China Open-Source, Compute Arms Race, Reordering Global Trade | BG2 w/ Bill Gurley and Brad Gerstner

15% on all goods coming from Europe, 0% on US goods going to Europe. So, an opening up of European markets, paying us 15%, and on top of that getting commitments like $750 billion—almost a trillion dollars—of energy purchases from the US. Or look at Japan, which they announced last week—another huge market. Again, similarly, they're going to pay tariffs to the United States, with no tariffs imposed on the United States, and they're going to invest $550 billion into the US in a way the president gets to direct.

I just think we need to give the president credit where credit is due. Everybody said this was going to lead to retaliation, that trade wars were going to be disastrous for the US, and all we've seen so far is deals, deals, deals, deals.

I have to say, if this was the CEO of one of our companies—let's say we had a board meeting at the start of the year and he outlined these plans, and we said, “Hey, we're really nervous about this. This is a high-risk, high-reward strategy. It's either going to backfire and we're going to fire you, or it's going to work really well and we give you a bonus.”—if we're measuring him halfway through the year, I would say that he's in line for a bonus.

All right, the summer pods are back in action. Good to see you guys. We have our good friend Sunny Madra, the COO of Groq, joining from—I don't know, Sunny, it looks like some fancy hotel in Saudi Arabia in the middle of the night. Good to see you.

I am borrowing Mitch Lasky's incredible podcast setup in our Woodside office with the nice high-end DSLR camera.

Nice. You're looking good.

Sunny, of course, our good buddy Sunny, is the COO of Groq. Probably a few hundred million in revenue, maybe doubling year over year. I just saw something: you're rumored to be raising $600 million at a $6 billion valuation. Maybe that has something to do with you being over in Saudi Arabia. Of course, you're hosting all the open-source models in your inference clouds around the world. Is that about right?

Sunny Madra

Yeah, you got it right. You touched on all the key points. We don't comment on speculation, but you touched on some good points.

Well, it's great to see you in D.C. last week. Our good friend David Sacks is really on a heater. First, it was the crypto summit a couple of weeks ago. Of course, the GENIUS Act got passed, which really teed up these stablecoins, and now the CLARITY Act around market structures is making its way through Congress.

Then, of course, last week was the AI summit, where the president laid out a multipronged strategic plan for American AI that both extends the government's investment and leadership, but also accelerates the distribution of the American AI stack around the world. Both of those things—the crypto and the AI summit that he put together—I thought are key contributions really to the next generation of American technology leadership around the world. It's amazing to see that much progress in 6 months. Congrats to David Sacks and the rest of the team. Michael Kratsios, Dean Ball, Sriram Krishnan—they really all brought the heat.

1. Open-Source Models in China

Sunny Madra

Yeah, for sure. Of course, the AI Action Plan was focused on maintaining global AI leadership, particularly over China. I hate to say it, but we have really been our own worst enemy. It seems like excess regulation on everything from energy production to model development to semiconductor chip distribution has really been a bit of an unforced error by the US over the course of the last 24 months and handed a lot of momentum to China. I think it really threatened our leadership.

So this is kind of a 180 to get America back on track. We've underestimated Huawei and Chinese AI development—model development, really—at every step of the way. I want to kick off today talking about recent developments with the base models and reasoning models coming out of China because we had this freakout moment earlier in the year with DeepSeek that we all remember. NVIDIA stock plummets. Everybody in Washington's talking about it, but since then, people kind of forgot about DeepSeek.

The reality is China's been on a roll. They're dominating the global landscape for open-source models. We've seen 6 or 7 high-quality open-source model providers, many of the fastest-growing in the world. And this at a time when American open source—Llama 4—has been sputtering a bit, really losing its mojo around the world.

So Qwen, the open-source model out of Alibaba, has passed, I think, 400 million downloads. Of course, that's released, like these other models, under the Apache 2.0 open-source license. So, very open, as Bill's talked about.

But, Sunny, you tweeted—and one of the reasons I wanted to get you on the pod this week is you tweeted—that all of these open-source models are really coming together in China. They're leveraging one another. They can distill and generate synthetic data on each other's work. You went so far as to suggest this might allow them to pass the best proprietary models coming out of the US yet this year, maybe by Q4 of this year.

So why don't we dig in there? What is your theory of the case? Why is China doing so well in open source? And should US model companies like OpenAI and Anthropic be concerned?

