Grok 3、AI 记忆与语音、中国、DOGE、公开市场回撤｜BG2 与 Bill Gurley、Brad Gerstner

I witnessed, almost daily, people who are either in government or even friends of ours saying, “We have to win the AI war with China.” I don’t know what that means. I can’t imagine an end state where we control all the AI and they don’t have any. It’s already too late. They’re smart, and I think the reality is that we just need to focus on running our fastest race.

We need the Teslas, we need the OpenAIs, we need rockets that land themselves—we need all of this. But to think that they’re not going to have BYD building great cars, or DeepSeek building great models, or rocket companies that copy us and can land themselves, would be naive. It’s remarkably naive.

Bill, it’s good to be with you. Good to see you. We’re in this—wait, should we tell them that Steve Ballmer gave us this man cave? It’s a private cave in the—

The truth of the matter is, the hardest thing about this pod—I love this pod—is you and I getting our schedules to match and actually getting together. I’ve gotten a ton of feedback. I’ve seen people on Twitter asking, “When are you guys going to record the pod?”

First, thank you to the audience for encouraging us to do this, because I love doing it.

We would love to do it more. It’s just a little challenging to get together and do it. We have an ongoing dialogue pretty much 24/7 about the stuff going on in the world, and then occasionally we get to get together and share it with you all.

1. Grok 3

I thought maybe today, Bill, we’d kick it off with Grok 3. We’re now about 10 days out since Elon and his team unveiled, in record time, an unbelievable model. Maybe you can help us zero-base where you thought the model stood when it came out, where it stands in the rankings, and then we can have a conversation about the impact and what it means.

We’ve talked about this in the past, but everyone in the ecosystem was super impressed with how quickly they built the Memphis facility, exactly, and how big it was. It was the largest contiguous cluster in the world, and there was a lot of chatter about that ahead of time.

I can remember some of the investors saying, “This will prove that pretraining still has headroom, because this will be the biggest cluster ever trained on.”

Correct.

You can decide what your expectation was after that interpretation of the idea. The generic way of saying it is that Grok went right up near the top of all the benchmarks—on some, not on others. Some people argued about whether the reasoning component was real or whether they cheated or overtuned to a benchmark.

I don’t think it matters. The biggest positive takeaway is that there’s a new player in the model market. A lot of people had said this was a sport of kings and there were only going to be so many players. There’s a new one in the market that invested what it needed to invest, has access to capital, has a data asset that they argue is important and special, and was able to get to the front of the race. Let’s just call it that.

We’re looking at this Artificial Analysis chart that shows this clustering in the upper right. DeepSeek got up there a couple of weeks before Grok. What’s interesting is that they all seem to be coalescing, in an impressive way, around the top of these benchmarks.

When we say “they all,” we’re really only talking about 5 or 6 players who have a chance to be in this game at this point.

Correct.

I saw people interpret Grok’s fast rise as proof that pretraining still has legs. I had almost the opposite reaction, which is that they just slammed up against the ceiling that’s holding everyone in, although it’s still an incredibly capable model.

No doubt. I’ve said this for a while. I’ve been concerned that, given the way an LLM works and the way it’s optimized, building bigger clusters and adding more parameters won’t buy you much. Whether I said it or not, Ilya Sutskever said it, [name unclear] said it, and other people have said the same thing.

To me, this reinforced that point. I was expecting that if there were pretraining headroom, this would go through. I’ll qualify that: This was their first run. Maybe there were some tricks they didn’t know. They could very well back up and do another run on that same large cluster and shoot past these people, or maybe these benchmarks aren’t the exact right thing to be looking at.

I would say a couple of other things. Number one, it’s not just a pretrained model. They also have an inference-time reasoning component to the model that’s incredibly capable. We have this benchmark chart that I tweeted the other day, and I compared it to the search-index benchmarks that we all used to track.

The benchmarks are one thing, but the reality is: How do we feel when we’re using the product? What I will say is that Grok 3 rocketed to the top of all app downloads on the iPhone charts. At least my Twitter thread was full of people having great experiences and showing those experiences with Grok 3.

2. Grok’s Leverage of X Platform

It had a personality and an interaction with people that I think people were enjoying. Number one, it just has to be capable enough, and it clearly crossed the threshold of being capable enough. The real question now shifts to whether they can leverage the X platform, which reaches a massive and important audience, to really drive that.

The early indications, when you compare it, for example, to how Meta has used Meta AI—as incredible as I think Zuckerberg and Meta are and given the advancements they’ve made—I have not been particularly impressed by the productization of Meta AI. It’s basically just a search box stuck at the top of Instagram or stuck in my WhatsApp thread. When I’m on it, I never intend to be there. It’s not direct contact.

On X, they figured out that the first thing they would do is put that button at the bottom of the app, clearly distinguishing it as its own standalone application. They launched a standalone application, they’re using X to drive those app downloads, and now I opened my X app today and it said, “Go out and try the new voice for Grok 3.”

To me, the execution on the product side to drive consumer use has been pretty damn impressive, and it took them to the top of the charts.

Only DeepSeek and Grok, of all the others, have shown the ability to break into the top 10 on the App Store downloads.

3. AI Consumer Market & SEO

I think that while we all have a fascination with where they got to on the benchmarks, my own sense at this point is that this is going to be one of these battles, kind of like search was, where there are 5 or 6 players.

