有 AI 泡沫吗？Gavin Baker 与 David George

Are we in an AI bubble?

I do not believe we're in an AI bubble today. I was, depending on how you look at it, the privilege and the misfortune of being a tech investor during the year 2000 bubble, which was really a telecom bubble. I think it's really helpful to compare and contrast today to the year 2000. The year 2000 internet bubble, or telecom bubble, was defined by something called dark fiber. At the peak, 97% of the fiber that had been laid was dark. Contrast that with today: There are no dark GPUs.

And that brings us to our opening fireside chat. We're going to start with a taboo question right out of the gate. Are you ready for it? If AI is the biggest trend in the world right now, where is the evidence for it? Why is it only just beginning to show up in the economy? And as Andrej Karpathy asked, are agents really just ghosts?

To kick this off and to help us answer this question, please join us in welcoming Gavin Baker, managing partner and CIO of Atreides. Some of you may know Gavin as that really thoughtful guy on Twitter. Anytime some big piece of AI news comes out, I know more than a few people who count on Gavin to explain what the fuck is really going on.

A huge thank-you to Gavin for being with us today. Joining him is our very own David George, general partner at a16z. Who knows what that music was from?

Glad they got our pump-up music right.

Yes. Battlestar Galactica, the original 1977 one, in case we have to all fight Cylons in a few years.

It's a good segue into the topic, I guess. Thank you for being here. I always love talking to you.

Same. I'm really grateful to you for inviting me, and grateful to your colleagues for having me here. I'm really looking forward to the next 2 days. I think I'm going to learn a lot, so thank you.

Yeah. Okay. All right. The big topic is the AI bubble, kind of a macro view of things. Maybe just to start with a couple of stats to set the stage, and then I want to get your take on where we're at.

We have about $1 trillion of data centers in the U.S. The plan is to add $3 trillion to $4 trillion in the next 5 years. Over the past 3 years, we have already built out, in data center capacity, a larger amount of dollars than the entire U.S. interstate highway system, which took 40 years, just in terms of dollars. And that's inflation-adjusted.

OpenAI alone, I think, has more than $1 trillion of deals set up that they've committed to, and we can talk about that. At the same time, those are all big numbers on infrastructure. They're scary, and they say, "Oh, bubble." Google released a stat recently that they have seen a 150x increase in the amount of tokens processed over the last 17 months.

On the one hand, you've got this crazy, scary-sounding buildout. On the other hand, you actually have a bunch of usage that's happening. So, are we in an AI bubble?

I do not believe we're in an AI bubble today. I was, depending on how you look at it, the privilege and the misfortune of being a tech investor during the year 2000 bubble, which was really a telecom bubble. I think it's really helpful to compare and contrast today with the year 2000.

First, I think Cisco peaked at 150 or 180 times trailing earnings. NVIDIA is at more like 40 times, so valuations are very different. Most important, however, is that the year 2000 internet bubble, or telecom bubble, was defined by something called dark fiber.

If you're a veteran of the year 2000, you'll know what that was. Dark fiber was literally fiber that was laid down in the ground and not lit up. Fiber is useless unless you have the optics, switches, and routers that you need on either side. I vividly remember companies like Level 3, Global Crossing, or WorldCom coming in and saying, "We laid 200,000 miles of dark fiber this quarter. This is so amazing. The internet's going to be so big. We can't wait to light these up."

At the peak of the bubble, 97% of the fiber that had been laid in America was dark. Contrast that with today: There are no dark GPUs. All you have to do is read any technical paper. One of the biggest problems in a training run is that GPUs are melting.

There's a very simple way to cut to the heart of all of this. It's the return on invested capital of the biggest spenders on GPUs, who are all public. Those companies, since they ramped up capex, have seen, call it, a 10-point increase in their ROICs. Thus far, the ROI on all the spending has been really positive.

It's a really interesting and open debate about whether or not it will continue to be positive with the quantum of spend we're going to have on Blackwell. I personally think it will, but there's no debate that thus far, the ROI on AI has been really positive. Valuation-wise, we're just not in a bubble.

I couldn't agree more. The other thing that I would say is you can contrast the actual adoption and usage of the technology from then. The internet was actually really hard because you had to build a two-sided network. You had to build websites, and then you had to get users. It's much more difficult in the case of the AI tools. All you have to do is light them up via API or turn on ChatGPT on your website, and everybody has access to them, right?

They're built on top of cloud computing, on top of the internet, and you can get to instant distribution—a billion people right away.

The other thing is the counterparties. You mentioned this: They happen to be the best companies in the history of the world, right? I think collectively, the people who are coming out of pocket and writing checks for this capex generate around $300 billion of free cash flow a year. Is that right, directionally?