Sunny Madra

Yeah. So let's tie it into, I think, 3 important things that we see happening. The first one being, you know, the Chinese—and the president addressed this at the AI summit, right? He addressed the point around using copyrighted work, and he used a great example: if you read a book and you use it, you're not violating the copyright there.

He addressed that concern, and that was one of the major things that a lot of people didn't talk about, but I think it's important for the model makers. The Chinese have just been able to work around that because of their position on IP.

What we're really seeing here—and I think Bill teed it up even better off of my tweet—is that they're able to compound. So what you're seeing very quickly is both the open-source nature and the open-weights nature allowing them to basically compound on each other. Instead of working in silos and having to create giant training clusters separately, they can basically take each other's work and build on top of it.

Almost consider it a remix of someone's model. K2 is sort of a well-known remix of what DeepSeek had done, and now we're starting to see that happen really fast. We're seeing 2 dimensions of it going quickly. One, we're seeing the leading-edge models get quicker, and then we're seeing them distill down into smaller turbo models really, really fast as well.

There was a release today of a Qwen 30-billion-parameter model, which is performing as well as GPT-4o. So think about that, right? GPT-4o was world-class not that long ago. Those are the reasons that we're really seeing an acceleration right now.

I want to dig into this model development in particular. A year ago, we were talking about these models being stochastic parrots, and we really had to compress the entire internet. So you go back to GPT-4, and you're compressing the entire internet.

But now we really don't need to do it because we've trained them to use tools like the internet, right? They're true reasoning engines. When I ask a question today, it doesn't just spit out an answer immediately. It goes and uses the tool and searches the internet.

So if you don't have to compress all this Wikipedia information—take a subject like World War II—you just need to know how to go out and use the internet to find the information and summarize it in real time. How has that changed the pace of progress and the balance between open and closed?

Sunny Madra

Yeah. And so it's spot-on, right, Brad? What we see now—and you see it when you use these reasoning models—and what I suggest everyone do is, when you're using a reasoning model, you can usually expand out its thought process.

When you ask a question, it'll say, “Oh, the person is asking a question about this. What should I do? Let me go and maybe search the internet. Let me go do a few different things.” It can have a lot of different tools, and so the push has been towards really, really strong reasoning models.

We have to give credit there: OpenAI really started that with o1. That was really the first reasoning model that was put out there. But I think, on the back of the research and what everyone talked about—and the pod's good friend Noam Brown was the leader on that program—everyone's been able to look at that and say, “Let's reframe the problem.”

This allows us to build stronger reasoning models that don't have to compress, like you said, all the internet's information. Once they're coupled with strong tools, you start getting these really, really incredible results that don't even just show up in benchmarks, because none of the benchmarks really allow you to use a tool to answer the results. And if they did, we're going to see a whole bunch of new results there.

Hey, Bill, in many ways I think this validates what you were arguing over the last 6 to 12 months. You said this was likely to happen. Everybody knows you're one of the biggest proponents of open source in the world.

And you were telling people, you know, the Chinese are going to use open source to their advantage, and now we're seeing that in a really profound way. Why is China so successful here, and what can we learn from them?

Well, we've talked about it in the past. I won't dwell on it, but China got excited about open source about 20 years ago. It's not a new thing that's happened. You can imagine when most of the world accuses you of IP theft, embracing something like Linux and all the other open-source products seems very appealing, right?

I think it became kind of a common way of operating within China. It's a country that hasn't prioritized IP protection the way we have around patents. I could make an argument that there's way more prosperity if ideas are shared instead of protected. But I don't know—the one thing I don't know is, in the current AI situation, was the government promoting open source and encouraging it, or did it just develop through competitive forces?

But now you have a scenario where these companies are—first of all, there are new ones popping up. I almost feel like an idiot when Kimi comes out of Moonshot, and this week—I don't even know how to pronounce it—Zhipu releases a model, and I go on PitchBook and look it up, and they've already raised $1.4 billion. It shouldn't have been a secret, but I didn't know about it. All of a sudden, they're in the leaderboards on OpenRouter, and I'm like, “Oh my God, they're coming out of everywhere.”

What it shows you is that when you have a competitive dynamic where every single player—and I think there may be 7 or 8 deep-pocketed players with open models in China—they all learn from each other extremely fast. In this case, unlike software, you can use one model to distill the other and make it better. So it's almost like an accelerated form of that, and you just get massive, quick co-evolution.