OpenAI is releasing ChatGPT 4.5 literally as we’re about ready to go on. They’ve hinted in this presentation at ChatGPT 5 or 6—I guess that was shown on a screen. If you look at 4.5, one of the important distinguishing elements they’re pitching is that it’s more humanlike and gives better, more concise answers. It’s not a big breakthrough on evals, although there are some improvements in the early looks against the evals.

Ultimately, I think we’re going to measure the success of these things by how many people are using them.

Let me say one thing that would be good for the audience. I know you’ve said it in the past, but you’re an investor in OpenAI, and I think you have a theory about their prowess and lead in the consumer market. Why don’t you reiterate that?

I’ve shown this chart before. In the search wars, we had Google, Yahoo, AltaVista, Lycos, Ask Jeeves, Excite, Infoseek, and others. They all did pretty damn well on the benchmarks, but the reality is that didn’t get them to any value creation, because ultimately all the consumers aggregated around Google.

The real question is whether that same pattern plays out—winner-take-most in consumer AI. It did in search and it did in social, but it’s not necessarily a follow-on that it will happen in AI. X has an incredible installed base that it can market into. Meta has an incredible installed base. Google has one, and it’s existential for those companies to market to those consumers.

I don’t think it’s going to be winner-take-all as much. I don’t think we’re going to see a 99% monopoly here, but I do expect that we’re going to see 70% or 80% share go to the winner.

If we look at the numbers today, I think Sara reported last week that OpenAI has crossed 400 million weekly active users. That’s a user number, not a paid-user number. The number of paid users is a fraction of that. I think they also reported something like $11 billion or $12 billion in expected revenue this year.

You can reverse-engineer your way into what percentage are paying for that, but more importantly, I think the monthly active user number must be somewhere in the order of magnitude of 700 million to 800 million monthly active users. You and I have followed consumer for a long time, and there’s this magic number around 1 billion. I already think they’re nearing escape velocity, but at 1 billion monthlies, you can funnel all of those people into weeklies, and then funnel the weeklies into paying subscribers or people who are consuming advertising.

What I’ve seen is everybody else catching up on the benchmarks. What I have not seen is people catching up on the consumer velocity.

Let’s handicap some of the other players. Who do you think is closest from a user standpoint? Is it probably Gemini, if you count the Gemini searches within Google Search to get to a number that’s close to OpenAI?

Let’s start with Google. There have been a lot of reports over the last couple of weeks from public companies reporting that their Google organic clicks are down 20% to 40% year to date.

The question is why. Why are Google’s clicks down so much? If I do a Google search today on my phone, half of the page is taken up with an AI answer to whatever my query is, and the rest are all paid links.

I think that’s the right decision for Google to make. If you want to compete, you ultimately have to be willing to take the innovator’s dilemma head-on and cannibalize your product with AI. If they do that—and if you’re one of these people who thinks SEO wasn’t already dead, which I would have declared it dead a while ago—SEO is really dead.

Those free links were the core product that used to attract everybody to Google. The idea that SEO is basically gone is pretty remarkable.

I’ve been remarkably frustrated with Google’s organic links for the past 5 years. You go in and search for your favorite team’s schedule, and all the ticket sites are up front. The link you’re looking for is buried, and you have to hunt for it.

Let’s talk about that for a second. The obscure link or obscure information that we may be looking for may be on page 3, 4, 5, or 10. You and I are never going to get to page 3, 4, or 5.

What’s so interesting about OpenAI’s Deep Research is that if I launch a query using Deep Research, it will go to page 4, 5, 10, or 100 and find those obscure pieces of information. I don’t want to go do that deep research myself, so I think the evolution of Google actually provides acceleration to the Deep Research projects.

Google has a massive installed base. The number of people going there who have the inertia to continue going there will carry them for a long time. But I’ll say this: You can search for this on Twitter or anywhere else, and certainly with my own behavior, the amount of activity I used to do on Google has been 80% cannibalized by ChatGPT because there’s search embedded within ChatGPT. I’m getting all of that information and all of those answers.

I think Google is going to be formidable. I think they’re being bolder than they’ve been, but they’ll have to continue to do that. We’ve talked about this, but I think some of their assets are remarkable. You have the YouTube data set and all the search queries over all the years, as well as their understanding of structured data around a lot of the consumer verticals.

They built that out in airlines and other areas. They should be able to do those agent-type queries better and faster. Their velocity on product has not been impressive, and their velocity on consumer has not been impressive.

They’ve had these assets for a long time, Bill. They had ChatGPT before ChatGPT. They also have Android, which is a massive asset, and they have their own browser. Both Perplexity and OpenAI have started toying with the idea of either having a browser or, in the case of Operator, using a browser in the cloud to go do this work.

Google has so much. I still think they have a bit of the innovator’s dilemma, in that they still have to try to maintain those paid links on the page.

This chart here plots Google’s paid-click growth against OpenAI’s weekly average user growth. It’s not going in the right direction.

It benefits from the fact that informational searches are what ChatGPT cannibalized first, not commerce searches, which is where most of the money is in the paid links.

Although, again, we’re going to see this from X and from everybody else: The entire domain of the internet is the domain of agents. If you think about Operator as one of the first agents rolled out by OpenAI, what does Operator do? It mimics me as a human going out and researching a hotel, booking a hotel, or whatever else I might do on the internet.

