AI 最大瓶颈：能源与散热

It was a very simple premise when we started: tech markets are bigger than ever, and companies are staying private longer than ever. As a result, the opportunity set for us is huge. I was looking at it last night, and I think the 6 most valuable companies are US-based tech companies. It’s definitely 5, and then sometimes it bounces around at number 6. Seven or 8 of the top 10 are US-based technology companies.

Technology has swallowed the whole market, and I think it will increasingly take market cap over time. We’ve got some slides showing this whole trend, and Databricks was an appropriate way to kick off talking about the trend of companies staying private longer than ever. That’s obviously a double-edged sword for us. It gives us an opportunity to invest in companies more while they’re in the private markets, but we’re also very mindful about generating returns and DPI.

The big thing that’s changed since we started the growth fund is AI. We’ve got some slides on it. It’s massively expanding the market. AI companies are getting bigger faster than anything we’ve ever seen. The investment amounts are bigger than anything we’ve ever seen, so it looks to be a huge tailwind for us over the next 10 or so years as we look to make new investments.

Let’s jump into the details. I mentioned AI already. You can go to the next slide. The groundwork is being laid in a way that’s very different from previous cycles, and the groundwork being laid is bigger than anything we’ve ever seen before. I’m all over the team saying, “This is too conservative. These numbers are going to end up way bigger,” because I think the big tech companies alone, if you run-rate their capital expenditures from the latest quarter, are at around $400 billion of annual capex, and most of that is going into AI infrastructure and data centers.

What that means is that the infrastructure is going to get built for all of the training and inference needs that the market is going to need. This is great for all the companies building on top of it. The best part is that it’s mostly the large tech companies bearing the burden of the buildout. You’ve probably all seen the charts of capex spending as a percentage of their overall sales. It turns out they’re probably the best companies ever created: companies like Google, Facebook, Amazon, and Microsoft. They can bear potential capacity overbuild and things like that.

If you take a conservative view—which, again, I think is going to end up being way bigger than this—the buildout is massive, and this bodes very, very well for our portfolio companies that are building on top of it.

Next slide. At the same time, the input cost and input quality are getting remarkably better, faster than Moore’s law. You don’t need to look at the details of this. Just trust me when I tell you that the cost of accessing these models has declined 99%, or a little more than 99%, over the last 2 years. That’s roughly a 100x decline, greater than the decrease predicted by Moore’s law.

At the same time, the models have been improving in frontier capabilities by a double factor every 7 months. There’s been a massive decline in input costs at the same time that quality is going way up, and this bodes really well for building new things and new capabilities on top of AI.

Our house view now is that AI is going to end up like electricity or Wi-Fi. If you’re accessing electricity at somebody’s house, you’re not saying, “Let me chip in a few pennies for sitting in a room with light in your house.” I think it will end up being the same thing, in the fullness of time, with AI.

The market opportunity for AI is so much greater than the software market, and I think that’s really exciting. If you look at the previous cycle we went through—mobile phones plus cloud computing—the big story was basically creating $10 trillion or so of new market value across software companies, internet companies, and megacap tech companies. I think AI is going to be much larger because I think the impact on the economy is going to be much larger.

If you look at the simple math we have on the page, US software spend is about 1% of GDP, while US white-collar payroll is about 20% of GDP. There are a lot of areas where I think we’ll see augmentation, potential cost savings, efficiencies, or replacements using technology.

There’s always a question, when these things happen, of how much the new companies are able to capture versus the end customers. My rule of thumb is that 90% of the value goes to the end customers, and 10% of the value goes to the companies serving them. It turns out that’s a massive amount of market cap if you’re the 10% capturing that value.

The examples I always give are: What does your iPhone cost? The latest one is, give or take, $1,000. If you had a gun to your head, what would you pay for an iPhone? If you’re on the higher end of the income spectrum, probably far more than $1,000. The difference between that and the $1,000 you pay is the surplus, and it turns out Apple is still a really great business.

Or take Google’s properties, like Search and Gmail, and increasingly its AI products. You get them for free, which is massive surplus, but Google is probably monetizing you at around $200 per year or something like that. There’s a tremendous amount more value delivered per user than that, I would argue.

I think the big story is going to be massive new surplus created. A ton of it gets captured by end customers and end users, whether they’re businesses or consumers, and a massive amount of new market cap goes to companies that are capturing that opportunity.

There was a great meme the other day where someone said, “Imagine if Google had known that users were willing to pay $100 or $200, as they are with ChatGPT. If they had known that people would pay for a similarly magically delightful product like Google, God only knows what Google’s market cap would be today.” Obviously, it might have an even more dominant market share than it already does.

It is amazing when you think about the reconfiguration of how we will view monetization in this new world.

Before we go even further, David, I think these last couple of slides set up probably 95% of the questions around the market. If you go back to slide 7, I think that slide is daunting for a lot of folks, in part because you see it bundled with the headlines, but also because it feels like many of us lived through the early 2000s, when it ended in a very less-than-desirable picture.