I came up with a little analogy for people. I'll try and do it quickly, but imagine you had 2 communities. They're both farming communities, and let's say there are 10 to 20 farms in each. In one community, they come into the farmers' market once a week and just compete by selling their products, but then they go back. In the other community, when they come into the farmers' market, in addition to competing and selling, they're forced—I don't know who would force them, but they're forced—to share all their best practices from that week with everybody.

Everyone does it, and everyone shares their best practices. If you ran that exercise over 2 years or whatever, obviously the community where the best practices are shared across all farms is going to have a higher global output for the community than the one where you've just got proprietary ideas driving the individual farms, and competition without the idea sharing. Even me saying that may cause some people to scream, “That's socialism,” or they may not understand open source or how it works or why.

I do think you end up with a higher fitness level for a community that's behaving that way overall. You may end up with a lot less chance of a breakout monopolist like we've had in many of the sectors in American technology.

Well, let's just assume the Chinese government is, in fact, encouraging this in whatever ways, right? If you look at the release of the AI Action Plan last week, the Trump administration had a section—

They did.

—which is about encouraging open-source and open-weight models in the US, saying that these could become standards in some businesses and academic workloads. It's important they're built on the American AI stack.

As an aside, the Chinese quickly followed the American AI plan. I think they released theirs a couple of days later, where they called for the establishment of a global AI cooperation organization, which I thought, again, was interesting.

So, Bill, how do you feel about this? We haven't seen that much traction in the US labs on open source. Obviously, Llama has probably been the market leader there, but this is for both of you. Handicap for me, if you will, how you think this plays out. First, Sunny, maybe you start. What do you see at Groq? Do you see a lot of demand for these Chinese open-source models? And if so, what would it take for an American open-source model to catch up?

Sunny Madra

Yeah, one of the things that we should pull up is the chart of intelligence to price. One of the things that you see with the leading open-source models now, which are the Chinese, is 90% of the quality in terms of intelligence, but at a 90% price discount. Whenever you offer that to anybody, you're going to see people want to use that, whether it's individual developers or enterprises. So we're seeing that.

Let me interrupt there real quick. If you're looking at this chart, in the top right of that chart, you'll see a cluster of these Chinese open-source companies. The vertical axis here is really intelligence, and the horizontal axis, from left to right, is the cost per million tokens. You really want to be in the top right of that model: high intelligence, low cost.

What this chart shows is that, to Sunny's point, you can get 90% of the intelligence right for 10% or 20% of the cost. The result, I assume, Sunny, is that you're seeing huge demand at Groq and in Saudi Arabia, where you are right now, for these Chinese open-source models.

Sunny Madra

We are. Taking that forward, what do people want? They want some accountability. That's what you'll get. That's what you sort of got out of Linux and, say, Red Hat, right?

As great as Linux was—and Bill was touching on it—the majority of the enterprise was using a distribution which they could go and point to someone if they needed something. I think the world wants models that they can get from companies that they can go to.

So, to answer your question, what happens? I think if we look at Q4 this year or Q1 next year, I'd be willing to say a top-three worldwide model will be a US-based open-source model. We've got 2 big efforts happening there. We know we have the OpenAI open-source effort, which a lot of people have been working on at OpenAI, and even Sam has commented on that and its release later this summer.

Then we have all the efforts by Meta. If you combine both of those things together, I don't think you end up with something further down the list in terms of intelligence and/or price.

One thing I wanted to highlight about the China situation that I think might inform the US situation: I was having a conversation with this extremely young AI entrepreneur that I know, and he was poring over the Zhipu paper. He asked me for some information, so I went on PitchBook and sent him who had funded it.

2. Future of American Open-Source Models

He asked me, “Why is Alibaba funding all these things when they've got their own model?” They're in several of the other competitive plays, and it reminded me of a lot of the points that I've made about open source: if you're not confident you're going to win on offense, you want to play defense.

For any large tech company, commoditizing a potential threat is actually quite valuable. You look at what Facebook did with the Open Compute Project inside of its data centers. It may very well be that Alibaba just wants to make sure there's no ByteDance equivalent in the AI space. That would be pretty rational.

