was decade after decade after decade of excessive optimism followed by disappointment.

When I was in college in the 1980s, there had been a famous AI boom-and-bust cycle in venture and in Silicon Valley. It was tiny by modern standards, but at the time it was a big deal. By the time I got to college in 1989, AI was a backwater field in computer science departments, and everybody assumed it was never going to happen.

But the scientists kept working on it, to their credit. They built up this enormous reservoir of concepts and ideas, and then we all saw what happened with the ChatGPT moment. All of a sudden, it crystallized: “Oh my God, it turns out it works.” That’s the moment we’re in now.

Significantly, that was less than 3 years ago. That was Christmas of 2022. We’re roughly 3 years into what is effectively an 80-year revolution—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.

The great news with this technology is that it’s already ultra-democratized. The best AI in the world is available at ChatGPT, Grok, Gemini, and these other products that you can just use. You can see how they work. The same thing is true for video: You can see state-of-the-art models with Sora and Veo. For music, you can see Suno and Udio, and so forth.

We’re seeing that happen now, and Silicon Valley is responding with this incredible rush of enthusiasm. This gets to the magic of Silicon Valley, which is that Silicon Valley long ago ceased to be a place where people make silicon. That moved out of California not long ago and ultimately out of the United States, although we’re trying to bring it back now.

The great virtue of Silicon Valley over the last 80 years of its existence has been its ability to recycle talent from previous waves of technology into new waves of technology, and then inspire an entire new generation of talent to join the project. Silicon Valley has this recurring pattern of reallocating capital and talent, building enthusiasm, building critical mass, building funding support, building human capital, and building everything needed for each new wave of technology.

That’s what’s happening with AI. I think the biggest thing I could say is that I’m surprised, essentially, on a daily basis by what I’m seeing. We’re in the fortunate position of getting to see it from 2 angles. One is that we track the underlying science and research work very carefully. Every day I see a new AI research paper that completely floors me—some new capability, discovery, or development that I would never have anticipated. I’m just like, “Wow, I can’t believe this is happening.”

On the other side, of course, we see the flow of all the new products and all the new startups. We’re routinely seeing things that, again, leave me with 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, and this is an industry that routinely gets out over its skis and overpromises. 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,” or whatever.

But against that, I would just say the capabilities are truly magical. By the way, I think that’s the experience consumers are having when they use it, and I think that’s the experience businesses are having for the most part when they’re working on their pilots and looking at adoption. Then it translates to the underlying numbers.

We’re seeing this new wave of AI companies growing revenue—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 very compelling products, are growing revenues faster than any way I’ve ever seen before.

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 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, so I think we probably have a long way to go.

Maybe on that topic: one of the big knocks is, yes, the revenue is immense, but the expenses also seem to be keeping pace. What are people missing as part of that discussion?

Start with the core business models. You’re right: this industry basically has 2 core business models—the 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. Sometimes people ask us, “Is AI like the internet revolution?” It’s a little bit, but the thing with the internet was that we had to build the internet. We had to actually build the network, and ultimately that involved enormous amounts of fiber in the ground, enormous numbers of mobile cell towers, and enormous numbers of shipments of smartphones, tablets, and laptops in order to get people on the internet. There was an incredible physical lift to do that.

People forget how long that took. The internet itself is an invention of the 1960s and 1970s. The consumer internet was a new phenomenon in the early 1990s, 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, which is fairly amazing. We didn’t get mobile broadband until around 2010.

People actually forget that the original iPhone came out in 2007. It didn’t have broadband; it was on a narrowband 2G network. It did not have high-speed data or anything resembling high-speed data. It wasn’t really until about 15 years ago that we even had mobile broadband.

The internet was this massive lift, but the internet got built and smartphones proliferated. The point is, now you have 5 billion people on planet Earth who are on some version of mobile broadband internet, with smartphones all over the world selling for as little as $10. You have amazing projects like Jio in India that are bringing the remaining population of the planet that hasn’t been online until now online.

We’re talking about 5 billion or 6 billion people, and the consumer AI products could deploy to all of those people basically as quickly as they want to adopt them. The internet is the carrier wave for AI to proliferate at light speed into the broad base of the global population. That’s a potential rate of proliferation for 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. This is what we’re seeing: the consumer AI killer applications are growing at an incredible rate, and 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 that 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.

I think a lot of companies cap their opportunity by capping their pricing too low, and I think the AI companies are more willing to push that, which is good. I think that’s reason for considerable rational optimism about the scope of consumer revenues we’re going to be talking about here.

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, you can do even the most prosaic things, like raise your customer service scores, increase upsells, reduce churn, or run marketing campaigns more effectively. All of those are directly relevant to AI. These are direct business payoffs that people are seeing already.

If you have the opportunity to infuse AI into new products, 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? Again, you observe it and you’re like, “Wow.” The leading AI infrastructure companies are growing revenues incredibly quickly. The pull is really tremendous. It feels like incredible product-market fit.

The core business model is actually quite interesting. It’s basically tokens by the drink—tokens of intelligence per dollar. By the way, the other fun thing is that if you look at what’s happening with the price of AI, the price of AI is falling much faster than Moore’s law.

I could go through that in great detail, but basically, all of the inputs into AI, on a per-unit basis, are collapsing in cost. As a consequence, there’s this hyperdeflation of per-unit cost, and that’s driving a more-than-corresponding level of demand growth through elasticity.

It feels like we’re just at the very beginning of figuring out exactly how expensive or cheap this stuff is getting. There’s no question that tokens by the drink are going to get a lot cheaper from here. That’s going to drive enormous demand, and everything in the cost structure is going to get optimized.

When people talk about the chips or whatever the unit input costs are for building AI, the laws of supply and 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 that you have a shortage of GPUs, inference chips, data center capacity, 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. There’s going to be an enormous build-out of all of this. There are hundreds of billions, or at this point possibly trillions, of dollars going into the ground in all these things.

And so the per-unit costs of the AI companies are going to drop like a rock over the course of the next decade. The economic questions, of course, are very real, and there are microeconomic questions around all these businesses. But the macro forces, at least here, I think, 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 that people can use this in their lives and businesses, it's really hard for me to see how it wouldn't both grow a lot and generate enormous revenue.

Yeah. Actually, I think it was 2 or 3 weeks ago when AWS was saying that the GPUs they've been using have been able to extend to even 7-plus years. So the shelf life of the GPUs they're using is also extending in ways that they can optimize better than perhaps in the last couple of cycles. Is that the right way to think about it as well?

