The New Rule for Picking AI Winners | The a16z Show

Anthropic and OpenAI are adding more revenue per month than Meta, Google, or Microsoft. And I wouldn't be surprised if the combination of those 2 companies is doing $200 billion in revenue run rate.

Between 2020 and 2024, a top 1% exit started at $10 billion. We updated those numbers in February this year: $20 billion.

Yeah.

We've grown 10x over the space of 24 months.

As models get really good and the products built around them get really good, you see this takeoff in usage happening. I thought we had an AI bubble. I feel pretty confident saying that we're not in a bubble right now.

I can't think of a time in my career when I have changed my mind about things at a faster clip, which is good but also humbling, right?

Two big areas are scale and value capture. On the scale side, the world changed in November as it relates to our business, and I think productivity in the workforce. The way that we thought about much of the AI work that was happening before that was a nebulous promise in the enterprise, but we were probably contextualizing it around things like the cloud in software companies and productivity enhancement.

On the consumer side, you could think about AI companies as a consumer business: how many users they have, what the price is, and how big that can get. By the way, I think that's going to be much bigger than people expect, too, which we could talk about. But as of November, I think all of our priors shifted around what is actually going to happen in the enterprise.

Just to contextualize what's happened since then, Anthropic and OpenAI are adding more revenue per month than Meta, Google, or Microsoft. They are already at that scale of revenue being added, and actual diffusion of this technology into the real economy is tiny. It's less than 5%.

Yeah.

Now, within coding and tech-forward companies, it's much more advanced. But as it relates to every other function in the enterprise, full utilization of the capabilities is nowhere right now. So, if you pair that with the fact that they're already getting bigger in terms of revenue added than the hyperscalers, and you're at less than 5% diffusion into the economy, I think the outcomes are going to be extraordinary.

The thing that we've started to try to look at to gauge what can possibly happen—what's the upper bound—is that enterprises are going to have to pay for this somehow.

Yeah.

If you look at the Fortune 500 or the S&P 500, they're actually pretty close. They generate roughly $2 trillion of profit per year collectively.

And I wouldn't be surprised if the combination of those 2 companies is doing $200 billion in revenue run rate by the end of this year.

Yeah.

Not to mention people using open source and other vendors, so you can add even more on top of that. We're already talking about roughly 10% of profit into the Fortune 500. I think the upper bound is going to be where the dollars come from to buy this stuff.

One of the implications of this is that we had all these theories about why open source and local were going to be really important. It turns out that cost is going to hit us in the face and make them really important sooner than we thought. We've updated our priors to get really pilled on this outcome, on the size of the prize and the scale. You can see the early signs of it in the numbers.

But basically, there's almost no diffusion into the real economy. It's going to get great for all these other functions. By the way, what's happened in coding, you can start to see in some other white-collar jobs. It's starting to happen in legal. The legal space is much smaller than coding, obviously, but when the models get really good and the products built around them get really good, you see this takeoff in usage happening. I think it's going to happen in a bunch of different functions in organizations and verticals over the next 12 months.

How much of that do you think is going to be native AI applications? I always go back to Chris Dixon's point that, for the first 3 or 4 years, you see these skeuomorphic applications come in. We've seen that at the minute: most people are using AI to do their existing job in a way that's more efficient, faster, and cheaper. But we're starting to see some of the native applications come in, particularly around generative AI. How do you think that alters the landscape?

I think the big thing that's going to change in enterprise is that we're kind of nowhere on how companies are run differently today. What's happening with some of the layoffs we're seeing is trimming of previous fat. I don't think it's actually efficiency gains.

By the way, there's a really interesting thing happening inside these companies: most of the resource allocation, at least for really good companies, is actually on product and new things, as opposed to automating the way they're run. They only have so many resources, and the best ones know that the size of the prize for getting something right on the product side is enormous. The best people at those companies—the best engineers—want to work on that side of things. The size of that prize, and the fact that the best people are going to work on it, are driving where most of the work is happening.