The reason I think that's an interesting data point when you think about the US is that there are several big tech companies that seem not to be on the bleeding edge of AI. You've got Microsoft—maybe, I mean, they have access to OpenAI right now, but they might lose that or whatever. You've got Amazon, and you've got Apple.

If I'm at any of those companies, I'd be funding an open-source competitor rather than funding, like Amazon did with Anthropic. I think you're in a much better position to encourage open source.

But Bill, are there a bunch of open-source startup models in the US?

That's where I was going next. I think you're going to see new entrants pop up that try to co-evolve with the Chinese models. Is Linux American? Is Linux Chinese? No one thinks of it as having a domicile, right?

This is just me predicting, but I think you're going to see—just like you saw Kimi and Zhipu pop up—I wouldn't be shocked if you see other new entrants pop up that are trying to be sanctioned or cleaned-up versions of these things. Use Red Hat as an example, Sunny.

If you start with access to those Chinese models, it wouldn't take you long to move into a nearby position, and you wouldn't have to spend the kind of money the foundational model companies have. Then you may see a big fight around regulatory capture, where someone tries to say that's not allowed or whatnot.

But I do expect to see that. And if I'm the Mistral team, if you're not distilling on these Chinese models, I don't know what you're doing right now, but I don't have any data on that front.

OpenAI is rumored to be launching its open-source model any day. Sunny, what would they have to do? So, to your point, you predicted—go back to the intelligence and pricing chart, right? If OpenAI was in the top right of that chart, i.e., if they were able to deliver something at the intelligence of, let's call it, Qwen, and also deliver it to market at 20% of the cost, it would seem to me that actors around the world—certainly, as part of the American AI Action Plan, we'd want everybody in the world to use that model.

Do you believe they have a shot at outcompeting and being in the upper right by the end of the year? And if so, do you think that will be the outcome? These companies that are using Qwen on Groq, do you think they would prefer to use OpenAI so long as it was equally capable and equally price-performant?

Sunny Madra

The 2 things that we see are brand and being U.S.-domiciled, or having someone that they can kind of point at who wins. If that shows up, it will win, because if you're a company, at some point you have to get your teams to sign off on what it is that you're using, what the risks associated with it are, and who is liable if something goes wrong.

I sort of feel like with OpenAI's release, and if Meta charges back, or even if some of these startups emerge that we can point at, I think we'll see a huge shift back toward those models versus the Chinese ones.

Yep. And we really don't know at this point, unless you guys have some inside knowledge, when Meta makes its second push with all these hires that they've made, whether they're going to remain committed to open source or even be more open. We really don't know yet at this point.

Sunny Madra

Right. You've certainly seen some of those rumors out there. I've seen rumors on Twitter that they were debating whether or not they should back away from open source. My hunch is that that's a misread. My hunch is that they're going to stay very committed to open source, but they may complement it with a proprietary model.

That would be my guess, as opposed to scrapping open source altogether. And Brad, can I throw one thing out there? I think part of the pitch to get everyone there is that it's open, because everyone that they're pulling over is coming from closed places. If you're really passionate about the work you're doing and you're passionate about where this is going to go, there's only 1 company that can fund that to that scale and do it open, and it's those guys. So I think it's part of the pitch.

Yeah. Some of the researchers have a religious belief in it, which was evident in the interview with the DeepSeek founder. It's oddly higher on their Maslow hierarchy than the money.

Sunny Madra

Of course they're getting the money.

Of course it's both. This is, I think, a really important point I want to come back to, Sunny: you're seeing massive demand for these Chinese open-source models today precisely because enterprises around the world can utilize them. They have 90% of the capabilities at 10% or 20% of the cost. It turns out that if you deliver something really powerful and really cheap, that's more important to these players than being aligned with American values.

But if you gave them something that was super powerful and super cheap and aligned with Western values, that would be the winning formula for an American OpenAI model to top the distribution leaderboards around the world. Is that what I'm thinking?

Sunny Madra

I think, yeah, on a place like OpenRouter, where you can see where this is happening, we would see it rise to the top in a few days.

That's music to David Sacks's ears because clearly, in the strategic plan, they're worried about Chinese open-source models dominating globally. If you just watch the pace of releases, the quality of the releases out of China, and the cycle time on the innovation in the open-source community out of China, it's faster and better at the moment.