Round numbers.

Yeah. And they have $500 billion of cash on the balance sheet. So whenever people are like, "Oh my God, it's a bubble. Is it going to pop?" I'm like, "I think it's kind of fine." It costs like $40 billion or $50 billion to light up 1 gigawatt.

Yeah, if you're on NVIDIA chips.

On NVIDIA chips.

Yeah.

So there's kind of an $800 billion buffer growing by $300 billion every year.

Free cash flow at some of them has begun—

Well, this goes to your point on return on invested capital. There is a little bit of a mismatch at the buildout. We should see that next down a little bit.

There is a little bit of a mismatch at the buildout.

But Larry Page apparently internally said, "I'm happy to go bankrupt rather than lose this race." I think that is the mentality for sure at Google and perhaps Meta. It's just seen as existential, and you have to win.

Okay. So lots has been written about these round-tripping deals. Give me the—because round-tripping is a very scary concept from the internet buildout. That was a big problem. What do you make of it here?

It is objectively happening. Money is fungible, so NVIDIA, if they sign a deal with OpenAI, can say, "Hey, you can't use our money to buy our chips," but money is fungible. It's happening at a very small scale.

I think what is driving this isn't the need to finance GPU or data center purchases, but it's actually competitive dynamics. NVIDIA's biggest competitor is not AMD, it's not Broadcom, it's certainly not Marvell, and it's not Intel. It's Google.

More specifically, it is Google because Google owns the TPU chip. This is by far, perhaps today, the only alternative to NVIDIA for training and maybe the best inference alternative. Google is a problematic competitor because they also own a company called DeepMind, and they have a product called Gemini.

I think you could argue that they are the leading AI company today. I think they've taken 15 or 20 points of traffic share in the last 2 or 3 months, and that does not include AI Overviews. I suspect, on an actual traffic basis, Google is bigger than OpenAI, Anthropic, or anyone today. That business is going to run on TPUs.

Then we have 3 other labs that are relevant today. There's Anthropic, and that's an Amazon and Google captive. Anthropic is really going to run on TPUs and Trainiums. So you're left with xAI and OpenAI at the forefront.

If Google is going to a lab like Anthropic and saying, "I'm going to help you fundraise and give you chips," I think, for competitive reasons, it's very hard for NVIDIA not to respond. As Jensen said, he thinks it's going to be a good investment. So I think the round-tripping concerns are pretty overblown.

What NVIDIA really needs is Meta to get their act together, or another American open-source player to emerge, or maybe some sort of détente with China and AI.

When people ask me about NVIDIA and all the moves and the round-tripping, my reaction is that everything they've done is completely rational.

100% rational.

Long term, sure, things they do may not have as high a return on capital as other things, but strategically, I think they're all kind of the right moves. Jensen's one of the 2 best CEOs, along with Elon, I have ever known. I think he's playing a strong hand really well.

Yeah. All right. You started getting into the model companies. Let's just talk about the model.

So, we can come back to chips, memory, and networking because I want to get your take on that. But since we're on the model side, what do you think happens with market structure? Who wins where? Who are you most optimistic about, and where do you have concerns?

I think humility is an important virtue for an investor. If we're going to make an analogy and say that ChatGPT is to AI as Netscape Navigator was to the internet, at this point in the internet boom, Google had not been founded. Mark Zuckerberg was in middle school. Travis Kalanick was in kindergarten, so it's just very early.

I think it's important to be humble about making high-confidence predictions at the application layer. It's one reason I think the infrastructure layer is often maybe a safe place to be at the beginning of one of these new technology waves. Well, actually, let's talk about the role they play at the infrastructure layer. There's a piece of them that obviously serves as an infrastructure layer, powering other application providers, and then they also have their own applications.

I would draw the distinction.

Yeah, that's most true of Google. But I just think it's hard to have high conviction other than to observe that the internet was a very disruptive innovation. I think there are reasonable arguments that AI could be a sustaining innovation because the raw ingredients—data, capital to buy compute, and distribution, which is what you need—all of today's biggest tech companies have in spades.

As long as they execute well, hire good people, and have a sound strategy, I think you could see it be a sustaining innovation for a lot of members of the Magnificent 7. On the other hand, I do think it's existential, and if you don't execute, IBM might be a good fate.

Yeah, that's tough. Data, distribution, compute, dollars, talent.

Yeah.

They have every right to win. It seems now more than before that they're taking it quite seriously.

Yeah, maybe Google in particular, but obviously Meta is making the dramatic moves they're making, too.