We’re in a very embryonic state. I agree with you that we’re not there yet, but it’s very clear what the roadmap is going to be. It’s going to want your credentials, and whether you give it your credentials or not is going to matter, because it’s searching against an—

Let’s talk about Meta. I know that—

Actually, one last thing on Google. There was a point when Facebook went public at $40. We had some exposure, so I was paying attention. Zuckerberg, as he has many times, got woken up on mobile.

Everyone thought he was dead because he had also built in HTML5 and didn’t believe in native apps. There was a whole thing that they weren’t going to be able to monetize mobile. He was on the cover of Barron’s magazine—the cover of Newsweek magazine—and it was, “Meta’s dead,” or “Facebook’s dead.”

But he woke up and fixed it. Can Google do that here? Is that possible? What would it look like, and what would it take?

I’ve said publicly that Google’s moat was not a technological moat with search. Their moat was a distribution moat. Their moat was a mind-share moat. We googled everything when we wanted to know anything.

The only thing that could attack Google was never anything head-on. It had to be an orthogonal attack from something that was 10 times or 100 times better, because it gave us answers instead of blue links. That’s why it was such a mortal sin for Google to ever allow anybody else to go first. The only thing that could give you a trillion dollars’ worth of free mind share was going first with something that was 100 times better.

That’s exactly what ChatGPT did at the end of 2022.

Go to Meta.

Meta has 3 billion users of its products. I think it has products that are tailor-made for chat-oriented AI, whether it’s Instagram, with shopping agents and co-shopping agents, or WhatsApp, with a bunch of agents living inside my WhatsApp channel. It feels natively much better positioned for AI, and we know Zuckerberg is in complete beast mode.

I am surprised, though. We’re now about 18 months into the Llama thing, and the manifestation of it into the product has been slower than I expected.

Back to your product point, exactly. Meta hasn’t—and I will say, even though we know he was irked by DeepSeek blindsiding Llama with the release of R1, I would say it’s not just a product issue for them.

I heard from several inference players that we’re friends with that all of a sudden, DeepSeek rather than Llama is the enterprise open-source model of choice that everybody’s experimenting with and playing with. That becomes a real problem for Meta as well.

I think 2025 is a critical year. I think they will come through. Remember, when it comes to almost all product stuff—whether it’s Stories copying Snapchat or Reels catching up with TikTok—they’ve always shown up to the party late, but they are grinders and they always deliver the product.

It’ll be interesting to see what they do. Who else would be on the list? Anthropic has really not been—

No, they pretty much ceded the game on consumer. There was a product announcement yesterday that they’re going to be powering Alexa, but now we’re stretching. The idea that Amazon did a big Alexa launch yesterday is pretty late in this game.

Alexa occupies a different space in most consumer minds. It is not what ChatGPT does. To dislodge something with the momentum ChatGPT has, you have to go at them and do better than what they do at the thing they do.

This is why X is so interesting to me. They have a platform that is the number-one news platform in every country on the planet. The people who are most actively engaged use this platform, and they go there for information. They go there for answers, and they go there to engage.

I think it’s an audience that’s very well suited for AI. The integration they’ve done is as good as anything. They’ve done this in a very short period of time. Everything from the logo, to some tweets where the logo will pop up and summarize or do more research—I’m really impressed by the velocity of not only catching up on the benchmarks, but also catching up on the consumer-product side.

They’re number one on the App Store, and that stands for something. They came out of nowhere, and people said Elon couldn’t do this. I never doubted that they would catch up on the benchmarks if they got a big enough cluster, because Elon set a mission that people become messianic about. His engineering capability to build out the cluster and do all those things was never the question for me.

The real question was: Can he close the gap in the consumer race?

The odds-on favorite there has to be OpenAI. I think they continue to widen their gap. I think they’re accelerating at scale, but that’s where the race is. It may very well be that coming in second place, with 20% share, is a pretty good place to be.

I want to mention one more company, and then I’m going to make a guess at 4 ways someone could try to win this game. The one I want to mention first is Perplexity.

I’ll give them credit for being product-centric and innovative in ways that the others haven’t, and for doing it on their own terms. They don’t have near the usage of OpenAI, so there’s a question about whether they look like an acquisition candidate. I don’t know if anyone can agree on the price, but for one of these other players that hasn’t been as successful from a product standpoint, you could imagine Microsoft buying Perplexity and now having a consumer brand to go battle it out.

You and I just had this conversation. You could imagine a world in which Microsoft were to buy Perplexity and then have a consumer brand to go battle it out. We know how much Satya wants to win in consumer. Now he owns a bunch of OpenAI, so he has some potential channel conflict there.

The bigger issue is that with Elon out there, you’re probably more likely to be able to do a deal like that. But when founders are raising at an $89 billion valuation, it becomes a much more difficult decision for a company like Microsoft. I’m not saying it couldn’t happen.

4. AI Memory

When it comes to punching up—being innovative, scrappy, and having product velocity—the founder there, Aravind, and the team have been super impressive to watch. I think they made other people better. But the numbers, as we look at them today, are really powerful but still much smaller.

I have 4 things I’m watching that could potentially lead either to further lock-in by OpenAI or to a window for someone else to do something. I’ve mentioned some of them before.