Maybe you can summarize your case for why it’s different and particularly talk about the timing cycles. Incidentally, despite the massive broadband buildout and then the glut, we are, of course, beneficiaries of that today.

Well, we grew into it. We ended up growing into it. It was just a time lag in that case, and it wasn’t the strongest companies in the world building it out. The thing to watch will be what role leverage plays.

I read an article this weekend asking whether there’s systemic risk in data center buildout. First of all, most of the people with their necks on the line are private capital. Private capital is playing a role right now, and the biggest funder of private capital is actually banks, or private debt. Banks are the ones funding the private debt companies, and increasingly, they all have insurance companies. Maybe insurance companies are backdoor funding this buildout. That’s a really good sign for the stability of the buildout.

That’s the nature of the supply side, which, again, feels different this time given who is actually doing the buildout and who the tenants are. The demand side is the bigger, more interesting thing.

I read a statistic yesterday that the time to get to 365 billion searches on ChatGPT was 2 years. The time for Google to get to 365 billion searches was 11 years, so it took 5.5 times longer.

The big story on the demand side for me this time around is that AI is built on the back of the internet and cloud computing. Because of that, it allows for immediate global distribution. If you look at the way Google and Facebook started, for example, they started much smaller. Both had a network-effect dynamic, which just takes longer. We didn’t have full internet proliferation across 5.5 billion people, smartphones in everybody’s hands, and the ability for everybody to access the internet.

Because of the nature of this technology, you don’t have to deliver a new hardware product. We have global internet buildout and cloud computing, so the whole world can access this. If you take ChatGPT again, they got to their current scale 5.5 times faster than Google, which is staggering.

There are probably—I don’t know—a billion monthly active users. The latest figure is that there are more than a billion monthly active users, and there are probably another billion or so people who have tried it. If you add up all the different platforms, a lot of people have probably tried Google products and Facebook products.

Probably well over half of the global internet population has used AI tools already. We know that there are probably, in some shape or form, somewhere between 1.5 and 2 billion active users of these products. The speed at which they got to distribution is unlike anything we've seen before.

That is heartening to me, because the supply buildout will be utilized in a more predictable way than broadband in the early internet buildout days, just because it's built on the back of the previous infrastructure. I also want to comment on the Google thing you said: imagine if they could get $200 or something from users.

The beauty of Google, Facebook, and Apple for us as consumers is that there's not a clean way for those companies to price-discriminate. They don't know that I'm willing to pay more, and they can't charge me more than the other user of an iPhone. But in the case of AI, because of the way the business model is structured—and we've already seen some proof points of this—I think there will be greater success.

So today—or I guess yesterday, technically—OpenAI released its India subscription product, and I think it's something like $3 or $4 a month. That makes total sense. In the U.S., there are high-end subscription products that are $200 to $300 a month that are flying off the shelves. Consumers can't buy enough of them.

To me, the real story of the growth in this market is that there's going to be an evolution of the business model that allows these companies to address the user base and price-discriminate in an effective way. They can do a combination of subscriptions for higher-end users and get to the point where those users are willing to pay more. They will probably also end up with premium products that they monetize through some form of advertising.

It's hard to speculate now on what that would look like. I think it's probably some form of an affiliate-type thing. That has a dirty connotation because that's kind of a backwater industry in today's internet, but I think it will end up looking like that.

One way to see it in the product, if you haven't done this already, is to go into one of the deep-research products—OpenAI, Grok, or whatever it may be—and have it do a really sophisticated shopping-research project for you. It comes out with incredible stuff. I had it do this for my son's baseball bat, literally, because it required a bunch of different specifications and things, and I wanted to look at the year-over-year value.

It came back with extraordinary answers. This is a far superior experience to researching by typing into Google, clicking around, and having 7 sponsored links above you before seeing anything organic. I think there's going to be an opportunity for them to monetize free users.

The big story in the case of OpenAI is that there are probably 30 to 40 million paying users today. The other platforms are kind of a rounding error relative to that, so maybe add another 10 million. There are roughly 40 million people paying for this stuff today at some level, and there's probably 2 billion people using it.

Facebook and Google monetize their properties at, for U.S. users, between $150 and $200 a year per user. There's a massive amount of opportunity to monetize. On the consumer side, there's probably going to continue to be tremendous surplus. I find tremendous surplus in it today.

Daily active users of ChatGPT already spend around 28 to 29 minutes a day on the product. To put that into context, I think Instagram is around 50 minutes a day, and TikTok, sadly, is around 70 minutes a day. This is real time spent and real consumer value already.

I know the question was about risk on the supply side and the infrastructure buildout, but the actual usage and distribution that we've seen this time around makes me think that it's different. We have a really good view of demand signals that took many years to develop in the case of the internet, or even in the case of mobile phones, because you had to manufacture phones and convince people to buy them.

In that way, it's built on the back of the previous technology cycles, which is great. But it also, in my mind, derisks the forward-growth potential of the AI companies that we're investors in.

Totally. Pulling up the ChatGPT website and being able to do it accessibly—almost like the party trick of, “Hey, let me show you the cool thing that I just did”—that usability is obviously very different from the prior cycle, where we literally had to wait for the infrastructure to be built and for the device and hardware to catch up as well.