Yeah, that's right. That's one really important question and observation. By the way, that also gets to this other question where there are different theories, 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 the small-model revolution is happening at the same time. If you track the capability of the leading-edge models over time, what you find is that after 6 or 12 months, there's a small model that's just as capable. You can get these charts from the various research firms, but if you track the capability of the leading-edge models, you see this happen repeatedly.

There's this chase function happening: The capabilities of the big models are basically being shrunk down and provided at a smaller size and therefore a lower cost quite quickly. I'll just give you the most recent example, which came out over the last 2 weeks. Again, this is a thing that's just kind of shocking.

There's a Chinese company—I forget the name of the company—that produces a model called Kimi, one of the leading open-source models out of China. The new version of Kimi is a reasoning model that, at least according to the benchmarks so far, is basically a replication of the reasoning capabilities of GPT-5. These new models—GPT-5 was a big advance over GPT-4—and of course GPT-5 costs a tremendous amount of money to develop and serve. All of a sudden, here we are 6 months later, and you have an open-source model called Kimi K2 Thinking.

I think they had it shrunk down to be able to run on either 1 MacBook or 2 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 whatever GPT-5 cost, or you don't want to have it hosted and want to run it locally, you can do that.

Again, it's just another breakthrough. It's another huge advance. It's another Tuesday. It's like, “Oh, my God.” Then, of course, it's, “All right, well, what is OpenAI going to do?” Obviously, they're going to go to GPT-6, right? There's this kind of layering happening where the entire industry is moving forward. The big models are getting more capable, and the small models are chasing them.

The small models provide a completely different way to deploy these systems at very low price points. 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. Therefore, you're always going to want the most intelligent thing, because why would you ever want something that's not the most intelligent thing for any application?

The counterargument is that there's a huge number of tasks that take place in the economy and in the world that don't require Einstein. A person with a 120 IQ is great. You don't need a 160 IQ PhD in string theory. You just need somebody who's competent and capable, and that's great.

We've talked about this before. I tend to think the AI industry is going to be structured a lot like the computer industry ended up being structured. You're going to have a small handful of basically the equivalent of supercomputers, which are these giant, so-called God models, running in giant data centers.

My working assumption is that you then have this cascade down of smaller models, all the way down to the very small models that run on embedded systems—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. It's what happened with computers, which became microchips, and it's what happened with operating systems and a lot of everything else that we built in software. 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 what we call robust market power. But in time, that draws competition, and that's happening right now.

Nvidia is an absolutely fantastic company. It fully deserves the position it's in and the profits it's generating, but it's now so valuable and 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 in AI chips.

That's already happening. You've got other major companies like AMD coming at Nvidia, and you've got the hyperscalers building their own chips. A bunch of those big tech companies are building their own chips, and of course the Chinese are building their own chips as well. It's pretty likely that in 5 years, AI chips will be cheap and plentiful, at least compared with the situation today. Again, I think that will tend to be extremely positive for the economics of the kinds of companies that we invest in.

Yep. Startups are also starting to go after new chip designs, which is exciting.

Yeah. Well, that's the other thing: You have these disruptive startups. Actually, just for a moment on chips, we were not really big investors in chips because it's kind of a big-company thing.

It's a little bit of historical happenstance that AI is running on quote-unquote GPUs, which stands for graphics processing unit. For people who haven't tracked this, there were basically 2 kinds of chips that made the personal computer happen.

The so-called CPU, or central processing unit, which classically was the Intel x86 chip, is the brain of the computer. Then there was this other kind of chip called the GPU, or graphics processing unit. That was the second chip in every PC that did all the graphics—3D graphics for gaming, CAD/CAM, Photoshop, or anything else that involved lots of visuals.

The canonical architecture for a personal computer was a CPU and a GPU. The same thing was true for smartphones, by the way. Over time, these have kind of merged. A lot of CPUs now have GPU capability built in, and a lot of GPUs now have CPU capability built in. This has gotten fuzzy over time, but that was the classic breakdown.

The fact that this was the classic breakdown meant that, while Intel had a monopoly for a long time on CPUs, there was this other market of GPUs. Nvidia basically fought the GPU wars for 30 years and came out the winner as the best company in the space. It was a hypercompetitive market for graphics processors. It was actually not that high-margin and not that big.

Then it turned out that there were 2 other forms of computation that were incredibly valuable, happened to be massively parallel in how they operate, and happened to be very good fits for the GPU architecture.

Those 2 highly lucrative additional applications were cryptocurrency, starting about 15 years ago, and then AI, starting about 4 years ago. Nvidia very cleverly set itself up with an architecture that works very well for this, but it's also just a little bit of a twist of fate that if AI is the killer app, it turns out that the GPU architecture is the best legacy architecture for 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, I think, would be much more economically efficient.

And, Erik, to your point, there are startups that are building entirely new kinds of chips oriented specifically for AI. We'll have to see what happens there. It's hard to build a new chip company from scratch. It's possible that 1 or more of those startups makes it on their own, and some of them are doing very well. It's also possible, of course, that they get bought by big companies that have the ability to scale them.

We'll see exactly how that unfolds. Of course, we'll also see the Koreans play here for sure. The Japanese are going to play, and the Chinese are going to play in a major way as well. They have their own native chip ecosystem that they're building up.

There are going to be many choices of AI chips in the future, and that's going to be a giant battle that we observe very carefully and that we make sure our companies are able to take full advantage of.

While on the topic of international issues, you mentioned Kimi earlier. It seems like some of the best open-source models today are from China. Should this be worrisome to folks?

How are you thinking and talking about this topic with folks in DC? I know you were just there last week. How much of this is a concern for US companies, particularly having just seen the rise of China do unnatural things in solar markets and car markets? Are they flooding the ecosystem so that they can eventually take share and increasingly own the ecosystem?

A couple of things. You want to start these discussions by saying, “Look, there's vigorous debate in the US and around the world about how much we're in a new Cold War with China and exactly how hostile we should view them.”

It's very tempting, and I think there's a very good case to be made, that we're in a new Cold War that's, in a lot of ways, like the US versus the USSR in the 20th century. The counterargument would be that it's 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 that the USSR never really made anything that anybody else needed, other than weapons. The USSR's primary exports were literally wheat and oil. China, of course, 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 anything—a computer, a smartphone, or whatever—it's got a lot of componentry in it that was made in China. There's a much tighter interlinkage between the American and Chinese economies than there was between the American and Soviet economies.