Memory is still this thing that could tie you to something. OpenAI has probably done more with memory than anyone else, but no one has really gotten to the place where I’m telling it to remember things, store things, and create lists, where it starts to become an executive assistant for you. I haven’t seen that yet. I still think that’s a dimension that could be really important.

Voice is another. We’ve talked about it, and they’re all playing with it. Voice also ties in with device type. If the voice were spectacular, I might not have to carry the phone around as much.

You need an earbud.

You need an earbud. The third one is nebulous, but someone could focus on a feature that no one has focused on to date. Right now, the game looks like everyone is running to the same place, whether it’s benchmarks or voice. That’s not an easy thing to say, but it would have to be really out of the box.

The fourth thing I’ve been thinking about is that no one has really thought about a network effect. I wonder how you could make the quality of the AI experience a function of your user base.

Let me give you an example of a network effect that I think is happening around model improvement. If you have 700 million or 800 million monthly active users, the diversity of information in their questions, answers, follow-ups, and so on is much higher. That data is now being fed back into the models to improve them.

Some users may have seen this—I know I have. You get 2 answers, and OpenAI asks you to rate them. I think that’s an example of OpenAI very actively attempting to build network effects in terms of the quality of the model and the quality of the answers.

There could be a more intense form of network effect if you found a way to leverage the user base as part of the value proposition.

Let me go back to your first point, memory. You and I have talked about this a lot. If you get memory, the switching costs explode.

Exactly.

5. AI Voice

I would argue that not only do the switching costs explode, but the conversion rate from free to paid probably also goes up, simply because of the value delivered.

I was with my 89-year-old mother last Sunday. My mom has wanted to write a story of her life for a long time, but the reality is she’s never going to sit down and write the story of her life.

When I’m with her, podcast-style, I’ll ask her questions and record it on my phone so I have it and can perhaps go back to it later. Then I started thinking about it and said, “I don’t need to be the interviewer. Advanced Voice Mode could be the interviewer.”

So I was sitting there with her last weekend, and here’s the prompt I gave Advanced Voice Mode:

“I’m sitting with my 89-year-old mother tonight, who wants to write her life story. I want you to interview her about her life, asking questions about her childhood, having kids, working, growing up in the Depression, her love of computers, and travel. Remember everything you talk about, and then compose a story of her life that her grandchildren would like to read.”

Advanced Voice Mode just started asking her questions. My mom was really nervous at the start, but then a little tear welled up in her eye because she realized, “Oh, my God, this could be a massive unlock.”

Here’s the thing: Advanced Voice Mode and ChatGPT already have memory. You can already do these things. The problem is the nature of the product. You don’t know that it can do those things.

Part of the challenge of designing a product where the prompt is your way in is that you have to help people imagine what’s possible. You and I could have imagined, in the age of the internet, somebody building an internet website that just did that thing.

I think that’s one of the challenges all these companies face: the innovation around the top of the funnel, and creating a prompt that can help people better get into it.

I’ll give you another example: deep reasoning, which is really fascinating. They basically took the o3 series of models and fine-tuned them end-to-end based upon all these browser interactions. But the more specific the prompt, the better the Deep Research report is going to be.

A lot of people are using o1 to help them build sophisticated prompts that they then feed into Deep Research. I think there’s something there where we’re effectively using AI to get us to the point where we’re better at prompting.

One of the ways it will be very simple: Once I have this assistant and I’m having an interaction, I just say, “Assistant, my mom wants to tell her life story. I’m not sure how to go about doing that. Do you have any ideas?” It would say, “Yes, just use this prompt.”

6. Future AI Assets

I also think there are other assets that could play a role, like a contact database and email.

Yes.

I can even imagine moving my email to one that’s integrated inside, because contacts are a great one. If it just cleaned up your contacts, knew my contacts, knew your contacts, sent an email, sent a text—there’s a lot there.

For anyone who works with content, there’s some app or repository. All my writing and everything I’ve done for the past 10 years has been in Quip, but some people use Notion.

Yes. Where does that story land? Where is it stored once you’ve done it? Do you have to take it out of OpenAI, or would you rather just have a place for it?

Correct. Now you have Projects in OpenAI and other things.

This brings me to a point. When you think about these research labs and look at the number of people who work there, the fact that we even call them research labs is interesting. You and I haven’t called Google a research lab. Nobody called Google a research lab. It was a company. It had product teams, marketing teams, finance teams, and so on.

A lot of these people came out of research, so they’re still very small teams, heavily tilted toward building toward the benchmark. OpenAI now has thousands of people. I know Kevin Weil, who runs the product team over there, and you look at all these companies: If you’re going to win this race, you have to do all the things great product teams do.

You have to build all the things you’re talking about. You have to be thoughtful, growth-hack, and get customers to use the product more and more. That’s hard.

One thing came out this week that I don’t know whether it was intentional or not. Someone published OpenAI’s internal forecast, which included, I think, losing $20 billion in 2025 and another $20 billion in 2026.

Someone said to me, “Why would they publish that? Why would they show that?” To me, there’s an information war out there trying to scare capital in or out. To lay down a statement that if you want to be in this game—and keep in mind, there’s a variable cost every time you serve a Deep Research query—if people think this is winner-take-all, just like we had with Uber and Lyft, they’re going to go hard at trying to win.