The more mature companies would probably be better suited to automating the way their business is done internally, but they're slower adopters. There's a latent opportunity in our portfolio companies to drive efficiency gains and stuff, but it's not the best people working on it, and it's not where the incremental dollar is going to go just yet.

The most cutting-edge folks inside those companies who are trying to do this, and whom I've talked to, are in the documentation phase. That's just: turn everything into Markdown files, capture as much context as you possibly can, and then see where you can still manage your business appropriately, not make sacrifices on customer experiences, but drive efficiency. We're very, very, very early in that.

The native AI companies run themselves totally differently. The founders are just built different. One of the things that we've observed about the previous generation of founders, if you look at SaaS companies, for example—I’ve written about this—is that we didn't realize how inefficiently they were running until much later.

More quickly they could grow as well.

Yeah, or how much more quickly they could grow. And by the way, it turns out that the magnitude of their market is so small compared to what we've seen in the models. The model companies are adding more than the entire public software universe in terms of revenue added, combined.

They're not particularly tightly run, but they had great business models. They could grow and do well, and everyone had a mandate to buy more software, headcount grew, and everything worked out. The new companies are very lean, very aggressive, and they work all the time.

Mhm.

And so, it's fun to see the most cutting-edge companies when you go in. All their researchers are sitting there, whispering in, and running swarms of agents. David George

Yeah, so they're not typing agents.

They're not even typing. They're efficient, and I think that's kind of going to be the future. It's just really early.

I think the skeuomorphic phase is, I would say, everything that is reactive today. I think there's going to be a shift to proactive engagement, both in consumer and in enterprise. We're starting to see it in some of the cutting-edge, early-stage companies that we're working with, but it's really, really early.

Yeah. When I think of our prior from 10 to 12 months ago, there are a couple of things that I think have changed. One has been reinforced: we always thought that the largest companies were going to continue to be an order of magnitude larger than we'd seen in prior cycles.

Yes.

And if anything, that's accelerating. We've put out some data around the size of a top 1% exit doubling every 5 years or so. Between 2020 and 2024, a top 1% exit started at $10 billion. We updated those numbers in February this year. A top 1% exit for 2025 in the first 2 months of 2026 was then $20 billion. We just updated them yesterday.

If you look at just the exits that have closed, it's now at $32 billion. So, where is the threshold for the top 1%? And then if you think about OpenAI and Anthropic coming in, potentially we could be north of $100 billion by September.

It's incredible.

Which is just—so we've 10x-ed—

Yeah.

—over [laughter] the space of 24 months what a top 1% exit looks like.

Yeah. I mean, just the combination of those large companies, I think, is larger than the entire Russell 2000, if I'm not mistaken.

Yeah.

And so, the magnitude of these companies has just grown so great. And look, we've built our firm in response to that. We believe the next generations of companies that get bigger as new trends happen are going to be bigger than their predecessors.

We actually did a similar analysis where we looked at all of the VC-backed IPOs that happened over the last 6 years. If you sum all of them up, they're a little over $1 trillion. That's probably going to be smaller than any one of the 3 large IPOs that we expect to happen. So, I'd say the observation is the outcomes keep getting bigger, but it's happening much faster. The pace of value creation is getting faster.

Particularly something like Wiz and Cursor, you'd kind of like 4, 5, 6 years to get from nothing to, well, $30 billion and then potentially $60 billion.

I would say, similarly, there's a lot that we talk about all the time about deployment pace and how big our funds are and things like that. If you extrapolate out and say, "Hey, previous trends are kind of 10x smaller and the outcomes get much bigger." And by the way, there's a tremendous amount of concentration in the companies that are the winners. Now we believe it's a great time to be in the market investing.

Yeah.

The ChatGPT moment is, I think, less than 4 years ago. So, we're just now seeing some of the most interesting things happening on the back of the foundational technology.

Yeah.

We could have a long talk about who captures it, which is another thing where priors change all the time. But we believe now is the moment where the companies are getting created that are going to be the generational companies of the next 10 years.