Maybe Adam Smith or whoever might say that's good news for peace, because both countries need each other. The other part of that argument is that Chinese governance is based on high employment. All the geopolitical people say that if China ended up with 25% or 50% unemployment, that would cause civil unrest, which is the 1 thing the CCP doesn't want.

The corresponding part of the trade pressure is that China needs the American export market. The American consumer is about 1/3 of global consumer demand. China needs the US export market, or a lot of its factories would suddenly go bankrupt, causing mass unemployment and unrest in China.

It's a complicated, intertwined relationship. Having said that, the mood in DC for the last 10 years, on a bipartisan basis, has been that the US needs to take China more seriously as a geopolitical foe.

Under that school of thought, there's the military dimension, which is the risk of some kind of war in the South China Sea and the risk of some kind of war around Taiwan. That has everybody in Washington on high alert. There's also the economic question around the deindustrialization of the US, potential reindustrialization, and what that means about dependence on China.

Then there's this AI question. The AI question is an economic question, but it's also a geopolitical question. AI is essentially only being built in the US and in China. The rest of the world either can't build it or doesn't want to, which we could talk about.

It's basically the US versus China. AI is going to proliferate all over the world, and the question is whether American AI or Chinese AI will proliferate all over the world. Generally, across party lines in DC, that's how they look at it.

The Chinese are in the game for sure, with software. DeepSeek fired the starting gun in the software race. Then you've got, I think, 4 primary models: DeepSeek, which is an AI model from a hedge fund in China; Qwen, the model from Alibaba; and Kimi, from another startup called Moonshot AI. Then there's Tencent, Baidu, and ByteDance, which are all primary companies doing a lot of work in AI.

There are somewhere between 3 and 6 primary AI companies, and then there are tremendous numbers of startups. They're in the race on software. They're working to catch up on chips. They're not there yet, but they're working incredibly hard to catch up.

As an example of that, the common understanding in the US is that the reason you haven't seen the new version of DeepSeek yet is that the Chinese government has instructed them to build it only on Chinese chips, as a motivator to get the Chinese chip ecosystem up and running. The main chip company there is Huawei, although there could be more in the future.

Then there's everything to follow, which is AI in robotic form. There's a global technological and economic robotics competition that's kicking off. China starts out ahead in robotics because it's ahead on so many of the components that go into robots.

The entire supply chain of electromechanical things moved from the US to China 30 years ago and has never come back. That's the DC lens on it, and DC is watching it quite carefully.

The big supernova moment this year was the DeepSeek release. The DeepSeek release was surprising on a number of fronts. One was just how good it was. Along these lines, it took the capability set that was running in large models in the cloud and shrank it into a smaller version of equivalent capabilities that you could run on small amounts of local hardware.

It was also a surprise that it was released as open source, and particularly as open source from China, because China does not have a long history of open source. It was also a surprise that it came from a hedge fund. It didn't come from a big R&D or university research lab, and it didn't come from a big tech company.

It came from a hedge fund and, as far as we can tell, it was this somewhat idiosyncratic situation where you had an incredibly successful quant hedge fund with all these super-geniuses, and the founder of that hedge fund decided to build AI.

At least the external indications are that this was a surprise even to the Chinese government. It's impossible to prove what the Chinese government was or wasn't surprised by, but the atmospherics are that this was not exactly planned. This was not a national champion tech company at the time DeepSeek was released. It came out of left field.

That, by the way, is very encouraging for the field. It was possible for somebody who was unknown to do that kind of thing. That means that maybe you don't need all these super-genius, superstar researchers. Maybe smart kids can just build this stuff, which I think is the direction things are headed.

The success of DeepSeek, and the success of DeepSeek from China as open source, kicked off a sort of trend in China of releasing these open-source models. The cynics in DC would say, “Yeah, they're dumping.” They're obviously dumping.

They’re trying to commoditize it right out of the gate. The Chinese industrial economy does have a history of subsidized production that leads to selling things below cost in some cases. But I think that’s almost too cynical a view because it’s like, all right, wow, they’re really in the race—open source, closed source, whatever. They’re actually really in the race.

We’ve talked in the past, I think, on LP calls about these policy fights that we’ve been having in D.C. for the last 2 years. There was a pretty big push within the U.S. government 2 years ago to basically restrict or outright ban a lot of AI. It’s very easy for a country that is the only game in town to have those conversations.

It’s quite another thing if you’re actually in a footrace with China. I think the policy landscape in D.C. has improved dramatically as a consequence of an awareness now that this is actually a two-horse race, not a one-horse race.

For sure. I’ll jump ahead here to policy and regulation, because the current stance on 50 different sets of AI laws by state seems like a catastrophic way to put us effectively with one of our hands tied behind our back in terms of the AI race. What’s the state of play on that? Are folks recognizing that would be catastrophic for progress and development? Where do most people at least stand on that topic today?

It’s a little bit complicated. I’ll rewind to say that 2 years ago, I was very worried about ruinous federal legislation on AI. We engaged very heavily at that point, which we’ve talked about in the past, and I think the good news is that the risk of that, sitting here today, is very low. There’s very little mood in D.C., on either side of the aisle, to do anything that would prevent us from beating China. On the federal side, things are much better now. There will be issues and tensions in the system, but things are looking pretty good.

That has translated, to your point, into a lot of attention to the states. Basically, under our system of federalism, the states get to pass their own laws on a lot of things. A lot of well-meaning people are trying to figure out what to do at the state level, and then, of course, there’s a lot of opportunism where AI is just the hot topic.

If you’re an aggressive, up-and-coming state legislator and you want to run for governor and then president, you want to attach yourself to the heat. There’s a political motivation to do state-level stuff.

Sitting here today, we’re tracking on the order of 200 bills across the 50 states. Not just the blue states, by the way, but also the red states. For the last 5 years or whatever, I’ve spent a lot of time complaining about what Democratic politicians are threatening to do to attack AI. Republicans are not a bloc on this. There are quite a few local Republican officials in different states who also have misinformed or ill-advised views and are trying to put out bad bills.

It’s a little bit weird that this is happening. The federal government does have authority to regulate interstate commerce, and technology—AI, by definition—is interstate. There’s no AI company that just operates in California or just operates in Colorado or Texas. Of all technologies, AI is obviously national in scope. It’s sort of obvious that the federal government should be the regulator, not the states. The federal government needs to assert itself and step in.

There was an attempt to add a moratorium on state-level AI regulation that would reserve the right of the federal government to regulate AI and prevent the states from moving forward with these bills. That was part of the negotiation for the “One Big Beautiful Bill,” and there was a deal behind that. The deal blew up at the last minute, and the moratorium didn’t happen.