Our early-stage team also did a great post on this topic if you're curious about the future of commerce and what that looks like with AI. Even recently, if you have folks on this call who spent time on the public side, one of the things that has been significantly observable is the number of public companies that have reported a decline in referral traffic and engagement.

Largely because, with Google Search now, they're just doing the AI-summary version of the results flow. We could talk about the implications and what that means in terms of the downstream effects, but that is certainly top of mind. We're hearing a lot about it from folks in the Fortune 500, given this reorientation of how they engage with the consumer, when you could actually do a really detailed search, as David described with his son's baseball bat, without ever having to go to any website at all.

I told my son, “You have no idea how much research I did, man. I scoured the end of the internet for you, buddy.”

That's all dads in their desire to have endless amounts of research on arbitrary things, to say the least.

Here, I'll send that. There was a good little snippet on X about some of that Google Search traffic stuff. It was like AI [?].

Yeah.

Oh, yeah. The earnings call that Martin dropped the other day. Just so folks could see, again, it's the folks you usually would expect, right? It's folks like Groupon, for example, who are seeing the impact in real time.

There's one question from Chris we should take while we're just on this slide here. I'm going to throw in another one that's related to it around the shifts in bottlenecks. Right now, there's obviously this massive bottleneck as it relates to compute, but Chris's question was, “Is there enough energy to actually power this buildout as well?” We should lay into what we think the next bottleneck actually is beyond energy as well.

I mean, yeah. As of right now, through our current means of energy production, energy is a bottleneck. We've made investments on the nuclear side, and I'm quite optimistic that we now have an embrace of nuclear power. I think Three Mile Island is going to get powered back up.

The big tech companies are building data centers near nuclear power plants. We've figured out that there's a lot of natural gas in places like West Texas that can be used to build large training clusters very near to them and power those data centers pretty efficiently. But we're going to need different sources. That's the short answer, and we're probably most optimistic about nuclear.

For sure. And then there's the ability to just build these things. They're massive-scale operations.

One of the most remarkable things—and we're large investors in xAI—is that xAI stood up the biggest data center at the time in roughly a quarter of the time that anyone else had done the same thing. They had to do crazy, unnatural things, like get every backup generator in the multistate region bought out and buy labor off of different projects. But they did it.

The actual construction and build is massive. My view on chips and infrastructure is that production capacity will typically scale to meet demand. There's always a dislocation, and you've seen it. I think energy ultimately, in the next 5 years, will probably be the bottleneck, and that's why we're so excited about nuclear and making investments in that area.

Absolutely. And just to extrapolate out, once we figure out that piece—which inevitably, with any technology, you always do—the bottleneck just shifts to another.

The big component that I think most folks have not yet realized or zeroed in on is the cooling piece. And so you'll see a whole wave of innovation around that part as well. Once you figure out how to generate all this energy, how to actually cool all this stuff down without boiling our oceans and making the world melt down.

And making the chips melt down.

And making the chips melt down. Yes, that's right. There's one question here, David, if I can interject, because you are the business model snob. So this is a perfect question for you.

There's a lot of debate about whether the gross margins of a lot of AI companies should be more scrutinized. In particular, there's a lot of turmoil around, for example, the relationship between Cursor and Anthropic, and whether the growth for a lot of companies might actually be masked by a reliance on some of these models. Also, what are the actual unit economics that are considered best in class? Maybe help us distill how you and the team think about that, and particularly what this topic around gross margins is, where you're willing to make some short-term exceptions for a long-term benefit and output on the other side.

Yeah, I love this topic. I would just say that the reason that this industry and this job are so fun right now is because the range of outcomes is so much greater than before. The variance is so high. I was talking to the team at the offsite yesterday, and I was like, “Do you remember in 2000?” Some of them weren't in the industry, but the ones that were in the industry—

But I wasn't born yet, either.

I know, some of them weren't born yet. Yeah. So, some young folks who are very AI-native and very smart. But, you know, late cycle, there are questions that you are trying to answer that are interesting, but they're far less interesting. So you're like, “Oh, how big can Datadog get? Or just how big can—pick your SaaS app—how big can it get?”

And now we have all these questions around business quality, market power, who the winners are going to be, and even if you are a winner, whether there is going to be value that accrues to you and where you are in the stack. So the range of outcomes is so much higher. And I think if we do a good job in that period of time, what that means is hopefully we can get a greater degree of variance in our own outcome as investors. And so I'm very excited about that.

So, yeah, what are we thinking about in the business model? One, the value proposition to the customers is the number one thing that we care about. Is there customer love of your product, and is that love enduring? If you made me pick 2 top-line stats to look at to assess the business model, it would be gross retention rate.

Gross, because we obviously get to look at net retention rates, too, but with gross retention rates, it's sort of: Are people getting value out of your product? If you have 100 customers, in absolute terms, dollar-weighted, how many of them are sticking around? We look for things where 90% or more of the customers are sticking around. And hopefully they're expanding their usage, which would be expressed in net retention. But core value proposition is shown in gross retention.