You probably need to be willing to lose $20 billion a year to step into this game. X looks like it has the potential to raise that kind of money. I don’t know whether Microsoft or Amazon are prepared to lose that amount of money incrementally.

This is such a fascinating segue. There’s one company we didn’t even mention: Apple. When we went through all this, we didn’t mention Apple at all. That’s pretty shocking.

Why didn’t we mention them?

Apple has self-selected out of the race. They’re not building a big model. They’ve been very public about thinking that they can be a late mover here.

They did the Apple Intelligence integration with ChatGPT, and now they’re going to do one with Gemini. As I shared with an Apple executive the other day, here’s the only integration that matters: My ChatGPT app on the front page of my iPhone.

That’s mean.

It’s the truth. I just don’t use any of the integrated features on the phone, which I think creates vulnerabilities for Apple.

This is the first time they’ve been faced, I think, with this level of product risk. They have so much lock-in around this device, but somebody else—Huawei, for example—could build a device and perhaps ship better AI phones around the world. They’re not going to be able to ship them into the United States.

We’ll see what happens with the Google ruling at the end of the year. If Google is no longer allowed to be the default search app because of this consent decree, do they really turn Android into the thing that it potentially could be?

There’s a lot of potential risk.

We didn’t talk about Apple. I’d say the other company we didn’t really talk about, because we’re so focused on the United States, is China.

When you look outside the United States, you really have to look to China. I would say the acceleration and velocity of AI in China is off the charts. We’ve talked a lot over the last few weeks about DeepSeek clearly coming out of left field and very efficiently building a frontier-quality open-source model.

Most people have quietly ignored probably the company that’s the leader in AI in China, and that’s ByteDance. Their AI equivalent of ChatGPT is number one in China, and they’ve been using AI to drive TikTok globally for a very long period of time.

I know you have strong opinions on this. It seems to me that the United States has underestimated China in AI. Now we’re at this inflection point where a lot of people say, “They must be smuggling GPUs into China,” or something like that.

The reality is that China is going to have frontier AI, and almost all the things we do to try to slow them down and stop them are backfiring on the United States.

I couldn’t agree more. I witness almost daily people who are either in government or friends of ours saying, “We have to win the AI war with China.” I don’t know what that means. I can’t imagine an end state where we control all the AI and they don’t have any.

It’s already too late. They’re as smart as possible, they’re innovating, and you look at all the other products they’re crushing it in.

I just don’t understand it. I think the reality is that we need to focus on running our fastest race. We need the Teslas, we need OpenAI, we need rockets that land themselves—we need all of this.

But to think that they’re not going to have BYD building great cars, DeepSeek building great models, or rocket companies that copy us and can land themselves would be naive. It’s remarkably naive.

It’s going to lead to people making decisions that either slow us down ourselves a lot—much of the AI regulation would do that—or provoke them in ways that aren’t helpful. It’s not going to slow them down.

7. Regulatory Challenges

Let me give one example, and then I want to move on to talking about the arms race, if you will. During the Biden administration, the Commerce Department passed something called the “AI Diffusion Rule,” which we’ve mentioned on this pod before.

It created a convoluted set of rules by which U.S. semiconductor companies could export outside the United States. This wasn’t exporting to China; we already have export restrictions with respect to China. It created all these tiers and classifications for how much you could distribute and whether you had to distribute it through a hyperscaler.

The whole idea was somehow to prevent these chips from getting to China, but what it really does is cause us to compete globally with Huawei with one hand tied behind our back. It almost guarantees a Huawei-led Belt and Road Initiative around the world, and the world is going to run on Huawei AI chips. That gives them the demand they need to build a frontier AI chip.

Again, it may have been well-intentioned by the Biden administration, but it totally backfires. Hopefully Howard Lutnick and this administration will throw that out and start over.

There are a remarkable number of people in Washington, on both sides of the aisle, who have a perspective about China and use words like “enemy” and “threat” and “we have to win the AI war.” Those terms are so loaded.

I think they believe they can achieve something.

8. AI CapEx and Investing Dynamics

If I owned NVIDIA, my number-one concern would be excessive regulation coming out of Washington. My number-one concern.

Let’s shift gears for a second. You talked about OpenAI losing $20 billion a year. I’m not going to share anything I shouldn’t share, but I think we always have to keep in mind the difference between operating expense and capital expense.

There’s a variable cost of serving a ChatGPT query. I would posit that those variable expenses are not very high at maturity, although an o1 Pro search or Deep Research could cost 20, 40, or 50 times more than an ordinary query.

Correct.

I would posit that you’ll be able to come up with a variable expense structure using the right mix of models that will produce great margins. They may not be as high as search was for Google, but they’ll still be great margins.

What people are conflating is when you decide to spend $20 billion a year to build out Stargate, build out clusters, and do all these things. A component of that is the capex needed to serve inference, and a component is capex to build future products.

If we’re looking at Facebook, Google, or Microsoft, Microsoft is spending, I think, 80% of its free cash flow on capex. We don’t quote that as its profitability. It has its net income, and then it has net income less capex.

I would keep that in mind. These companies are very committed to continuing to invest aggressively in a future they see as big. But we heard likely Satya Nadella on the Dwarkesh Patel podcast, in what many are characterizing as a pushback against these high levels of spending.

I think of my fleet even as a ratio of the AI-accelerated storage to compute. At scale, you’ve got to grow it.