In fairness, the critics of that moratorium were probably right that it was too much of a stretch. It was definitely too much of a stretch to get enough support to pass, but it was also probably too much of a stretch in terms of restricting the states from certain kinds of regulation that they really should be able to do. It just didn’t quite come together.

We’re having very active discussions in D.C. right now about the next turn on that. The administration is very supportive of the idea of the federal government being in charge of this as part of it being an actual 50-state issue and an issue of national importance. Most members of Congress on both sides of the aisle get this. We just have to figure out a way to land it, but I think that’ll happen.

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 in and around Denver and Boulder. A year later, they’re actually trying to reverse their way out of that bill.

A year later, what were some of the nuances of it—like algorithmic discrimination and how to mitigate it—and what were some of the extreme versions of what they had proposed?

The really draconian one that we fought hard was the one in California, which was called SB 1047. It was basically modeled after what was called the EU AI Act—the European Union’s AI Act.

This is the backdrop to all the U.S. stuff: The EU passed this bill called the AI Act about 2 years ago, and it basically killed AI development in Europe to a large extent. It’s so draconian that even big American companies like Apple and Meta are not launching leading-edge AI capabilities in their products in Europe. That’s how draconian that bill was.

It’s a classic European thing. They have this view that, “If we can’t be the leaders in innovation, at least we can be the leaders in regulation.” They literally say this, by the way. Then they pass this incredibly ruinous, self-harming kind of thing. A few years pass, and they’re like, “Oh, my God, what have we done?” They’re going through their own version of that.

When I talk about Europe, I tend to be very dark about the whole thing. The darkest people I know about Europe are the European entrepreneurs who moved to the U.S. They’re absolutely furious about what’s happening in Europe on this stuff.

It’s so bad in Europe—they shot themselves in the foot so badly—that there’s actually a process now at the EU to try to unwind it. They’re trying to unwind the GDPR. For people tracking Europe, Mario Draghi, the former prime minister of Italy, did this thing about a year ago called the Draghi report, which is a report on European competitiveness. He outlined in great detail all the ways that Europe was holding itself back, and part of it was overregulation in areas like AI. They’re trying to reverse out of that, or at least making gestures. We’ll see what happens.

In the middle of all that, California inexplicably decided to copy the EU AI Act and try to apply it to California, which might strike you as completely insane. To which I would say, yes, welcome to California. It was basically this Sacramento political dynamic that got crazy.

It would have completely killed AI development in California. Unfortunately, our governor vetoed it at the last minute. It did pass both houses of the legislature before he vetoed it. To your point, it would have done a whole bunch of ruinously bad things. But one of the things it would have done is assign downstream liability to open-source developers.

And so we talked about this Chinese open-source thing. You’ve got Chinese companies out there with open source. Now you’re going to have American companies that have open-source AI, and you’re also going to have American academics and independent people developing open source in their nights and weekends, which is a key way that all this technology proliferates.

This law would have assigned downstream liability for any misuse of open source to the original developer of the open source. So you’re an independent developer, an academic, or a startup, and you develop and release an AI model. The AI model works fine the day you release it; it’s great. But 5 years later, it gets built into a nuclear power plant, there’s a meltdown at the nuclear power plant, and somebody says, “It’s the fault of the AI.” The legal liability for that nuclear meltdown, or for any other practical, real-world thing that would follow in the out years, would then be assigned back to that open-source developer.

Of course, this is completely insane. It would completely kill open source. It would completely kill startups doing open source. It would completely kill academic research in its entirety—anything in the field. That’s the level of playing with fire that these state-level politicians have become enamored with.

Like I said, I think the good news is the feds understand this. I suspect that this is going to get resolved, but it does need to get resolved because, as a country, it just doesn’t make any sense to let the states operate suicidally like this. That’s what we’re doing. We talk about this—we call this our little tech agenda. We’re extremely focused on the freedom for startups to innovate. We’re not trying to argue many other issues.

We operate in a completely bipartisan fashion. We have extensive support on both sides of the aisle and for both sides of the aisle, so it’s a truly bipartisan effort, very policy-based, and I think very much aligned with the interests of the country broadly. That is what we’re doing.

The other question we get—in some cases from LPs, but in a lot of cases from employees—is, “Okay, why us?” With any sort of policy question like this, there’s always this collective-action question, which is the tragedy of the commons. In theory, everybody—every venture firm, every tech company, whatever—should be weighing in on these things. In practice, what happens is that most of them simply don’t.

At some point, it falls on somebody’s shoulders to fight these things. Ben and I just concluded that the stakes here were way too high. If we’re going to be the industry leader, we just have to take responsibility for our own destiny. For better or for worse, I think that’s the cost of doing business and being the leader in the field right now.

Before we get off the topic of AI, I want to go back to one question that was submitted. Do you think usage-based or utility pricing is the right way to price AI compared to seats?

That is a fantastic question. This is one of these giant questions on my list of what I call the trillion-dollar questions, where, depending on how this is answered, it will drive trillions of dollars of market value.

Usage-based pricing is actually fairly amazing if you think about this from a startup or venture standpoint. It’s fairly amazing what’s happened, and I’m not really talking about this in public because I don’t want it to stop. I think it’s quite amazing. You have these technology companies—these big tech companies—with incredible R&D capabilities building these big AI models, these big AI models with this incredible new kind of intelligence.

Then it turns out that they were already in a war. They were already in the cloud war, right? They were already in the war for cloud services. This is AWS versus Azure versus Google Cloud, along with all these other cloud efforts.

What actually happened was that there’s an alternate universe in which they basically just kept all of their magic AI secret and captive and used it in their own businesses, or used it to compete with more companies in more categories. Instead, they’ve basically—“commoditized” is too strong a word—they’ve proliferated their magic new technology through their cloud businesses.

These businesses have incredible scale components and hypercompetition between the providers, with prices that come down very fast. You’ve got the most magical new technology in the world, and it’s basically being served up by those companies as a cloud business, made 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 it means you can start easily. There’s basically no fixed cost for a startup building an AI app. They don’t have giant fixed costs because they can just tap into the OpenAI, Anthropic, Google, Microsoft, or whatever cloud’s tokens-by-the-drink intelligence offering and get going.

From the startup standpoint, it’s this marvelous thing where the most magical thing in the world is available by the drink. It’s absolutely amazing. That model’s working, and those companies are happy. They’re growing really fast, happily reporting massive cloud-revenue growth, and they’re happy with the margins and so forth. Generally, I think it’s working, and those businesses are likely to get much larger.