No, to me, ChatGPT was Pearl Harbor for Google, and we're going to see how they responded. They're slowly starting to respond.

Yeah. And then, what's your forecast for the platform piece of their business—the infrastructure piece? How do you think it shakes out in terms of business-model market structure? Do you think they end up as high-margin businesses like the cloud businesses or like aircraft manufacturers, or do you think they end up very competitive and low-margin businesses like airlines?

I don't think they will be airlines, but anybody can just look at the P&L of a SaaS company circa 2021 and 2022, and you see 80%–90% gross margins. The nature of AI, because of scaling laws and Richard Sutton's “The Bitter Lesson,” is just more compute-intensive, so their gross margins are structurally going to be lower. But that doesn't mean they can't be great businesses.

I just think it's going to be a long time before we see a truly AI lab, a frontier lab, with gross margins anywhere near SaaS or internet-era margins. Their OpEx can be a lot lower, and maybe that's how you square it, but the gross margins are fundamentally different. Until scaling laws change, and the importance of test-time compute and things like that change—which I don't see happening—they are going to be lower-margin.

Yeah. Okay. So, let's talk about the application layer. You just got into it a little bit with the SaaS businesses. I don't know if you've waded into this fight on Twitter, but every few months it comes up: SaaS is terrible, and it's dead, and it's all going to go away. Then, with Andrej's Dwarkesh interview he just did, the market's reacting positively to it. It's a whipsaw reaction. So what do you think happens with SaaS and software?

I think I first said, probably in early 2024, that I thought all of application SaaS might be a zero, different from infrastructure SaaS. I would say I have a more nuanced view now, and I think there could be some really big application SaaS winners, especially if you serve a more fragmented SMB customer base.

Google has made it really easy, if you're a customer of theirs, to use your data and essentially make any SaaS app you want, and then your data isn't shared with anyone else. But the critical mistake that I think a lot of retailers made in dealing with Amazon is they looked at Amazon's margins and said, “We don't want to be in that business.” That was obviously a terrible mistake.

Here we are 25 years later, and Amazon has really healthy retail margins. I worry that application SaaS companies are trying to preserve their existing gross-margin structures because they believe that if their gross margins go down, their stocks will go down. It is definitionally impossible, given what we just discussed, to succeed in AI without gross-margin pressure.

I don't know why they have concerns, because we have an existence proof in Microsoft and Adobe that a software company can deal well with declining margins. It used to be that companies were scared to go from on-premises to the cloud because margins were lower. Cloud margins are lower. They're still good.

Microsoft transitioned from on-premises perpetual licenses with maintenance to a cloud model, and it was a pretty good stock for 10 years. So if you're an application SaaS company, what I would say is: don't be scared, and look at declining gross margins as a mark of success rather than a badge of shame or something to be feared.

It's actually so funny you say that because whenever we have these discussions about companies, basically every company that comes to present to us is like, “We're an AI company.” We always look at the gross margins, and it's become a badge of honor for them to actually have low gross margins because, “Oh my God, people are actually using your AI stuff.”

Yeah.

But if you show up and you're like, “I'm an AI company,” and it's like, “I got 82% gross margins,” you're like, “I don't think anybody's really using it.” So, yeah, it's interesting. If you're one of these public companies, would you rather have $10 of revenue with 90% gross margins or $50 of revenue with 60% gross margins?

Not hard.

It's not that complicated. It's hard to do in the public market.

It's hard to do in public, but if you communicate it and draw parallels to the cloud transition, I'm an investor and I would be excited about it, and I don't think I'm alone in the world.

The big advantage these legacy application SaaS companies have is they do have these really profitable existing businesses. So you can run your new AI products at break-even and catch up to the leaders, and I'm just surprised more people have not done that.

Why are none of the public coding companies even trying to compete with Cursor? The reality is Cursor now has a trillion tokens, and there will be a point where they have enough coding tokens that it's tough to catch them. But I think today, if you're a public coding company and you said, “I'm going to lean in. I'm going to run it at break-even. I have an existing business. I'm going to attach it to everything,” hey, you have a chance. The prize is clearly really big. I see Martin is skeptical.

Martin's shaking his head. You have a chance.

I said a chance. I said a chance.

That's like Dumb and Dumber. You're telling me there's a chance, not like a real chance. You're telling me—

You're telling me there's a chance.

Yes, exactly. I totally agree. We actually saw it with Figma, for example. When they went out, they had extremely high gross margins, and they were like, “Hey, we're going to pretty aggressively distribute our AI tools, and our gross margins are going to go down.” Investors asked a few clarifying questions, and then they were like, “Oh, that actually would be a good thing.” So I'm surprised more people in the public markets aren't doing it. It worked out okay for them.