Yeah. So, the other thing where my priors have shifted a little bit as well is just around the speed of change and what happens to the defensibility of the leading companies. Because we've seen in prior generations that it's not necessarily the first movers that ultimately captured the economic value of a market. Think Google wasn't the first search engine, and Facebook wasn't the first social media site.

One of the things we track is that every year Forbes comes out with its AI 50 startups list. What was really interesting was that from last year to this year, 40% of the companies that were on that list last year dropped off. Wow.

So, the half-life of these companies feels incredibly short.

Yes.

So, I think where our priors have evolved a bit is, yes, we think the outcomes are going to be much larger, but trying to predict who captures that feels like it's getting much harder.

Yeah.

Is that something that you guys are seeing internally in your portfolios?

Yeah, it is getting much harder because the shift in the technology has happened so much faster. We always talk with our founders about the shifting sands underneath you. That is very, very true. And our priors have been updated a ton about where value is going to be captured.

Yeah.

When we first invested in OpenAI, as you know, it was before ChatGPT. There were moments in the early days where we said, “Model companies are going to be everything. There's never going to be any more application companies. They're all going to go away.”

Then we went through a cycle where we said there's going to be application companies for everything, and the model companies are just going to be APIs. And now we're back in this moment where the model companies are leveraging their way up into the application. This is their biggest way to drive stickiness.

As it relates to assessing something's place in the world, first of all, right now, you have to be in the token path. That is the number one thing that we're looking for in our companies.

Yeah.

And the reason that's so important is what I had said earlier. There's actually cost pressure happening among buyers of technology already. It's happened very fast. They're not going to be increasing their budget for things that are previous-generation software. In fact, they can't even cover the growth in their costs that is happening from AI with reductions in that. There's going to be pressure on those.

Honestly, it's probably going to have to come from either higher prices that they can charge or restructuring of the labor force. The biggest driver of where value is going to get captured right now is, I would say, something that is totally unknowable, which is: What is the market structure of the model companies? How much competition is there?

If there's a couple at the frontier, token prices will probably be higher. If there are 5 at the frontier, token prices will probably be lower. Token prices being lower probably is better for the overall economy because there's not this pressure to restructure the labor force as quickly as things get really, really big.

Yeah.

Right now, the number is smaller. It's not 5. There's a tremendous amount of inelasticity for frontier intelligence right now. There's also the question of how much that changes over time. Can a lot of the jobs be done fine with previous generations of models? That's not the way anyone is consuming tokens today.

That's an unknowable. The market structure is an unknowable. What role does open source play? That's a tenuous situation. How much can you run locally? How much can you run with small models? These are all the open questions that I think will determine who captures value. But for the broader ecosystem to thrive, it's probably competition that keeps token prices lower.

So, a couple of my colleagues are in China at the minute, and it's been really interesting just getting their feedback relative to what we're seeing in the US. One of the things that they were saying was that the leading LLMs in China are probably 6 months behind where we are in the US in terms of the capability of their models, but they're 10x cheaper.

Yes.

So, one of the unknowns, I think, at the minute is—what percentage of the market will those types of companies capture? How much of what we end up doing over the next decade will need to be done through the very frontier models, and what can be captured by that next level down?

It's the classic innovator's dilemma, isn't it? You get the next-generation product that can do 80% of what the frontier product can do but at 10% of the cost. And over time those capabilities extend, and it's harder to be at that frontier.

Yeah. As of right now, we've been surprised at how voracious the appetite is for the absolute frontier. That's probably partially because we're not in the optimization phase yet, but the optimization phase is probably going to happen sooner than we would have expected, is my sense.

There's all these other open questions about the future of open source. How capable are these players of distilling the big models? The big model companies don't want their models distilled.

Yeah.

And so, it probably costs in the order of 2% of the actual training cost—the pre-training cost—of a model to distill it. If that continues to hold and be possible, that probably bodes well for open source. If not, it probably doesn't bode well for open source.