That doesn’t mean that the optimal pricing model for all of the applications should be tokens by the drink. In fact, I very much think that’s not the case. We spend a lot of time working with our companies on pricing. We actually have dedicated experts on pricing at our firm because it’s this magical art and science that a lot of companies don’t take seriously enough.

A core principle of pricing is that you don’t want to price by cost if you can avoid it. You want to price by value, right? You want to price so you’re getting a percentage of the business value, especially when you’re selling to businesses. You want to price as a percentage of the business value that you’re creating.

You do have some AI startups that are pricing by the drink for certain things they’re doing, but many others are exploring different pricing models. Some are simply replicating SaaS pricing models, but other companies are exploring pricing models based on the idea that if the AI can do the job of a coder, a doctor, a nurse, a radiologist, a lawyer, a paralegal, or a teacher, can you price by value? Can you get a percentage of the value of what otherwise would have been literally a person?

Equivalently, can you price by marginal productivity? If you can take a human doctor and make them much more productive because you give them AI, can you price as a percentage of the productivity uplift from the augmentation—the symbiotic relationship between the human being and the AI?

What we see in startup land is a lot of experimentation happening with these pricing models, and I think that’s super healthy. I was giving this little speech on this: High prices are really underappreciated. High prices are often a favorite of the customer.

A lot of the naive view on pricing is that the lower the price, the better it is for the customer. The more sophisticated view is that higher prices are often good for the customer because a higher price means that the vendor can make the product better, faster. Companies with higher prices and higher margins can invest more in R&D, and they can make the product better.

Most people who buy things aren’t just looking for the cheapest price. They want something that’s really going to work well. Often, high prices—the customer doesn’t ever say this, and it’ll never show up in a survey—can actually be a gift for the customer because they can make the vendor better, make the product better, and ultimately make the customer better off.

And so, I’m very encouraged by the degree to which the AI entrepreneurs are willing to run these experiments. We’ll have to see where it pans out. But at least so far, I feel good about the attitude of the industry about it.

Awesome. I actually had probably 10 more follow-up questions as you were going through that, but I’m going to go back to a topic you had briefly touched on: the trillion-dollar questions. Will open source or closed source win? It feels like we’ve come out on this debate—or where do you put that?

No, I think this is still open. I think this is still very open. The closed-source models keep getting better.

Generally, if you just take the temperature of the people working at the big labs on the big proprietary models, what they’ll tell you is that progress is continuing at a very rapid pace. There’s this periodic concern that shows up online or in the market: maybe the capabilities of these models are topping out. There are certain areas in which people are working, but the people working at the big labs are like, “Oh, no, we have 800 new ideas. We have tons of new ideas and tons of new ways of doing things.”

“We might need to find new ways to scale, but we have a lot of ideas on how to do that. We know a lot of ways to make these things better, and we’re basically making new discoveries all the time.” So I would say generally the people working across all the big labs are pretty optimistic. I think the big models are going to continue to get better very quickly here. Overall, the open-source models continue to get better.

And like I said, every month or something, there’s another big release of something like this Kimi thing. It’s just like, “Wow, that’s amazing.” They really shrunk that down and got that capability on a very small form factor.

And maybe the third thing to bring up is that the other really nice benefit of open source is that it’s easy to learn from. If you’re a computer science professor who wants to teach a class on AI, a computer science student trying to learn about it, or just a normal engineer in a normal company trying to learn this new thing—or somebody in their basement at night with a startup idea—the existence of these state-of-the-art open-source models is amazing, because that’s the education you need. These open-source models actually show you how to do everything.

And what that’s leading to is that the knowledge about how to build AI is expanding very fast, again as compared to a counterfactual world in which it was all basically bottled up in 2 or 3 big companies. The open-source thing is also just proliferating knowledge, and that knowledge is generating a lot of new people.

As you guys have all seen sitting here today, AI researchers are at an enormous premium. AI researchers today are getting paid more than professional athletes. That’s a supply-and-demand imbalance: There aren’t enough of them to go around. But, again, shortages create gluts.

The number of smart people in the world who are coming up to speed very quickly on how to build these things is growing. Some of the best AI people in the world are 22, 23, or 24. By definition, they haven’t been in the field that long; they can’t have been experts their whole lives. They have to have come up to speed over the course of the last 4 or 5 years, and if they’ve been able to do that, then there are going to be a lot more in the future who are going to do that. The spread of the level of expertise on this technology is happening very quickly now.

So, yeah, I think it’s still, as I said, still a race. And, by the way, the long-term answer may well just be both. If you believe my pyramid industry structure, then there will certainly be a large business of whatever is the smartest thing, almost regardless of how much it costs. But there will also be this giant volume market of smaller models everywhere, which is what we’re also seeing.

Yep. The other question you had posed at that point in time was whether incumbents or startups would win. At that point in time, I think there was a mixed bag in terms of how incumbents were approaching AI. I think that’s radically changed in the last 2 years. On the counterexample, with the blossoming of startups—many of them increasingly migrating into the incumbent category—how big have they become since that time? Do you want to take that question and give your assessment of the state of the world?

Yeah. So, look, big companies are definitely playing hard. Google is playing hard. Meta is playing hard. Amazon and Microsoft are playing hard. There are a bunch of these companies that are kind of in there very aggressively. And then you’ve got what we call the new incumbents, like Anthropic and OpenAI.

But even in the last 2 years, you’ve had the birth of brand-new companies that are almost instant incumbents. You could say xAI is one of those. Mistral, by the way, is the great outlier to my Europe thing from earlier. Mistral is actually doing very well as the sort of European, French-national, continental AI champion—the exception that proves the rule.

There are a bunch of these now that are doing quite well and are becoming new incumbents. And then, of course, there are tons of startups. There are actual foundation-model startups.

We funded Ilya Sutskever out of OpenAI to start a new foundation-model company. We funded Mira Murati, also out of OpenAI. We funded Fei-Fei Li out of Stanford to build a world-model company. There are new swings—all early, but very promising—to build new incumbents quickly.

That’s all happening. On top of that, there’s just this giant explosion of AI application companies. These are basically companies—usually startups—that take the technology and field it in a specific domain, whether that’s law, medicine, education, creativity, or whatever.

But it’s amazing how sophisticated things are getting very quickly. Let’s talk about the application companies for a moment. A classic example of an application company is Cursor.

They take the core AI capability, which they purchase by the drink from Anthropic, OpenAI, or Google—in other words, tokens by the drink—and then build a code editor, what we used to call an IDE, or integrated development environment—in other words, a software-creation system. So they build an AI coding system on top of Anthropic, OpenAI, Google, or whatever big models they use, and field that.