It's working out well—a long game to play. What about on the consumer side at the application layer? Obviously, Google was the portal to the internet, and it kind of still is. The whole business model was predicated on taking some intent and directing you to someone else's website, where they would do stuff with you.

It's kind of not going to be that way. It already isn't that way with AI. Although I tried the browser today and tried to do some pretty basic shopping stuff, it's still some work to do, but I think it will get there.

So what do you actually think happens with the market structure of the consumer internet companies? Do they get subsumed into a component of a chatbot interface, or do you think it's something else?

So, one: humility. Hard to say.

I would just say I think the AI companies that have launched these AI browsers may come to regret it. There’s something called Chrome that has, whatever it is, 5 billion users, and if you’re Google, you can just go look at what happened with Google Buzz. They’re very cautious. They’re currently in litigation with the government, and they could easily do this and probably do it even better, but they didn’t want to be first.

So now you have 2 AI-native companies with their own browsers. Let them run for 3 to 6 months, get a little head start, and then, wow, here we are: We had to do this. I don’t know how that’s going to work. Maybe for the companies other than Google that don’t own Chrome.

Yeah, I guess data and distribution are pretty powerful in that.

Yeah, hindsight’s 20/20. And the one thing I would say is I do think it’s tough to bet against the companies with large existing user bases today. I also think reasoning has fundamentally changed the economics of these frontier models. Pre-reasoning, I often said, if you are a frontier model without access to unique, valuable data and internet-scale distribution, you’re the fastest-depreciating asset in history.

I think reasoning really changed that because the way RL works during post-training, having a big user base now kind of unlocks that flywheel that was at the center of every great consumer internet company: You have a good product, you get a lot of users, the users make the algorithm better, the algorithm makes the product better, and it just spins. It’s not quite spinning yet in AI, but you can squint and see it. And so I think that fundamentally changes economics for Anthropic, for xAI, for OpenAI. But Mark Zuckerberg’s trying hard.

Yeah.

We’ll see.

Yeah. Yeah.

Yeah. A lot of smart people in there now.

Yeah, for sure. I think the worry is—and I think this is another interesting thing—is if you don’t—like, in a strange way, the Chinese open-source model ecosystem is a godsend to any American company that’s trying to catch those 4 leading labs. Because the problem is, if you don’t have Gemini 2.5 Pro or a later checkpoint of it, or a later checkpoint of Grok that we don’t see, or a later GPT checkpoint, when you’re training the next model, you’re at a disadvantage.

Oh, by the way, one thing I just want to say that drives me crazy is all these people who say that GPT-5 is the end of scaling laws. GPT-5 is a smaller model. It was not designed to be better. It was designed to be more economical for OpenAI and Microsoft to run it. Any reference to GPT-5 and its scaling laws is crazy. Sorry. Rant over.

We’ve got the pedestal up here if you want.

Yeah, exactly.

Shaking your hand.

Yeah, we could.

That’d be good. Do you want to talk about chips?

Sure.

So, okay, I know you love NVIDIA. Talk about your view of NVIDIA, AMD, TPUs, ASICs, and how you think the market structure shakes out there—the competitive advantages that the various players have.

I think it’s really a fight between NVIDIA and the Google TPU. Something that I don’t think is broadly appreciated is the extent to which Broadcom and AMD are effectively going to market together.

NVIDIA is no longer just a semiconductor company, as I’m sure you’ll hear from Jensen tomorrow. It was a semiconductor company, then a software company with CUDA, now a systems company with these rack-level solutions, and now arguably a data-center-level company with the level of architecting they’re doing with scale-up, scale-out, and scale-across networking. The networking, the fabric, and the software are all important.

What Broadcom is saying to companies like Meta is, “Hey, we will build you a fabric that can theoretically compete with NVIDIA’s fabric, which is a mixture of NVLink and either InfiniBand or Ethernet. We’ll build it on Ethernet. It’s going to be an open standard. And, hey, we’ll make you your version of a TPU, which, by the way, took Google 3 generations to get working. And you know what? If your ASIC isn’t good, you can just plug AMD right in.”

But I personally believe most of those ASICs are going to fail.

In the fullness of time, like over a period of time, or in the fullness of time?

In the next 3 years, I think you’ll see a bunch of high-profile ASIC programs canceled, especially if Google starts selling TPUs externally, which has been all over X. Who knows exactly how that would work? If you’re Anthropic, it’s rumored that Anthropic wants to buy tens of billions of TPUs. If you’re Anthropic, maybe you don’t want Google seeing your secret sauce, but there are ways around that.