And then ease of customer acquisition. You see that through organic customer demand, high value of the dollars that they're willing to pay relative to how much it costs to acquire them, either via marketing or sales. If you have things that are being pulled off the shelves and you have high endurance of the customer relationship, that's typically the best thing that you can get for business-model quality on the top line.

You mentioned gross margins. We care a lot about gross margins, and there's a bunch of debate right now around some of the AI-native application companies and their gross margins. I think our hypothesis and hope in the market is that if there are multiple model providers that are somewhat close to par, you're going to see input costs go down significantly over time. If you recall, there was a 100× decline in the input cost over the course of 2 years. The hope would be that that continues, and all indications on our side suggest that it will continue. Maybe it abates a little bit, but it will continue pretty significantly down over time as long as there's competition at the model layer.

Coding is one of the areas that people have spent a lot of time scrutinizing in this area and assessing the business models and business quality. I would say relative to mature SaaS apps, we probably are a little bit more lenient in assessing a company's gross margin today because we strongly believe that their input costs are going to go down over time. And because of the model improvements, they'll be able to harness better models and deliver better products to consumers over time.

So they won't need to increase price, but they'll deliver a lot more value and stickiness while their input costs go down. That's the hypothesis. I think that's subject to there being multiple players in the market that serve these models. Now, we're thrilled that GPT-5 is out, and it's a very credible alternative that will put pricing pressure on Anthropic. Google is also very focused with its Gemini models on coding, and we've seen a bunch of really good improvements and promising progress out of them.

As long as there are multiple players in the market, I think you'll continue to see costs go down. Again, in light of that, relative to mature-industry SaaS or infrastructure companies, we're a little bit more lenient in assessing a company's gross margins today. We don't want to invest in a bunch of companies with zero gross margins, and we don't do that. But if you took gross margins, retention rates, and ease of customer acquisition, I'd place far more emphasis on making sure that we feel like there's greatness in those latter 2, and give them a little bit more benefit of the doubt that they can improve on the first.

For sure. Yeah. I flip back to this because I've seen this myself a dozen-plus times, but something you said there really just clicked for me. A lot of people try to make comparisons to the dot-com era, and they're like, “Oh, remember we measured eyeballs as well?” Isn't that akin to this in terms of retention or usage?

This is reaching such an immense scale so quickly because of all the reasons you alluded to earlier, David, in terms of the ability to just pull up a website and actually try it. But also, people are paying for this. It's an interesting juxtaposition when you think about the last cycle of things we paid for, so to speak, like a Spotify subscription or a Netflix subscription, where as soon as pressures come from a budget standpoint, those are the first to go.

But this, you'd probably sacrifice a few things for, given how much of an impact it has made, either professionally or personally, by accelerating your productivity and, hopefully, your time as part of that.

Thomas's question. I'm happy to take Thomas's question.

It's from Thomas. So it seems like there's downward pressure on the likes of OpenAI on consumer pricing, yet the cash burn of OpenAI is ramping up to levels not seen before—reports of $1 billion-plus per month. How does cash burn moderate in the future relative to what you think, in your opinion?

Yeah. So what I was saying is that I think, effectively, there's greater consumer stickiness than you would think, and there have been a tremendous number of free alternatives thrown at consumers over the last 12 months, and it hasn't had any impact on their business. Now, that could change over time, but so far what we've seen is no effect that creates price pressure.

And if you think about what I had described earlier, which is, call it, a billion people using it and only 30 million people paying for it, I think there's way more upside to monetize the base than there is risk of price pressure on today's 30 million people paying for it. So I think there will end up being—if it's a P times Q, and we're talking about ChatGPT, I would say it probably applies to most of the consumer-facing stuff in the industry.

If there's a P times Q, which is price times quantity, which is the really simple way to think about these consumer internet businesses, Q is, at this point, they've gotten so big on the monthly active users. Over the course of the next 5 years, maybe it gets to 2 billion or something. I don't know. But the Google and Facebook properties are in the 2 billions, so it only gets so large.

But I think there's a tremendous amount of room to run on the upside on the P. So again, if you think about how they're monetizing today, it's 30 million people out of a billion at a modest subscription that is not really reflective of real price discrimination yet. So I suspect the P is probably the thing that we get surprised on the upside by.

I think about lessons learned from previous internet-era companies, and I remember looking at internet companies 10 years ago. We would always look at Facebook and Google, and we're like, “Okay, Facebook and Google, they're monetizing their users at X. That's the max we could get to.” And the big story about what's happened over the last 10 years is they've 8×'d their own monetization of their users.

And so I suspect that if they have some thoughtful ways of monetizing free usage while still maintaining trust, there's probably more upside than downside on pricing, to the question.

On the burn, the actual—most of the burn comes from research and development, and future investments. We could apply this to the whole industry of model companies, but we could talk—if it's specifically OpenAI, I'm happy to address that one. On the OpenAI side, I think they're in an advantaged position because they have the consumer base, and that's stickier. My family—my parents in Kentucky—use ChatGPT. If there's some slightly better model that comes out, they're not going to switch. And so I think that's pretty durable.