The critique of those companies in the industry has been that they’re what are called “GPT wrappers,” which is kind of the pejorative. The idea is that they’re not actually doing anything that’s going to preserve value because the whole point of what they’re doing is surfacing AI, but it’s not their AI. The AI being surfaced is from somebody else, so these are pass-through shell things that ultimately won’t have value.

It turns out what’s happening is kind of the opposite. The leading AI application companies, like Cursor—first of all, what they’re discovering is that they’re not just using a single AI model. As these products get more sophisticated, they end up using many different kinds of models, custom-tailored to the specific aspects of how the products work.

They may start out using 1 model, but they end up using a dozen models, and in the fullness of time, it might be 50 or 100 different models for different aspects of the product. And then, second, 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, by the way, with open source, they’re also able to pick up and run open-source models.

And so, 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, building their own AI. Some of them are also able to pick up and run open-source models.

Small models, though, right, Marc? When you think about big models versus small models, as you were describing that, would that be small? Would you categorize that as small?

Well, some of them—I will let them announce whatever they're doing whenever it's appropriate—but some of them are now also doing big-model development. Again, this is part of the learning just in the last 2 years.

Here's a big learning from the last 2 years, which is very interesting: 2 years ago, or 3 years ago for sure, 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.” They came out of OpenAI, so they had all the secrets and knew how to do it. 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. xAI is maybe the best example of that. xAI—Elon's company—is the company name, and Grok is the consumer product version of it. xAI basically caught up to the state-of-the-art OpenAI and Anthropic level in less than 12 months from a standing start.

Again, that argues against any kind of permanent lead by any one incumbent that's just going to be able to lock the entire market down. If you can catch up like that, even with far fewer resources, that changes the picture.

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 4 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 possible, it seems not to be that hard for other people to catch up, even people with far fewer resources. I don't know what that does. Maybe it makes you slightly more skeptical about the long-run economics of the big players. On the other hand, maybe it makes you more bullish about the startup ecosystem.

It should certainly 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 probably make you more excited about China. On the other hand, Chinese competition putting pressure on the American system not to 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. I should say this because 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. The strategy has to be logical and coherent, or the company kind of collapses into chaos. Companies need to answer these questions, and if they get the answers wrong, they're really in trouble.

Venture has our issues, but a huge advantage that we have is that we don't have to choose. We can bet on multiple strategies at the same time. We are betting on big models and small models, pretrained models and open-source models, foundation models and applications, and consumer and enterprise.

The nature of the portfolio approach is that we are aggressively investing behind every strategy that we've identified as having a plausible chance of working, even when that strategy is contradictory to another strategy that we're investing in. One reason is that the world is messy and probably a bunch of things are going to work, so there aren't going to be clean yes-or-no answers to a lot of this. A lot of the answers are just going to be “and.”

The other reason is that 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. We're going to have multiple ways to win.

That's the goal. That's the theory of why we're taking the approach we're taking in this space. 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.

It's a good segue to a16z questions, because we've gotten a few in so far, and we had a few that were sent in ahead as well. I'll start with a broad topic: What is something you and Ben disagree and commit on?

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

Where the romance is dead.

The romance is long dead. Yes, yes, yes, yes. The fire has long since gone out. We're in the park squabbling all the time.

We debate everything. We argue about everything. That said, one of the things that's made our partnership work is that we do tend to come to the same conclusion. Each of us is open to being persuaded by the other one, so we end up coming to the same conclusion most of the time.

I would say there aren't any—specifically, sitting here today, there are zero issues where I'm sitting here thinking, “I can't believe I'm putting up with this crazy thing on his part that he's doing, that I really disagree with, but I feel like I have to commit to.” I don't think it's the case for him, either.

Quite honestly, the biggest thing that he and I discuss—this is not the most important thing we're doing, but it is a topic since somebody asked the question—is basically the public footprint of the company: our presence in the world in terms of public statements, controversy, and how we vocalize and express our views on things.

There's a real tension there. It's maybe obvious, but it's a very important tension. Generally speaking, the more out there we are, the more outspoken we are, and the more controversial we are, the better for the business, in the sense that entrepreneurs love it. The founders want to work with people who are brave, controversial, and willing to take controversial stands and articulate things clearly.

They want that for a bunch of reasons. One is that it's a demonstration of courage, which they appreciate. The other is that it teaches them who we are before they even meet us. That has proven to be an incredible competitive advantage.

Long-term LPs will know this is why we started with a very active marketing strategy from the very beginning, and it completely worked. The whole thing was that if we're able to broadcast our message and be very clear about what we believe, even to the point where it's controversial, the best founders in the world are going to understand us before they even walk in the door.

They're going to know us even before they've met us, as opposed to everybody else in venture—at least at the time—which was basically just keeping everything quiet. The founder had no idea who these people were or what they believed. That worked incredibly well, and it continues to work incredibly well.

It's generally true across the industry. It's generally the case. On the other hand, there are externalities to being publicly visible and controversial on many fronts. We are trying very hard to thread this needle. We're not backing off from generally being a company that does a lot of outbound.

Erik Torenberg and the team that he’s built, who we’ve talked to you guys about in the past, are already off to the races. We’re tripling down on the idea of being the leaders in articulating the tech and business issues that matter—the issues that people need to be able to understand. That’s proven to be very effective.

A fair amount of our comms are actually aimed at Washington. If you’re a policymaker in Washington and you’re sitting there 3,000 miles away, and your entire information source is East Coast newspapers that hate Silicon Valley, that’s bad.

Our ability to broadcast informed points of view on technology is important. We meet people in D.C. all the time who say, “Most of what I know about this topic I learned from you guys because I listen to the podcast, I read the articles, and I watch the YouTube channel.” We’re going to continue to do that. Overall, we’re on our front foot on that stuff.

On the other hand, Ben and I do go back and forth a bit on exactly how many third-rail topics we should touch, and how frequently. I would say we are trying to moderate that.

As Elizabeth Taylor said, as long as they spell our name right, it can oftentimes be good in most scenarios, particularly when it comes to Little Tech.

I also think embedded in that question is probably some degree of the relationship that you and Ben have, which is now going on 30-plus years. So much so that Marc has become one person representing both. Some people refer to Marc as Andreessen Horowitz now—Marc and Ben have combined into one person.

Yes.

That’s the result of 30-plus years working together. So, it’s been 2 years since you reorganized around AI and launched American Dynamism. What do you think you got most right? In hindsight, is there anything that you underestimated or missed in that decision-making process?