That infrastructure need for the world is just going to grow exponentially. It’s manna from heaven to have these AI workloads, because they’re more hungry for more compute—not just for training, but now we know for test time as well.

Here’s an interesting thing: When you think of an AI agent, it turns out that an AI agent is going to exponentially increase compute usage. You’re no longer bound by just one human invoking a program; it’s one human invoking programs that invoke lots more programs.

That’s going to create massive demand and scale for compute infrastructure. Our hyperscale business, our Azure business, and the businesses of other hyperscalers—that’s a big thing. On the podcast, likely Satya reiterated that Microsoft is going to spend $80 billion this year and more next year, but there’s not a world in which they’re going to have unlimited, unconstrained spending.

This week, it was rumored that Meta is out shopping for a data-center campus. The rumored amount is $200 billion, capable of building 6 to 8 gigawatts. That sounds a lot like Stargate, which is in that 6-to-8-gigawatt range.

Microsoft, I think, has 5 gigawatts installed and is probably going to build a worldwide—what, 5 gigawatts worldwide? Is that what you mean?

Correct. They’re going to build more.

Again, it seems to me that if you want to be in the group of 5 or 6, that’s the calling card. You have to have either a business or the ability to raise capital such that you can deploy enough to build out that level of compute.

In the case of OpenAI, Masayoshi Son is rumored to be leading a very big, $40 billion round, with a lot of people involved. We saw them announce it at the White House.

It is important. Many people interpreted Satya’s comments as a tapping of the brakes.

Yes.

He said he was happy that some of these were leases. I don’t know any other way to interpret that.

There are 2 ways you could interpret it. One is that he’s telling you, “I’m hedged against this being overbuilt.” The other is that he’s better off canceling a lease than sitting on infrastructure.

I would say it even a little bit more directly. Likely Satya said last June—we talked about it on this pod—that it was very likely that at some point there would be a supply-and-demand mismatch, and you had to build a resilient company that could go through a zone of disillusionment.

He basically said the reckoning is coming at some point. Now he goes on Dwarkesh’s podcast and sounds like he’s tapping the brakes a little bit.

I think the interpretation should not be that he doesn’t believe in AI. He very much believes in AI, but he’s running a public company and has made commitments to his shareholders. He’s saying, “I need to see a certain amount of inference revenue in real time to justify that level of capex.”

Everyone believes in AI. The amount of spending and capex is something we’ve never seen before. That’s why I’ve said it’s better than watching Succession. This is a massive sport of kings.

Some of these things—whether it’s the $20 billion losses or Satya saying he’s glad he has leases—might be part of an information war, with other players trying to talk capital in or out. It’s a high-stakes game, and it’s fun to watch.

Business-model resiliency is going to be critical here. What do I mean by that? It means liquidity.

We know there was a zone of disillusionment in the internet. We know there was one in social, and we know there was one in cloud—a period when prices and spending got ahead of revenue. Given the level of competition, some people describe it as a prisoner’s dilemma.

In the case of Google and Meta, they literally have a printing press in the back room spitting out billion-dollar bills, so they’re resilient. Microsoft is resilient. In the case of OpenAI, they have to raise money, so you need to have a big stack behind you.

In the case of X, they need to be able to raise capital. Elon is obviously the wealthiest person on the planet, and he can sell shares and do some things. But I think the most powerful thing Elon has is a global belief in him as an entrepreneur, which gives him an opportunity to raise capital from sovereigns around the world.

If you asked whether this is still an open sport, I’d say no way. I don’t know anybody other than Elon and Sam who can play that game at this point, although DeepSeek surprised everybody.

If you’re going to play that game, remember that DeepSeek spent more than the amount reported in its last training run. More importantly, to serve an explosive amount of inference, they would have to spend a lot of money to build—

I want to make a point that we’ll probably come back to much later. When you have a scenario with this much ambition, this much competition, and this much capex as part of the game, it’s easy to lose sight of the microeconomics. It’s easy to lose sight of the unit economics.

If you’re Anthropic and you’ve got training credits over here and capex over there, are you thinking about depreciation when you say, “This is profitable”? Is that how you price your API product?

You’ve got this razor’s-edge pricing dynamic that I’ve never seen before. Explain what you mean.

The price difference between today’s model and yesterday’s model is 20 times.

So it’s a fast-depreciating asset. The second you’re off the frontier—

Yes. It’s dangerous. These are all traps, and it makes this fascinating.

Maybe we can transition to the public markets a bit. There’s a lot of talk that we’re going to see a CoreWeave filing, and I’m excited to see the numbers and piece together more of the information.

There’s a rumor that CoreWeave is going to file for an IPO, so you’ll be able to see the numbers.

I just want to underscore the point you made, because there is some rhyming to Masa coming back into the scene. Masa is one of the greats of this industry over the last 25 years, but people would also describe him as somebody who’s a bit of a gambler and places gigantic bets.

Some people would say he’s a total visionary, and other people would say he’s just not price-discriminating. But clearly he’s shoving all in with OpenAI. I don’t think he knows any other way of operating.

The point is that we’re at a moment where the danger for a company of getting this volume of capital is that it’s hard to focus on building the muscle and grit and ingenuity required to drive unit economics.