And so it's probably a better position to be in to fund those research efforts for continued model development. There used to be a thing that enterprise companies are stickier than consumer companies. In this case, they're developer companies—these are developers buying these things on the B2B side, for the most part, today, buying raw access to the models—and that's not very sticky yet. I think it's possible that it does get sticky over time, but as of right now, it's not very sticky.

To the point about coding models, if there's a new coding model that comes along that's better than the latest version from Anthropic, our coding companies will just switch. It's pretty easy to do because it's an API call. So, interestingly enough, it's a little bit different this time, where the consumer is a little stickier. I think that gives you an advantage.

And I think the companies will not irrationally spend on new model development if there's not a financial return. One of the things that we've observed over the last almost 5 years of spending time with these companies is that a lot of the founders started as research-brained AI people who were like, "We're going to—there's going to be no economy, and everything's going to end because we're going to have AGI." And then, you know, what has happened is that competitive forces have kicked in, and they've become hard capitalists. My expectation, from conversations with them on an ongoing basis, is that they're not going to do totally irrational things on the research side if there's not going to be a financial payback for them.

Can David George also comment on the durability of revenue for many of the AI applications built on LLMs—that is, outside of OpenAI, Anthropic, and xAI?

Yeah. It depends on the nature of the use. I think some of them are really sticky. Companies using medical scribe software, for example, are pretty sticky because there's a bunch of doctor workflow built around it. Customer support is pretty sticky. Some of the high-end financial analysis type of stuff is pretty sticky.

Can you explain why you think those specific areas are stickier than others?

Yeah. I think the more stuff that gets integrated, and the more company-specific rules built around the model you have, the stickier it's going to be. Stickiness in software applications comes in the same way it has always come with software. I'm sure one of my early-stage partners has written a blog post about it because we talk about it all the time, but it's things like integrations, rules engines, workflows, and enterprise capabilities.

For a customer, for example, the rules and the sequence for how to troubleshoot are so embedded that you probably wouldn't experiment much once you've got a workflow across multiple different scenarios.

Yeah, yeah. And even the style with which something engages—these are companies that are customers of these things, and they're brands. They care about the way the customer support agent interacts. So I think that stuff's probably pretty sticky.

I think there's a bunch of stuff that's not sticky at all. Some of the emergent behavior that's not as sticky is experimental usage of tools to build some of the internal tooling software replacements, or very low-end prototyping of websites and things like that. I just think it's TBD who the players are and what the use cases are.

I think the market for that will probably segment out. We're already seeing it somewhat, where some of the tools are just being used for prototyping, and some of the cool tools are being used to actually build and deploy apps. I think you'll see continued bifurcation of those things.

But it's so early. I don't think that companies are going to vibe-code up their Salesforce.com. It's just not worth it. It's not a core competency. I hope they would—I wish we could vibe-code away our Salesforce.com.

Okay, I'm going to speedrun through some of these questions because I recognize we've opened the floodgates, and I want to make sure we get through as many as possible.

One question asks, "Given the speed of go-to-market for AI companies, do you think $100 million ARR is the right milestone to measure against, or are you starting to move the goalposts on what success is there?"

Yeah, it's so funny. It used to be—I can't even remember—one of the VCs coined the term "triple, triple, double, double" or something like that, and that looks very modest.

Yeah, something like that.

Yeah. That time compression has massively gone down. We've run the analysis on it recently, but it was like the top companies that we've seen have gotten to $10 million, then $100 million, 4 times faster or something like that.

For us, it's really important that we have market context and real-time market context to make judgments about what great looks like. I don't have an absolute answer for you, but if you were to sit there and listen to one of our investment discussions, we are comparing the growth of some new app to Cursor, and nothing looks like Cursor, so it's unfair, but also to Decagon, Abridge, and ElevenLabs—not Shopify, DocuSign, and those companies.

If you're not fully in the market, seeing everything, and meeting all the companies, you're not going to be able to have that context to assess what great is at any given time.

For sure. For AI companies, what do you think about seat-based pricing versus usage-based pricing, and what are some of the experiments you've seen startups do well around configuring pricing in this age of AI?

Yeah, this is the big question. I love this question because I feel like on X and in other blogs, people have written all these really long posts about the new prices that are going to happen with AI, and to me it's a little hand-wavy. One, it's subject to innovation on the technology side and the products getting better. Two, we've really only seen true business-model innovation in the early, early days in one area, and that's customer support, because you can definitively resolve a task.

Even with that, it's hard to do. The question, maybe to reframe it, would be that we had licenses—perpetual licenses with maintenance—and then we switched to SaaS, which was mostly seat-based. That was a huge innovation, very disruptive, and a huge enabler of startups actually beating the incumbent software companies because it was so disruptive at the time. Then we got to usage-based pricing.