So I think this is really a battle between Google and its TPU, enabled by Broadcom for now. Google can take the TPU away from Broadcom whenever it wants.

Yeah.

Now, they can’t do the Ethernet networking that Broadcom is doing, but they control the TPU. So it’s really Google and the TPU versus NVIDIA, with Amazon. That’s a very talented team, arguably the most talented silicon team at a hyperscaler—the Annapurna team. I think Trainium 3 will probably be a much better chip than Trainium 2. It took 3 generations to get the TPU right, and then AMD will always be kind of the second source. You need a second source.

All right, exciting. What do you think happens? Okay, so I want to go back to business models. One of the big things that is widely discussed as a source of disruption—and most of the CEOs in this room are CEOs of startups who are trying to go beat some incumbent or find some new market opportunity—is that the ripest opportunities tend to come when you have a big platform shift that is also accompanied by a business-model shift.

There are a couple of areas where I can see it in an obvious way. We’re investors in Decagon, customer support, so you can pretty easily see a business model that is priced on the resolution of a task because it’s so measurable. In coding, a lot of the business model has now shifted to consumption, and obviously, especially for developer-facing things, that’s comfortable and pretty well-known.

What about the rest of the industry? I feel like there’s sort of this hand-wavy thing that’s going on, which is, “We’re going to go get all of services,” but it’s like, okay, so how do you actually go do that? It’s going to be pretty hard. Do you have any prediction on how that plays out?

Well, I think what you’re seeing in customer service, which is kind of an easy first example, is where you have a lot of textual data that LLMs are good at. You can probably really easily run some RL to make sure that they get a good verifiable reward, with verifiable reward being a happy customer, first-call resolution, or whatever it is.

But I do think you will see that played out. Humans—we’re fundamentally paid based on outcomes, and a lot of AI will be augmenting humans, but probably also replacing some humans. That will involve being paid for outcomes.

Going back to the consumer business model, everybody’s talking about affiliate fees. For sure, I’m going to have my own AI. It will be a version of Grok because we’re both xAI shareholders. It will be a version of Grok that knows me and likes me.

When I want to go on vacation, it will know the hotels that I like to go to, and it’ll say, “Hey, 3 hotels. I have Gavin coming. Who’s got the best price and the best room?”

It’s going to massively upgrade the gifts that you give to Becky, just in case. Becky’s in the audience. She really appreciated your Dumb and Dumber reference, I’ll have you know.

But, yeah, and then there will probably be some sort of affiliate fee. Again, that’s just being paid for an outcome and kind of closing that loop, which will probably be a little bit of a business-model degradation.

Why did Google never start a marketplace? Because people systematically overvalue their ability, once they’ve acquired a customer through Google, to keep it as an organic customer. So they systematically overpay, and they continue doing that. That’s why Google never went to outcomes or a marketplace: Advertising leads advertisers to systematically overpay. So that inefficiency will be squeezed out, but, yeah, we’ll go to outcomes.

I think Elon tweeted today that work would become optional. Instead of buying your vegetables at a supermarket, you can grow your own garden if you want. Who knows how long it takes us to get there, but that doesn’t sound wildly implausible to me for how powerful this technology is.

And I was just struck by Karpathy, 2 days ago, being painted as a skeptic for saying AGI is 10 years away. Are you kidding?

Insane. 10 years.

Yeah. Yeah. Sign me up. We have shorter timelines, please.

Yeah. Well, that’s awesome. While we’re on the topic of very exciting futuristic things, robotics—do you have a view on—

Yeah, very real. And it’s going to be Tesla versus the Chinese in the same way it’s Tesla versus the Chinese in cars.

Electric cars. Yeah.

Yeah.

I would just say cars, not electric cars.

Yeah. Cars.

Yeah. Do you have a sense of the timeline?

You can all watch the Optimus videos. Every roboticist I know is extremely impressed. There’s a giant debate: Is it going to be humanoids or not humanoids? I think that debate is over because humanoids can learn from watching YouTube videos, and then it’s easier for a human being to put on a suit and show the robot how to do it. It’s kind of crazy to watch the video of all 50 Optimus robots doing 50 different tasks, and then it’s very simple: Did you put the glass in the dishwasher correctly or not?

This is so fun, Gavin. I always love chatting with you. Let’s give a hand to Gavin.

Thank you, David. Thank you.

All right. Next up, we have a very exciting panel on building out real-world infrastructure. But first, give us a few minutes. We have to do a quick stage change here. So, thank you.

Thanks, everybody. Thank you, man.