As of right now, you're exactly right. The per-token cost for like-for-like is going down more than 10x year over year. But the appetite for tokens on the frontier is massively exceeding that in terms of dollars.

Yeah. Yeah. How do you factor that in when you're then thinking about valuations of these companies? Because I think one of the concerns that I would have is a bit like in 2021. I thought 2021 was kind of peak emerging manager because a lot of these managers had done the seed rounds, established firms were coming in and writing things up 6 months after the seed round had been done, and there was basically a 0 loss ratio.

Yeah.

And we know that's not how venture works. It feels like we're in a little bit of that situation today, but with the more established firms, because it's the established firms that have been, by and large, capturing the early breakouts in the AI space.

Mhm.

But when I look at it historically, when we look at our early-stage funds, there's a 60% loss ratio. So, 60% of deals don't return the capital that was invested in them. If I was looking at the loss ratio of the last couple of years in the AI space, it's not 0, but it's probably single-figure percentages.

will go up.

And that's not sustainable.

Yeah.

So, how do you think about where we are in that cycle today? Because at some stage, the laws of gravity will reassert themselves.

Yeah. Maybe it's helpful to explain our philosophy at the early stage, because we also don't want to target a low loss ratio.

No.

We're not taking a great amount of risk if we have a low loss ratio. We joke all the time that there's a prominent VC around in our ecosystem, and one of his big points of pride is that he's never lost money on a deal. And we're like, that's not a point of pride. [Laughter.] That's a horrible data point. That's not what you want.

Yeah. That's a PE firm.

Yeah, exactly. And so certainly you can make the case that you're not taking enough risk if that's the way you approach it. The way we've approached it historically—and this is sort of a Khosla Ventures philosophy—is that any major space where there are multiple very talented entrepreneurs building, where we think there are tailwinds, and where we have a point of view on the technology that it's good, we should pick the best founders. We should try to back the leaders at the early stage, the market leaders.

If the space happens to work out and we've got the leader, excellent. If the space happens not to work out and we have the leader, no harm, no foul. Actually, that's part of our business. That's what we should be doing.

Yep. Yep.

The bad box of what I described is the space works out and we picked the wrong one. Those are the things that we really scrutinize, and we try to make sure that we get right. There are many examples of spaces that didn't quite work out, but we did back the leading entrepreneur. They're talented entrepreneurs, they were competing, and there were lots of players in the space. That's totally fine with us.

That's the philosophy that underpins how we can have a loss rate and how we think about balancing taking an appropriate amount of risk. Obviously, that's a little bit different at the growth stage, and so we shouldn't have as high of a loss rate. As of right now, everything is so early that we don't know. There are all these unknowns about who captures value, as you said.

I'm sure loss rates are going to go up over time. All we can think about is how we build the firm, and the results will play out over time. Again, we think there are just massive power laws. We talked about it. The winners are going to take care of themselves, and we'll do our best for the things that don't work out. The way we're building our firm, I think, is catering to what the entrepreneurs want. You asked about emerging managers versus large platforms like ours. The reason we built our large platform the way we have, with a lot of scale, is because that's what the entrepreneurs want. They express that in high win rates of deals and large ownership of things that matter. One consequence of how fast this AI wave has happened is that companies run into big-company problems very early in their lives. We need to adapt the way we built our firm. That's part of the reason that we've scaled up some hiring. We're building out a much broader platform that includes things like international and channel, where we've already got experts in pricing, how you scale a sales force, and all those things in addition to all the things that we've always done for companies. The reason is that the companies are staying private longer, and they need it really early in their lives. Cursor, as an example, is billions of dollars of revenue, and they're very small and it's very early in their life. The previous generation of technology didn't happen so fast, so they didn't encounter things like major business deals they had to negotiate, complex supplier relationships, cloud deals, and international expansion. It's all happening so much sooner, and I think part of the market share gains, if you will, that we've seen is just entrepreneurs expressing their preferences.