We may see some of these new startups, Bill; we may see a reboot out of Meta, but the one that has, I think, everybody really holding their breath and hoping that we see something really capable and powerful is this open-source model that's been promised out of OpenAI. Now, of course, Elon says he's committed to open source as well. Grok 4 is a great model; it's impressive what they put out there. Any idea, Sunny, about the open-source plans out of xAI?

3. Compute Arms Race

Sunny Madra

Yeah. So I think Elon's been pretty clear on Twitter that they'll always open-source 1 generation back. So while they were on Grok 3, they should have gotten to Grok 2, and now they're on Grok 4. I think the thinking is that they will get there.

My only guess would be that right now, if they were to open-source Grok 2 or even Grok 3, it's so far behind that what's the purpose in doing it? It may not even be utilized, and you'll maybe end up having to deal with just a bunch of internet or Twitter FUD.

Just coming back to OpenAI, I will tell you it's one of those things that you rarely see in the enterprise. It's almost like the demand for a Tesla Roadster or something, or a Model Y before it came out. Everybody asked for it. That's the model that everybody wants to use right now.

We can't wait till it comes live, and once it's on Groq, once it's all over the world, I think it's going to be a real big one for everyone.

While we have you, Sunny, there's really just been this explosion in the compute arms race, right? There was a big debate when you were on the pod a year ago. We had, with Bill, the question of whether we'd topped out on compute demand, whether we were entering an overbuild à la Cisco 2000, and it's really been remarkable. I think now that's very clear.

Just a couple of tweets from Elon and Sam Altman in the last few weeks on this compute demand have really caught my attention. If you look at this one from Elon, talking about the xAI goal being 50 million units of H100 equivalent, Clark Tang on my team tweeted something that broke that down, which showed that that reflected something like 4 million total GPUs and an energy footprint of roughly 11 gigawatts, right?

And then They talk about their deal down in Abilene for 4.5 gigawatts, and the fact that they're going to come in well above the $500 billion estimate that they had promised to the government. So these compute clusters that are now being talked about as being built over the next 5 years are massively bigger than what we were even talking about a year ago.

I think they reflect this move toward inference-time reasoning, agent-to-agent interaction, and reasoning engines—you know, Jensen's comment on the pod last year that inference was going to 1 billion X, and the consequence of what we were going to need in terms of compute power to power all that.

Maybe just reflect: you're in the middle of all this, you're building out your own inference clouds around the world. Is this a lot of hyperbole and chest-pounding, or do you actually see the dollars going into the ground in places like Saudi Arabia and around the U.S.?

Sunny Madra

Yeah, I'm going to just quantify it with Google for a second.

In the Mary Meeker BOND deck, they have a slide there that shows Google went from 5 trillion tokens a month to 480 trillion tokens a month. They had just put some press out that they crossed 800 trillion, and I saw something today: they crossed a quadrillion. I had to look that up. That's 1,000 trillion.

So in the course of a year and a few months, they've gone from 5 trillion to 1,000 trillion. So that's 200x right there. I mean, that just shows it to you without having to look at anything else.

Every single search query on the planet today is now an inference transaction, correct?

Sunny Madra

And so you see it in Anthropic, with their continued fundraisers going through the roof. That's happening because they're seeing the amount of token consumption. Anywhere we lay infrastructure, we fire up a rack and it becomes fully consumed within a few hours.

You're talking Groq?

Sunny Madra

So you have demand far outstripping supply even at Groq.

Yep, we do. So, Bill, you see these fundraising announcements that are being discussed. Just CNBC's reporting tonight: I think Iconic is going to lead a $5 billion round into Anthropic at $170 billion.

In the case of Anthropic, it's $170 billion on rumored $5 billion in revenue. xAI has been rumored to be raising at $150–$200 billion. You've never, in the history of venture, seen fundraisers like this.

One of the topics being hotly debated on Twitter is that there's massive intervening dilution in these rounds because of the employee option grants or the employee RSUs that need to be granted to keep the employees in these businesses.

Just observing this a little bit from afar, I don't think you guys are direct investors in any of these major labs. What do you see? What do you observe? What are your warning signs about the size and the demand that you see in these rounds?

Well, I've never seen anything like it. I saw, through the Uber and Lyft situation, a precursor to this, but these dollars are even bigger.