Obviously, all the cloud companies run that way. Databricks runs that way, Snowflake, and so on. The hope, which I've seen and heard about, is that with AI, you can just monetize the replacement of tasks humans do. When I talked about the customer-support piece, that's the furthest along in monetizing the tasks that humans do.

But that's super, super early in all the other areas. I think it's really early. End customers want to buy stuff on seat-based and consumption-based pricing, and you have to meet the market where it is. For now, we're not seeing hugely disruptive things.

As the capabilities get much better, and the measurability of task completion gets more objective, perhaps we get there. But I would say it's super early days, and I'm low-conviction that we end up, 5 years from now, with all the software companies monetizing in a completely different way.

Absolutely.

By the way, sorry, this is one more thing. This also goes to my point that I made earlier about surplus. It's going to be really hard to capture that surplus unless you can monetize based on the value of the completed task. But because that's going to be really hard to measure, and there's going to be competition in the market, I suspect that a lot of the surplus or savings may end up in the hands of the customer.

Again, you could probably still build really interesting, good companies that have really high market caps. But that is directly related to who captures the value and the surplus. I always say the steam engine got invented, and you didn't price the steam engine based on a calculation of how many humans it replaced in doing a specific task. Competitive forces kicked in, and it was priced at some appropriate, competitive level with a return on capital, while still capturing a lot of value but delivering much more value to the end customers or the users of the steam engine. I think the same will happen here.

For sure. Actually, it's a good transition for us to talk about what this all means for growth, given this setup.

I try to sit in on as many pitches in our early-stage practice as I can, just because I feel like I learn a lot about what interesting founders are starting to hone in on and in what areas they're building. Our early-stage team is doing a great job of doing the most exciting early-stage deals. So that, to me, bodes really well for the next 12 to 24 months of deal activity for us.

Looking out—alright, growth partner. This is a crazy-looking chart. I mean, it's not surprising because it was the first thing we talked about, which is that tech markets are big and, of course, companies are staying private longer. But you can see it in the chart here. It used to be that companies would go public within, call it, a 5- to 10-year window of inception, and despite the fact that companies are growing much faster and getting better quicker, they're staying private way longer. All that is to say, it's 14 years, and I think that's getting even longer.

If you take the market cap of private companies valued above $1 billion, we could argue that some of them are overvalued or undervalued, but that whole value in aggregate is like $3.5 trillion. That's like 11%. Three and a half trillion is like 11% or 12% of the Nasdaq, or 10% of the Nasdaq, something like that.

If you go back 10 years, that whole private market cap of $3.5 trillion was like $500 billion. So over the last 10 years, the market cap of these private companies has grown 7x. There's huge growth. Again, that is related to this point, which is that some of the best companies are taking longer, deciding to stay private longer. But it's pretty stark.

The other thing that's going on—I mentioned this earlier with Snowflake—is that the public markets are no longer the place of extremely high growth. Because of this, it follows logically, right? Something like 5% of software and internet public companies are forecasting 25% or more next-12-month growth. So 95% of the public-market universe in software and internet is growing less than 25%.

If you want access to the high-growth segment of new technology companies—and this is a little bit too bad because there are obviously implications for retail and stuff like that—the reality of what's happened is that the high-growth segment of new technology companies is all living in the private markets now, for the most part. I don't see that changing. I do think Figma and a bunch of good companies will go public, but this is a trend that's pretty long-standing, and I don't see it reversing anytime soon.

We have a bunch of really good AI companies. Some people have asked, “Isn't there crowding that's happened in some of the best companies?” What I would say is, one, we're getting into a lot of these companies much earlier than others are able to. And 2, even for the ones that are later, where there are bigger rounds, it's really important that we have that early relationship because it gives us ball control and access into the rounds. xAI was the first outside money in beyond Elon.

A lot of these are growth-fund kind of first money in, so it's important to get in earlier to generate returns, but also to position us with management and to have access to shape later rounds.

If I were to spend a second on how we've shaped our AI strategy, there's basically 2 parts to what we're doing. One is companies that are flying off the page, with undeniable momentum—companies like Cursor, Decagon, ElevenLabs, Abridge, and others.

The second bucket is very, very, very early deals in the very, very, very best teams in the market. And I say “very, very best teams in the market” with emphasis because I think we're talking about the top 5 teams in the world, and anything beyond that, we're not doing these companies. They're a different shape.

They're growth dollars earlier than normal, and there's a higher degree of variance in the business outcomes. But because the teams are so special and so great, we feel like business risk looks very different from capital risk. We feel like they're kind of asymmetric bets, where the asymmetry lies in the fact that we're probably downside-protected because the team quality is so high and there's so much demand for talent. We would be downside-protected even if it doesn't work out, and so there's high business variance but an asymmetric capital or returns profile that looks a little bit different than a typical early-stage investment.

We talk about timelines to exits because, on one dimension, folks may take the reaction of, “Gosh, private for longer means just an extension of hold periods for LSV companies.” Maybe the reframe is: Why is that great for us, frankly, as private investors? Why do we think it's actually advantageous, in some respects, for some companies to delay time to IPO? And how are you counseling folks as a part of that?