No, I mean, look, we made plenty of mistakes. I think those were the right calls.

The whole theory of venture that we’ve had from the beginning, and that many people before us have had as well, is that the money in venture is made when there’s a fundamental architecture shift, when there’s a fundamental change in the technology landscape. That’s been true for venture capital basically forever.

The reason is that if you have a fundamental change in technology, then you have this period of creativity in which very aggressive people can start new companies. They have this shot to come in and win categories before big companies can respond. If there’s no fundamental change in technology, it’s very hard to make startups work because the big companies just end up doing everything.

Venture lives or dies on the basis of these waves, these transitions. I think the best venture capital firms in history are the ones that were the most aggressive at being able to navigate from wave to wave.

I was a beneficiary of this when I came to Silicon Valley in 1994. There was no venture firm in 1994 that was the internet venture capital firm. It just didn’t exist. But there were a set of venture capital firms at the time, including our firm, Kleiner Perkins, that said, “This is a new architecture. This is a new technology change. It seems totally crazy. Everybody says you can’t make money on it. Whatever, whatever—these kids are nuts. But we’re going to make those bets.”

They were willing to invest. Kleiner Perkins in the 1990s invested not only in us, but also in Amazon and then Google, and in company after company after company. They invested in @Home, which basically made home broadband work. They invested in a fleet of companies.

They were a venture capital firm that had started in the 1970s, really around what was at the time called minicomputers, which was three generations of technology back. They had navigated from wave to wave. The same thing is true for Sequoia, and the same thing is true for basically any successful venture firm that’s been in business for 30, 40, or 50 years.

In this business, of all businesses, you need to get onto the new thing. It was pretty amazing that most of the venture ecosystem just decided to sit crypto out. The number of VCs that we talked to between, call it, the release of the Bitcoin white paper in 2009 and the beginning of the crypto bull run in 2021 who basically said, “We’re not going to do crypto,” was fairly amazing.

I never quite know what to do with the VC who says, “There’s a new wave of technology, and I’m very deliberately not going to participate in it.” I’m always like, “Is that not the job?” I was fairly amazed by the VCs that didn’t make the jump to crypto.

They looked briefly smart during the crypto wars, I would say, of the last 3 or 4 years, and I think they probably look a little bit less smart now. AI is another one of these areas where there are certain firms that are jumping all over it, and there are certain firms that are just sitting back and letting it happen.

There were also certain firms that never made it to the internet. There were firms that were very well known and very successful in the 1980s that just did not make the jump to the internet and basically petered out. In this business, of all businesses, you have to jump on the new wave.

I think we got the magnitude of it right—that this was a fundamental transformation inside the firm. American Dynamism is doing great. American Dynamism itself is also a beneficiary of AI in 2 ways. A lot of the kinds of products that American Dynamism companies build themselves benefit from AI, and AI is also a driver of demand in other sectors of American Dynamism, like energy and materials.

Crypto is back to being an exciting industry as a consequence of all the policy changes. I think there are going to be quite a few intersections between AI and crypto. Biotech and healthcare are also obviously going to be transformed by AI, both on the healthcare side and on the actual drug-discovery side, and that’s underway.

The individual efforts in the firm feel good and suitable for the time. The interactions between the teams and the hybrid ideas—the companies that are coming at these things from multiple angles—feel really good.

The corollary question is, what do we feel like we’re missing right now? I don’t think we’re missing a vertical. As of right now, there’s not a specific vertical where we think, “We just need the equivalent of a new unit or the equivalent of a new fund.” I don’t see that at the moment.

I think it’s more about executing extremely well in the verticals that we have in front of us, and being the best possible partner to the portfolio companies.

Actually, on the point of American Dynamism, there’s a lot of talk about AI taking jobs. Ironically enough, the jobs in American Dynamism sectors have never been more in demand in the physical world, related to energy and, obviously, data-center buildout. The pendulum, it seems, is also swinging from an accelerant standpoint, from a societal point of view.

You talked about the importance of society also needing to be ready for tech adoption. Have you seen that accelerate recently? What’s your sentiment on how to actually increase that, to also make sure the convergence of adoption falls in line with how quickly tech is being implemented?

We’ve talked about this before, but for a very long time, tech was just not very relevant. If you go back over 300 years, there are recurring waves of total panic and freak-out caused by new technology.

Or even go back 500 years, to the printing press, which basically was hand in hand with the creation of Protestantism, which really changed things.

And then, you go back to—you know, there were always continuous panics. There have been multiple waves of automation panics for the last 200 years. A lot of the foundational panic under Marxism was basically a fear of the elimination of jobs through the application of automation.

A lot of the same arguments you hear today about how AI is going to centralize all the wealth in the hands of a handful of people, and everybody else is going to be poor and immiserated—that’s basically what Marx used to say. I think that was wrong then and is wrong now, which we can talk about.

Even in the 1960s, there was this whole panic around AI replacing all the jobs. There was this great—it’s long forgotten, but it was a big deal at the time during the Johnson administration. You read these AI pause letters today, like this one that just came out a few weeks ago that Prince Harry headlined, of all people, talking about how AI is going to ruin everything.

In 1964, there was basically a group of the leading lights in academia, science, and public affairs. There was this thing called the Ad Hoc Committee on the Triple Revolution. If you do a Google search on “Ad Hoc Committee on the Triple Revolution Johnson White House,” or whatever, this thing will pop up. It was a very similar kind of manifesto: We need to stop the march of technology today, or we’re going to ruin everything.

Even in the course of the last 20 years, there was a big panic around outsourcing in the 2000s taking all the jobs. Then it was robots, weirdly enough, in the 2010s—which is amazing, because robots didn’t even work in the 2010s, and they kind of still don’t. There was a panic around that, and now there’s this level of AI panic.

I would just say that the way I would describe it is that we in Silicon Valley have always wanted the work that we do to matter. We spend most of our time, quite honestly, with people telling us that everything we’re doing is stupid and won’t work. That’s the default position.

Then, basically, that flips at some point into panic about how it’s going to ruin everything. It’s easy, sitting out here, to be cynical about that, especially when you see the patterns over time. My view is that we need to be very respectful of that, and we need to be very aware of it.

I use the metaphor of the dog that caught the bus. We always wanted to work on things that matter, and we are working on things that matter. People in the rest of society actually really do care about these things, and it’s our responsibility to think it all through very carefully and to do a good job—not just building the technology, but also explaining it.