And the amount of money that these companies are willing to lose in a year. I still think it's particularly interesting that Google has to compete with OpenAI because OpenAI is going to lose $7 billion this year, and Google won't—they would never allow themselves to do that. This started back in that previous era where, for the first time ever, you saw private companies have a competitive advantage in that they can be more risk-seeking with capital than public companies are allowed to be.

But I think in the past 12 months, we've seen OpenAI, Meta, and certainly xAI move into this place. I call them the cost-is-no-object, or CNO, group, where they're just putting out press release after press release and opening data center after data center. There are other people, I think—you look at Microsoft choosing not to extend its capex budget; you look at that Amazon example when we spoke to the Levant brothers at Codeium, where they're not keeping up with the NVIDIA purchases relative to their AWS share.

There are a few companies that are back on their feet, and there are a few companies that are really pushing the gas pedal. There are a few in the middle, and I can't tell whether Anthropic has the audacity and the means to raise enough money to start building data centers themselves. They haven't so far.

But it's a sport of kings. There has never been this amount of money spent right now. NVIDIA and others building in the stack, like Dell, which we spoke to a few weeks ago—they're the winners in a pick-and-shovel game that's got this amount of aggressiveness. Maybe SK hynix, too. I don't know.

Sunny, going back to the well-worn cliché that every shortage ultimately ends in a glut, do you have any evidence on the horizon where you see supply outstripping demand?

Sunny Madra

No. I was going to ask this to Bill as he was just saying it. Bill, on a daily basis, are you consuming more tokens, or are you consuming more traditional web lookups? I'd be willing to guess you're consuming more tokens.

Oh, it's insane.

Sunny Madra

And tokens are increasing. When you're using those reasoning models, you don't see all the tokens, right? They don't publish it, but there are 10 to 100 times more than in your very first AI search.

Yeah, for sure. Absolutely. Go ahead. You finish.

Sunny Madra

Even yesterday, Anthropic had to put this press release plus a product change out, saying, “Hey, we've got to throttle everybody,” right? This is because we have all these people using way too much of our services. So I think if you have intelligent models and you have the capacity for it, it's one of those things: people are consuming Jevons' paradox, or whatever. Sorry. Go ahead, Bill.

No, I was just going to offer one caveat to this super-exciting line of thought, which is that, because of the amount of venture capital out there, companies are not pricing the cost. None of the model companies—I don't think anyone, even in the verticals—no one's pricing the cost because they're pricing to take market share. You and I, Sunny, had a previous discussion about unlimited pricing, and inference has variable costs. So is that even sustainable?

And even when people say to me, “Oh, well, we're going to run out of power,” you wouldn't run out of power if you just took the price up. The thing that throttles demand is price, but no one here is raising prices. Anthropic doesn't need to throttle; they just raise prices, but they're not willing to do that because they're afraid to lose share.

Everyone's pricing to share, which means they're pricing under cost. There are rumors that even some of the best-known brands in AI have negative gross margins. I don't know when that settles out, but that will create a bump in the supply-demand curve if that ever has to be fixed. But for now, it doesn't.

Can I ask a question there?

Yeah.

Going back to the point that we made on open source, if you know in the back of your mind that there's something that's 90% as good but 90% cheaper, how does that factor in? We've also never had that factor as we're going through this growth curve. Since almost the beginning of the pod, I've routinely highlighted that the steepness of that price curve—as it becomes less cutting-edge—is something I've never seen before. I've never ever seen it.

I'm sure that a lot of people sit around and say, “Well, it's okay if I'm losing money here because, 6 months from now, I'll just use the older model.” We also talked in the past about how, in the internet age, all the startups began with Oracle and Sun, and eventually they all moved to Linux and MySQL. So there was a “we've got to win at all costs” phase, and then there was a phase where you started worrying about cost and optimization.

One day, we'll likely make that move. A few of the companies I've talked to that are running inference at scale are already starting to think that way. They're looking at it from that lens.

Well, and I think that's why Groq and Cerebras are doing so well. But, Sunny, give us an example. I would imagine that the Windsurfs of the world and the Cursors of the world, and all these folks who are building these coding agents, have massive demand for their applications, but they're paying through the nose to Anthropic or to these underlying proprietary model providers to be able to do that. What's the dynamic that you see there? Do you see them running to implement Qwen or some of these cheaper models?

Sunny Madra

Yeah, without getting into specifics of any one of them, multiple folks are building their own models based off open source.