Yeah. Yeah. So, it's funny: one of my colleagues was at a conference yesterday that was run by the UK Venture Capital Association, and they surveyed the audience, asking, "What do you think about AI valuations today? Too high, about right, or too low?" 80% said too high, and about 6% said too low.

As I think about that and the AI universe, it feels like that's probably about the right balance, because I think 80% of companies are probably overvalued today, given that we know historically that most companies aren't going to work. And there's probably going to be a small subset of those companies that are massively undervalued because they're the ones that are going to emerge as the leaders, and we'll see multiples of where they are being valued today.

I think, from an LP perspective, I really would struggle to be in your shoes today, because having to pick those individual companies—I know you can put a portfolio together—but one of the advantages, I think, of being in the LP seat is that we can have a really broad and diversified portfolio of the potential outliers in that AI space. We know historically that that basket will increase in value over time, even as the majority of those companies might fall away.

Yeah. Look, this dynamic is exactly why it's so important for our business to be centered around the early stage. We have to do the early-stage investments in those companies that end up working out, and many won't work out, but that's the nature of the beast. Our business starts and ends with how successful the early-stage business is.

At the growth stage, a lot of the stuff that we spend our time trying to think about is similar to the venture stuff that I described: our lens on the venture side, but also how much we invest in a given company in a given situation. Slugging percentage is very well covered as an industry topic, but we really have to get slugging percentage right because of that risk dynamic that you described.

Yeah. Yeah. We also get a lot of questions about whether we're in an AI bubble. One of the things that feels different today is that, typically, bubbles are characterized by excess supply destroying the economics. Today, we're in a situation where there's scarcity: not enough compute, not enough memory, not enough data centers, and not enough power. It feels like we are supply-constrained, not demand-constrained.

How do you think that changes the shape of the cycle?

First of all, it's probably a healthy thing right now, only in the sense that it probably makes it less likely that we have a bubble.

I feel pretty confident saying that we're not in a bubble right now. I'm less confident that we won't be in a bubble 3 years from now. But all I can speak to is where we are right now.

We're massively supply-constrained. You can't get data center capacity at scale until late 2028 or early 2029 right now. I think that's going to get harder. I think we're probably a year behind what people would expect for data center buildout in the US.

We're already behind. We're supply-constrained in pretty much everything in the data center supply chain. Part of that is TSMC showing restraint and trying to be balanced. Part of that is just other hardware components that are hard to manufacture and spin up to meet demand.

I think this data center resistance stuff is absolutely crazy. The arguments that I see are wild. The best data center operators are going into communities and saying, "We're going to fund a nature preserve, and we're going to fund high-speed internet in your school. We're going to make it beautiful, and we're going to create a bunch of jobs and a bunch of tax revenue." Those should all be good things, and then we're met with resistance: "Oh, it consumes too much water."

I'd rather eat 4 or 5 fewer almonds and make sure that I have the capacity to do all the things that I need to do. My yard consumes a lot more water than data centers.

We'll see if there's mounting resistance to this and if it has an effect on the ecosystem, but I think it's more likely that we remain supply-constrained for the next 3 years than that we end up in bubble territory. I would say the one thing that could shift that would be massively smaller models. That probably comes from an algorithmic breakthrough of some sort.

We do have companies that are working on that. If you just start with the human brain, the human brain is far more efficient at learning and requires less context for intelligence than models.

And so, I would expect there to be some shift in that. Everything won't be so token-consumptive in the future. If we had some massive, unexpected step change in that, maybe we could end up in an oversupply situation, but I think that's unlikely in the short term.

And then, if you look at the build-out expectations over the next 4 or 5 years, if we spend $5 trillion of CapEx, can you get $1 trillion or $2 trillion of revenue as a return on that? We could debate how much of a return you should get, but that's probably a reasonable expectation. If the 2 big model companies alone end this year at a $200 billion revenue run rate, I think everyone should feel pretty comfortable with that equation.

Yeah, over the next few years. Again, it's hard to say what's going to happen with the supply side. Supply is obviously, I think, what would drive a bubble, but I think we're so far from it right now that we feel pretty confident investing right now.