I think we have a real obligation to really explain ourselves and engage on these issues. In terms of how to measure how things are going, it’s the classic social science question. If you want to understand patterns of what people are doing and thinking, there are basically 2 ways to understand it: one is to ask them, and the other is to watch them.

Every social scientist, every sociologist, will tell you this. You can ask people, and the way you do that is through surveys, focus groups, and polls—what they think. But then you can watch them and do what’s called revealed preferences. You simply observe behavior.

What you often see in many areas of human activity, including politics and many different aspects of society and culture over time, is that the answers you get when you ask people are very different from the answers you get when you watch them.

You could have a bunch of theories as to why this is. The Marxists claim that people have false consciousness. The explanation I believe is simply that people have opinions on all kinds of things, particularly when they’re in a context where they get to express themselves, and they have a tendency to express themselves in very heated ways.

Then, if you just watch their behavior, they’re often a lot calmer, more measured, and more rational in what they do. That’s playing out in AI right now. If you run a survey or a poll of what, for example, American voters think about AI, they’re all in a total panic: “Oh, my God, this is terrible. This is awful. It’s going to kill all the jobs. It’s going to ruin everything.”

But if you watch their revealed preferences, they’re all using AI. They’re downloading the apps. They’re using ChatGPT in their jobs. You see this online all the time now: “I’m having an argument with my boyfriend or girlfriend. I don’t understand what’s happening. I took the text exchange, cut and pasted it into ChatGPT, and had ChatGPT explain to me what my partner is thinking and tell me how I should answer so that he or she isn’t mad at me anymore.”

Or, “I have a skin condition, and the doctors…” I take a photo, and I’m finally learning about my own health. Or I use it in my job: “I had to get this report ready for Monday morning, and I ran out of time. ChatGPT really saved my bacon.”

People in their daily lives are—you just look at the data. They are not only using this technology; they love this technology. They love it, and they’re adopting it as fast as they possibly can.

I tend to think the public discussion of this is going to ping-pong back and forth for a while, because there is this divergence between what people are saying and what people are doing. But I do think that what people are doing is ultimately the part that wins.

I think this technology is going to be exactly the same as every other one. What’s going to happen here is that it’s going to proliferate really broadly, freak everybody out, and then, 20 years from now, everybody’s going to be like, “Oh, thank God we’ve got it. Wouldn’t life be miserable if we didn’t have this?” Or maybe it will be 5 years from now, or 1 year from now, when people reach that conclusion.

I’m very optimistic about where this lands. It’s just that there will be turbulence along the way.

I’m smiling because I also witnessed that in the wild. Literally late last week, I was on a plane, and the guy next to me was talking to his ChatGPT. I could see him, and he was like, “Help me draft an escalation letter to United for the delay on this flight.”

I was like, “Sir, you are on the flight right now. At least wait until it’s over.” It was very good, though. I’m sure he had a great email crafted as part of that.

I’m going to switch gears to a few fun questions that were sent in, intended to be a lightning round. What is something you’ve changed your mind on recently? Bonus points if it was someone younger than you.

It’s like every day. It’s just a constant experience. It’s almost all about what’s in the realm of the possible.

I’m terrible at specific examples, so I don’t have one ready at hand. But, like I said, it’s always something. It’s often somebody showing up. It’s either something somebody writes or something somebody says, and it’s very frequently somebody who’s very young.

I would say it’s a routine experience.

Good way to stay young. Do you plan to be cryogenically frozen?

Not with current cryonics technology. The track record of that is not great, and the stories are somewhat horrifying. But we’ll see.

We’ll see. You’ve still got some time.

How do you stay grounded when your influence itself may distort reality around you?

The good news on several fronts is that the concern is real. It’s hard for me to talk about with my Midwestern—Midwesterners are either very humble, or we’re really good at faking it—but it’s hard to talk about and requires some introspection.

The reality-warping effect is definitely real. By the way, there is a very big advantage to the reality-warping effect, which is being able to get people to do what you want them to do. There is another side to it.

But it is a concern in terms of having an accurate understanding of what’s happening. I guess I would say 2 things. One is that my partners, including Ben, are quite forthright in telling me when I’m wrong. More generally, we are very exposed to reality.

And so, again, you mentioned that it’s a way to stay younger, make sure their hair never grows back, or whatever. We run these experiments because we make decisions about whether to invest or not invest, and we work with these companies and all their things, and reality kicks in quickly. The delusions don’t last very long in this business because these things either work or they don’t.

You have these long, elaborate discussions about theories on this, that, and the other thing, and then reality completely smacks you square in the face: “You idiot.” This is the ultimate frustration of the business, which is also very motivating: the number of times you think you’ve applied superior analysis, and then you’ve either invested or not invested based on that analysis, and it turns out the analysis was just completely wrong. You completely overrated your ability to epistemically analyze these things. You basically inflicted harm.

The question is always: Is any activity that we do value-add, or is it actually value-subtract? I think in this business, of all businesses, it’s kind of like that, and that applies to all of my own contributions as well. And then I would say maybe the final thing is just that I do have the entire internet ready to tell me that I’m an idiot, so that also doesn’t hurt. It does on a regular basis.

On the point you alluded to earlier about decisions on investing in companies, my favorite line—I think it was from the Cheeky Pint interview that you did—was: When you invest in a company and it doesn’t go well, at least it goes bankrupt. If it does well, and it does fantastically well, you hear about it every single day for the rest of your life.

Yeah. For the next 30 years, reality smacks you in the face, saying, “You fool.”

You had it. It’s literally—you had it in your office. All you had to do was say yes.

And by the way, this is the thing: These are the stories that VCs tell each other. Every great VC basically has this history: “My God, I had it. It was in my office. The thing was in my office, and I said no. If I had just said yes…”

So, yes, the constant reminders in The Wall Street Journal and on CNBC every day that you made a giant mistake are very good for the old humility factor.

Yeah, very humbling. It helps you stay grounded all the time. Last question: Do you plan to go to Mars if and when that opportunity presents itself?

Probably not.

My subliminal Zoom background wasn’t sending the positive vibes.

Well, I’m not even willing to leave California. I’m barely willing to leave my house. Maybe by VR.

Yeah.

And then we’ll see what happens. Having said that, I think Elon’s going to pull it off. I don’t know. I don’t want to predict—this is not a prediction—but I would not be surprised if, within a decade, there are routine trips back and forth. So, yeah, this may actually become a practical question. And, by the way, I do know a lot of people who are probably going to go.

Myself included. Put me on that.

Oh, fantastic.

The flights around the world have prepared me for the 6-month journey to Mars, so I will be just fine.