Think about what Elon had to do at Tesla. Capital was hard to come by, so he had to figure out how to make money on every damn car. How do I take costs out of manufacturing at every single stage of production?

9. Government Spending + DOGE

When you have excess capital, you lose that discipline. You don’t build that muscle. I think it’s an important admonition for the board and leadership at OpenAI and all these companies: It’s one thing to invest aggressively in the future, but you better make sure that along the way your unit economics work.

You’ve been thinking a lot about DOGE and what it means for the capital markets. It’s interesting to even say “if it happens,” because as I watch the press every day, there’s an equal number of people saying, “This is going to take out all these costs,” and others saying, “They’re just saying things, but they’re not actually going to happen.”

You and I talked about this on our pod around February 6. When you asked me about the markets, I said we had peak political uncertainty, because we have a lot of things changing. We have peak economic uncertainty—not just because of DOGE, but because we have tariffs and other things. We also have peak technology uncertainty. It’s hard to predict the future: What software company is going to be worth what in 5 years?

That causes discount rates to go up and multiples to come down. I said I was surprised by how resilient the market was in the face of all this uncertainty.

Now I would argue we’re starting to see a few cracks in that. If you look at this chart, Bill, it’s the Nasdaq since the election. It ran way up; the Nasdaq was as high as 10% post-election. Now we’ve come off 4 or 5 points from that high, but we’re still 4 or 5 points higher than we were on election night.

One thing I’ve been thinking and talking a lot about is the difference between stimulus and austerity. Over the last 3 or 4 years, we had massive stimulus in the economy. You and I both supported that in March and April of 2020, when we were in the depths of COVID. You had to prevent the economy from coming to a screeching halt.

The Fed went all in and Congress went all in to save the economy. But we were also very critical that the Fed moved way too slowly, that the second stimulus package was way too large, and that it led to the runaway inflation we saw. Inflation hit 9%.

The one thing all of that monetary liquidity did to the system was cause risk assets to go up in value. Now we’re in a period where we’re not talking about adding $1.5 trillion of liquidity to the system; we’re talking about pulling $1.5 trillion out.

Last year, we had $56 billion of tariffs imposed on other countries. That’s the amount of revenue we collected from tariffs. We’re talking about that going to $500 billion, a 10-times increase.

We know that some of those costs will be eaten by producers. The company producing something in China will just take a lower margin. But we know a lot of those costs will be felt by us consumers, who will end up paying higher prices for a Dell computer because Dell passes along the price increase for a computer made in Mexico, as an example.

That’s $500 billion. On the other hand, there’s DOGE. There’s no doubt in my mind at this point—and we’ll show the chart of the likely spending cuts—that they’re not only making big cuts, but the president just said last week that he wants Elon to be more aggressive.

They sent an email to every employee that said, “Respond to us or you’ll be deemed to have resigned.” They’re giving them more shots on goal, but the message is very clear. I think there’s going to be a downsizing of the federal government to the tune of 40% or 50%.

A lot of people have been giving DOGE a lot of grief, but I remind you—and I tweeted this the other day—that Bill Clinton did DOGE in the late 1990s. I don’t know the exact percentage of federal employees they let go; it was between 10% and 20%. But we had a balanced budget in 3 fiscal years and a $230 billion surplus.

It was helped by the internet, but now we’re going to be helped by AI. I think you can see some replay of that. It does mean that we’re probably going to take $500 billion to $1 trillion out of federal spending over the course of the next couple of years.

All I’m suggesting is that austerity has the reverse impact of liquidity from the government into the system. If you go back to the GDP calculation in macroeconomics—C plus I plus G—G is the amount of money the government is spending. The amount of money the government is spending is going down.

Tariffs are a headwind to the economy, and this austerity from the government is another headwind. I’m 100% in agreement that this is the short-term shock therapy we need to get our fiscal house in order. But you have to think of it like somebody saying, “You’re out of shape and you’re going to have a heart attack. You have to take this medicine, endure this short-term pain, work out every day, and get fit to avoid the heart attack.”

We need to get fit to avoid bankruptcy. All I’m suggesting is that it might affect the markets. My risk profile is lower than our standard risk profile. I own half as much as I would normally own at a given point in time.

Do I think that’s because the future is bleak? No. I believe aggressively in the future. But I think we’re going to have to take a little short-term pain, which means we could see a random, run-of-the-mill 10% to 15% drawdown in the market while the market gets its head around the fact that the economy is going to grow a little slower.

When the economy grows a little slower, companies grow a little slower. When they grow slower, earnings go down and the multiple goes down.

When Elon went into Twitter, one of the stories that came out was that they found software licenses for a whole bunch of people who weren’t using them, and they cut those licenses dramatically.

Do you anticipate that one of the outcomes of DOGE will be a headwind for a bunch of companies that have sold software or services to the government?

100%. There’s just no way around it. If you go from 3 million federal employees to 1.5 million federal employees, you don’t need as many licenses and you don’t have as much cloud consumption.

If you think about the multiplier, you take the federal employee’s salary, then add healthcare, benefits, pension, and everything else. Then you have all the ancillary spending.

I won’t name the exact company, but I talked to an airline the other day. At this airline, the number of government tickets sold year to date is down 50% already. That’s a 50% impact.