Yeah. We touched earlier on just the size of companies. What will that mean for the public markets generally, do you think? Is there enough capacity in the public markets to consume and digest that? And what does it mean for the next generation of companies that are coming along? Is there going to be some indigestion post those IPOs?

Yeah, look, I think having these companies get into the public markets while they're in hypergrowth is an excellent thing for the investor community. It's really, really good. There's been all this debate about the inclusion of those companies into indexes, for example.

Yeah.

And my parents' retirement funds are in index accounts. So, my hope is—and it seems like it's going to go that way—that there'll be index inclusion and broader ownership. So, I think it's a good thing. We've been going through this shift over the last 20 years where the number of public companies has shrunk by half.

So, I think this is going to be a good shot in the arm to bring some very high-growth, interesting stuff into the public markets. I've talked about this a lot. If you exclude the data center supply chain stuff right now, there are very few companies that are growing fast that are available for people to buy in the public markets.

The Mag 7 are all growing sub-30% at this point. All the software companies are growing sub-30%.

Palantir is the only one that seems—

Palantir is really the only one growing, 70% or whatever it is. So, I think it's good for the market to get some high growth.

Mhm.

And so, they just happen to be at larger absolute values. But again, I think the future of those companies is probably hypergrowth for many, many years, and we'll look back 10 years from now and say, “Wow, look at how big the biggest companies got.”

In the same way that we think about the Mag 7, where we say, “Wow, you never would have thought 10 years ago that we were going to have a $4 trillion company or $5 trillion company.” But here we are.

Yeah.

So, I think there'll be some shifting of ownership of things to make space for buying those companies. But I think the market's really going to be able to bear it. It's a great thing.

Yeah. One last thing I'm keen to get your thoughts on, David, is that if the optimistic case for AI is right, what do you think the VC industry looks like in 5 years' time?

That's a great question. There are so many unknowns that drive this.

If you can't speculate on a podcast—

I know, exactly. Yeah. Thought leadership of totally unknowables.

The number one thing that I think is going to drive the next 5 years' structure of our industry is what I had talked about: the sort of market structure of the model industry and the labs. The role open source plays, how much competition for tokens there is.

There's the Bill Gates quote, which I'll butcher, but it's effectively that if you're a platform, the value of the companies that are built on top of you needs to exceed the value of the platform itself. And so, if that's the future, I'm very optimistic that we're going to have a massive wave of really valuable companies that get built on top of tokens, AI, and intelligence. We're at the very early stage of seeing those. We just need to be in position to back those founders.

If you look at the health of our business, we measure it by whether we're seeing and doing the best companies at the early stage and then following on and backing those founders time and again, and that all looks really good.

But I think there's this sort of market structure question of the labs and what happens to token costs. That's probably the biggest driver of how value's going to get created in the VC industry in the next 5 years. I tend to think that there's enough smart people working on this that it's going to work out, and it's probably an end state where the labs are extraordinarily valuable and then there's this massive ecosystem of companies that are built on top of intelligence that are really valuable.

Lastly, I'd say some of the biggest outcomes—probably the biggest outcomes—tend to come from the consumer side. We spend a lot of our time talking about B2B. We're very early in shifts in consumer.

One of the things that I'm most excited about is that the last 10 or so years, pre-AI, has basically been a story of time spent getting captured by all the big tech companies, and competing with them was extremely hard. So, I'm optimistic that with all these technology changes and breakthroughs, we're going to see a shift in time spent and consumer attention, which I think will probably create really extraordinary outcomes.

Yeah. I've been investing in VC funds for 34 years, and this is by a distance the most exciting and scary time that I've been involved with. I just find that the pace of change is a real opportunity, but you've got to get things right as well.

Yeah, same here. The opportunity is so great. I think changing the way we live and work—I happen to feel strongly that it's going to make the way we live and work a lot better societally.

Mhm.

And so, I think the way that we do things is going to change a lot, and I think there's going to be a lot of value that gets created out of that.

Cool.