Yeah, this is a tricky issue. I said it right up front: We get paid on DPI, too. That's how we compensate ourselves and our teams, and so we care a lot about it. I would say we're in a fortunate position where our portfolio is pretty good, and we've had a number of companies that have exited to the public markets or sold themselves. So we're not in a position where we haven't been able to deliver some liquidity.

For those sort of champion companies that have been private for a long time and are getting a lot of coverage, I think each has idiosyncratic reasons why they've wanted to stay private, and I think I'm confident that they will go public. I think most of them will probably end up going public. It just will take a longer amount of time.

One of the big things we've observed is that the private markets have adapted to some of what you get from the public markets. We've fortunately been very active in some of those situations. Things like tender offers in the private markets are a big thing.

What's hard for private companies is competing for talent. Part of the reason why it's so hard is because public companies grant RSUs that typically vest on a quarterly basis, and so they hit your account. They're already tax-withheld, and it's basically like getting paid a lot of cash. It's hard for private-market companies in some cases to compete with that.

Some of the bigger, better ones have done more regular tenders, and we've been very active in shaping those with the companies. To me, it's a balance. We want what's best for the company. If there are strategic reasons why it benefits them to be private, that's great, and we'll help enable that. But we do recognize that our goal is to monetize great investments. We're not there yet, where we feel like there's a need to do unnatural things. I just don't see that being the case.

For sure. And I do want to call out as well—I'm going to brag on behalf of you, David, because I do think DPI generally is an issue and challenge across the industry. I don't think it's our issue, necessarily. I'm also going to cover a question here. Thank you for letting me do our commercial, David, on DPI. It's my favorite topic.

On the topic of some of the names that we have underwritten, I think there's this broader question: There are a lot of names—for example, the ChatGPTs and OpenAIs of the world, and also Databricks—that feel like they're doing more of these tender offers and more of these SPVs. How do you counter that portfolio composition of names like that versus names like Anduril or Flock, where it's probably very, very difficult to get any access? And how do you think about the configuration of the portfolio as a part of that?

Yeah, look, the OpenAI investment that we made—we were able to write exactly the check that we wanted to write. In terms of portfolio construction, it doesn't seem—I'm sure some folks have had access because other managers are trying to do SPVs and stuff, but I think that's increasingly hard. It's also our decision on how much to weight these opportunities and build a portfolio.

So, I happen to think that the return profile of that recent one in Databricks is very attractive. But we don't want an entire portfolio of things that we think are pretty safe 2xs, where I think the question is, can we make 5x on them, but we feel pretty confident we can make 3 to 4x, and 2x is pretty safe.

Awesome. Speaking of public companies, we had a question from Paige here. Well, actually, it's 2 related questions. What publicly traded software companies do you think won't be disrupted by AI and have real moats, and are there any obvious ones that you think will be? And related to that, I think we had some questions earlier around how much, if at all, you invest in public companies in the fund.

Yeah, probably very few public companies in the fund. We'd have to have a really, really strong thesis and relationship with the management team to want to do that. Again, the public universe is just way slower-growing, and there's really, really exciting stuff that we can do in the private universe. And so the bar is extraordinarily high for things in the public markets, and we've done a couple which have worked out, I would say.

But the bar is very high, and the opportunity cost is probably something that may be even higher growth in the private markets. The question of what companies are going to get disrupted in the public markets—I love this question. I don't know. All I would offer is this framework, which is maybe there are 3 interesting ingredients to consider when you think about the safety or durability of the public software companies.

One would be UI/UX. So, to the extent that we get a complete reimagination of UI/UX, that'll be really exciting. What is Salesforce.com? Salesforce.com is a set of really uninteresting checklists and forms that people fill out, for the most part, on the front end. The promise of this stuff—agents is a way overused term—is that the new technology should be able to do things for you as opposed to keeping your records for you.

And so you could imagine a completely reimagined UI/UX which is proactive, which is: “Hey, instead of you going and inputting things, I know what you're doing. I'm just going to tell you what to do.” A total reimagination of the workflow would be one ingredient for a startup to win versus an incumbent.

The second would be access to data. An entirely new form of data gets sucked in to take the actions on your behalf. With Salesforce, I mentioned the form thing on the front end; the reason it's really, really sticky is the database on the back end. So, to the extent that instead of using that sort of structured database on the back end, you take all of your unstructured data and dump it into Databricks, for example, and query it or access it through that, that would be another interesting opportunity for a startup to attack.

And then the third is business model innovation. I mentioned this already: it's super, super early days, but to the extent that startups can come up with a novel way of shifting the business model in a disruptive form against seat-based pricing at Salesforce.com, it would have a chance to win.

For startups to win, I think you need all 3. And that's just in the head-to-head stuff. I think startups are already finding interesting windows of opportunity in and around these systems of record, which we've made a bunch of investments in. We haven't found the startup that's got the killer idea for dethroning Salesforce.com yet. I hope we find it.

I know that's an unsatisfactory answer, but at least it's the framework I would use to think about those 3 buckets.