Yeah, because they said, “We don’t want you traveling. We want you in the office every day,” and all this other stuff. And so this airline has already been impacted. I think everybody in the ecosystem—if you have revenue line items, if you’re a business, if you’re a public company, and you have revenue line items from the federal government—it’s not just that the rate of growth is going to slow. It’s that they’re actually going to be negative on a year-on-year basis.

Now, again, I happen to think this is generally a good sacrifice for us to make. Those are our tax dollars. There is no government money; this is our money that’s being consumed. But I don’t think the public markets or investors generally, and certainly not Silicon Valley, have gotten their head around what this means. Now, what’s the flip side to this?

Well, in fact, I would say, ironically, this happens quite a bit in our world, but Silicon Valley and venture capitalists have just gotten comfortable with backing companies that sell to government.

Exactly. We see a lot of that. Interesting timing. People get excited about—well, you saw what happened to Palantir stock the other day when the president directed his cabinet member, the Secretary of Defense, to find 8% cuts in the Department of Defense every year.

Yeah.

Right? And so this austerity, again, is real. Now, that probably means we’re going to have a rotation of money out of the less technologically innovative folks into the more technologically innovative folks. But Trump went further. He suggested to Xi that China and America should both cut their military budgets in half.

Yes.

Now, maybe that’s provocative. Maybe that would be an amazing thought. That was extraordinary. Back to this idea: We’ve both blown up the world many times over. There’s a certain camp of folks—and I think likely Mearsheimer is in this camp—which is great-power politics: You just have to build, build, and build, and eventually you’re going to have a war, or something like this. Or maybe the fact that you have these stockpiles deters the ultimate war.

One thing that is just fascinating: I’ve never heard an American president in my lifetime suggest that he wanted to sit down at a table with China and Russia and talk about how they could collectively cut their military spending in half. Just from an entrepreneur perspective, it caused me to stop in my tracks and be like, “Hmm, that’s an interesting idea.” I thought it was a cool thing he thought of.

That’s an interesting idea.

Well, I will tell you, back on the public markets, the other interesting thing here: Warren Buffett just put out his annual letter. He’s going to have his annual meeting coming up here. He has a $400 billion cash stockpile and has been liquidating stocks, right? His biggest stockpile in ages. Stan Druckenmiller, Howard Marks, and Steve Cohen came out over the weekend and said, “I’m nervous about the markets,” for the same reason that we were talking about a month ago.

So I think there is a growing chorus of players now. What’s the flip side to this? Well, since Trump’s been elected, the cost of a mortgage or credit card, et cetera, is starting to come down. Why is that? Right, there are 2 reasons. The first reason, I think, is because we’re saying, “Okay, the economy is going to slow a little bit.” And if the economy slows, equities as an investment are a little less positive relative to a bond, so you rotate that into cash. And when the cash is sitting on the sideline, it’s invested in a U.S. Treasury.

Just to put it in perspective, the only anecdote I really ever hear about this is, “Well, China doesn’t want to own our Treasuries anymore.” China buys 3% of our Treasuries annually. It’s tiny. They used to buy—10 years ago, they bought 12% of our Treasuries, and everybody panicked that they were too big a buyer. So what I see is just the opposite: Every sovereign around the world and every domestic investor who’s starting to put more money into cash, who’s hedging a little bit—all of that’s going into U.S. Treasuries.

I just think that one should brace over the next 3 months. I think these tariffs are very real, they’re structural, and the president has committed to them. I think, number 2, the reconciliation package is now rolling, and I think they are very committed to balancing the budget within this president’s term. The only way you balance the budget is a trillion dollars has to come out of spending. Remember, the 2019 baseline was about $5 trillion in spending; the COVID high was $7 trillion. We’ve got to get that back down to at least $6 trillion, probably to $5.75 trillion, if you’re going to balance the budget. That means a trillion out in a year. That’s austerity, and that’s going to be a headwind to the economy, but it’s the right thing to do.

10. Golden State Warriors

Okay, that’s a tough note to end on, so I’ll switch to something more positive. I got invited to the Golden State Warriors game on Tuesday night. The Butler trade looks like it’s working. It’s incredible—6–1, I think, since the trade.

I happened to get an invite to the banner ceremony and dinner afterward for a good friend, Andre Iguodala. Steph gave an incredible speech, and I had Andre speak at our investor day maybe 2 years ago.

Yeah, I remember.

And 2 things Steph said that really stood out to me about Andre. Number 1, he said, “There is no this without Andre.” By “this,” he explained to me, he said, “He came at a moment in time. Even his decision to come to the Warriors made us believe in ourselves. Then he came here, and he did whatever it took.”

The second thing he said is, “Andre Iguodala always put excellence over ego.” He never pouted on the bench when he came off the floor. He was the first to get guys fired up. Steph talked about Game 6 in Boston. I remember that game; I was at that game. I remember Andre—he must have played 5 minutes in that game—and he was so fired up and really willed all the players to up their game. So I was so happy for him.

Yes. You know, our good friend Jason Chang and I—I never bet on sports. I never bet on sports. He talked me into it. We were at a Warriors game during the losing streak, and the odds were so great that they weren’t going to win it all. He talked me into placing a bet on them winning it all. At the time, it was like 40-to-1 against them, right? And all of a sudden, they’re on this 6-game winning streak. They trade for Jimmy Butler, and they may win this whole thing. Fingers crossed. It now has me with a focused mind.

Great to see you, great to be with you. Take care.