Awesome. Okay, I'm getting the hook because I know we only have a few minutes left of this webinar. So maybe in the last couple of minutes, we'll fast-forward here to talk about the team. I don't know why it took us over an hour to get to this point. So maybe we'll go through that, and then, David, also how you work with the broader a16z team. I'll bring us all back home with how this all fits in with the broader franchise and firm as well.

Yes, I love the team. They're very, very smart. They've come together in an awesome way. We have a really strong team subculture. We're very, very intertwined with the early-stage teams.

If you think about where we get alpha in our business, there's access, which I've talked about already. In terms of being already involved in the companies, 80% of the time when we make a new investment that's not an early-stage investment, one of our early-stage folks has some preexisting relationship. So access is a big piece of it.

Insights is the other piece. I think the way you get really outsized returns in our business is through market and product insights. I think you have to do all the analysis around business model and financials because you can make big mistakes if you get that wrong, and we go super deep in each of those. But I think the real outlier opportunities come when you have a market or product insight that maybe the rest of the market hasn't figured out yet.

And the fact that we're attached to the early-stage team gives us the opportunity to have those insights and gives us a tremendous amount of leverage to have a relatively small number of team members, given how much ground we cover.

I will also take the opportunity to brag on your team and my team because I think both of our teams got a 91% on the employee engagement score. So, not that we're competing across teams, but this is a good barometer for team subculture, no doubt, and something we always do around the firm year-round, which is very company-like but hopefully gives you a sense of what we're measuring ourselves on as well in terms of team culture.

But I want to spend a minute to illustrate the point that David alluded to around how we work in collaboration with the early stage, and also think about late-stage venture in terms of sitting across all 6 different early-stage buckets across the firm.

And so maybe, David, if you want to give just the high-level talk track on, if you were to hypothesize what the composition of the portfolio would look like across these 6 different buckets, that'd be helpful. And then I saw a question around how much we would expect around crypto investments, and particularly around token crypto investments. That would be helpful.

Yeah, absolutely. The best part about where we are is we've got people at the early stage figuring out where they think they should be spending their time. And typically, our world is 12 to 24 months downstream of that. So when you see the way we size our funds at the early stage, that's a reflection of what those early-stage teams think is their opportunity set, and ours sort of follows that.

So I think the largest amount of opportunity will continue to be AI infrastructure and AI applications. Next, American Dynamism. Our early-stage team is killing it in that area. Beyond just AI, there's obviously a pressing market and world need for American Dynamism companies.

At the same time, you have market need, and you have people that figured out how to do this inside of SpaceX and Palantir who have then left to start companies. So the talent is there to navigate a very complex go-to-market motion. And then you have advances in technology beyond GenAI, like autonomy capabilities and vision advances, that enable American Dynamism. So we're very excited about that area.

We're starting to see a little bit more interesting stuff in AI-enabled health. We've done a couple. You'll continue to see us active there, but it's probably a little bit less than the AI infrastructure and application side.

And then crypto—the way we're doing that is working hand in hand with Chris and the crypto team for their high-conviction bets where they're at a stage that fits the growth fund. I brag on them: I think they're the best crypto investors in the world, and we're happy to attach ourselves to that.

It certainly depends on the opportunity set. We're seeing really exciting stuff in the enablement of stablecoins, and so, to the extent that we see a massive takeoff in that market, it could be more.

I always say it would be great if 100% of our investment activity in the growth fund was follow-ons, because that would mean that our early-stage team is absolutely killing it and they're getting a lot of market share. But I think there will always be a place for us to do really attractive new things.

And we don't build the portfolio based on some target around new versus follow-on. It's really best ideas—where do we have access to win, et cetera.

Great. And then, in less than 60 seconds, do you want to take Annie's question? On the topic of portfolio construction, how much exposure do you want to top research teams with a wide range of outcomes versus opportunities with a more narrow range? How many more researchers do you think are actually out there, beyond the ones that you've already backed, that you'll get more exposure to?

There are extremely high-end researchers at some of the big labs that our early-stage team is tracking. And I should have mentioned this: when we do these, we do them with the AI infrastructure fund.

We rely on them heavily to make sure that we're doing the right assessment of the teams and the research ideas that they're pursuing.

In terms of portfolio construction, I quite like the way that we've done it. We have some absolute champion companies that we think actually have a lot of room still to run. On the downside, we think that if things really go poorly, we'd still probably make 2x our money. On the upside, there may not be the opportunity for 10x over 5 years, but we think there will be the opportunity for 5x over 5 years.

The research team with the high variance is a little bit of an output of that great opportunity set. Right now, there's not another Ilya floating around in the AI market, and so we would never try and say, “We want 10% of our portfolio to be in these, so we need to find someone.” It's more reactive when we do find those people who are really special.

I think the only shift that we'll see over time is that you're seeing a bunch of really, really exciting AI apps and American Dynamism companies at the earlier stage that are poised to become champion companies over the next 5 years. You'll probably see a lot of activity from us in those kinds of companies.

Awesome. With that, I'm going to close out here. Lastly, thank you, David. I appreciate it.

Great to hang, as always.