Right, so they could just go distill any one of these models.

Sunny Madra

Correct.

Right. And given that they're these very lenient Apache licenses, they can eliminate the entire cost of Sonnet that sits under it for 70% of use cases.

Yeah. And like Bill said, turn that into a premium offering, right? The gold, silver, and bronze are built off something that's, like I said, 1/10 the price.

That seems to me, Bill, if I had to forecast, that if I'm OpenAI, I'm running a consumer business with really high gross margins because consumers are less sensitive to what they're paying; their intensity of use is lower. Whereas if somebody's writing code, the variable intensity is high.

For them, it would make sense to launch an open-source model, back to where we started the pod, and price it really low to drive share. It reminds me a little bit of Amazon back in the day. Amazon had this monopoly retail business they could use to subsidize AWS, gain share for a decade, and then begin to take price. That would be a rational strategy for OpenAI to follow: take the profitable consumer business and use it to subsidize the market share in other applications that you hope to build.

Certainly a reasonable strategy. I'm not inside that company. You have way more knowledge than I do, but I pay $200 a month and I do every one of my searches on GPT-4.5, and I'm probably negative, I would think. Yeah.

And so I do think there'll be some rationalization where these models kind of self-pick which one they're running based on what you need.

Move more to a consumption logic. Yeah.

Move more to a consumption logic. They're already doing that in parts of their enterprise business. As they've transitioned to more of this consumption logic, I think it's led to some real unlocks for the business.

I think that makes it harder if you're in the lab game and you don't have a consumer product and you don't own an application that you can drive high gross margins. I think then it gets back to this question: How long can you run your business for share, hoping that someday? They all exist at the beneficence of the capital markets, and the capital markets are willing to provide an incredible amount of capital to these businesses today.

They sure are. But you and I have lived through these periods where that disappears quickly. Can I put something there just to hear your feedback on it, guys? Look at Google and the TPU. Google is clearly, by these numbers that we're seeing, putting out more tokens than anyone else, right?

Do you guys believe they have a strategic advantage because they have their own hardware? They're not having to pay an 80% margin on something that they can generate tokens with. How do you guys look at that business and say, clearly, they look to be the largest—at least openly saying—the largest processor of tokens?

Sunny Madra

I think the key data point in that case—which I don't have the data, but I'd be glad to repeat it if someone shared it with us—is how many non-Google applications are running on the TPUs. How many third-party customers are using them? Because what I've heard, or what the general perception is, is that most of their proprietary TPU transactions are their own applications.

Yep. But that's probably where most of their tokens are being processed anyway at this point, Bill, right? Like transcribing YouTube videos, and in Google Meet, you can turn it on, and all the searches.

I was just inferring from your question—maybe I shouldn't have been—that they'll have an advantage for Google Cloud. And in order for that to be true, they need to have this crossover moment.

One quick thing, Brad, on OpenAI: I've been writing a book, which I've talked about frequently, and I've been quite—although I guess there are some privacy things now you need to be worried about—quite open with OpenAI about the book and doing research along the way. It knows a tremendous amount about my book right now, and I can ask follow-up questions without having to put the whole book back in the prompt again because of that.

And so I continue to believe that OpenAI's most likely chance for long-term success comes from switching costs and lock-in more than it will come from staying on the edge of the model race, because I think—

And the pricing and the pricing power that comes with that brand dominance, right? No doubt. Because the fact of the matter is, you said you're paying $200, you're getting more than $200 in value. I don't know what the price is, but I know that if it was variable, you would pay a hell of a lot more money to use that service.

I would. I would, but the lock-in—once one of these systems starts to truly understand you and have all your historic knowledge—I think the switching cost will be very high at that moment in time.

Sonny, back to your question about the TPU and the advantage of that vertical integration to Google. I think it's too early to know. What I would tell you is that they've absolutely made some changes, I think, over the course of the last 3 to 4 months to accelerate the business. You've seen the news about OpenAI leveraging TPUs for some of the inference demand that they have.

Ultimately, what I've said all along about Google is they're, in many ways, the best-positioned company in the world, but a lot of their advantage right now is no longer much of an advantage, right? Namely, they were the dominant place where the consumer started every single query, and we know today that's just not true when people are looking for answers. Bill's book is not in Google; Bill's book is in ChatGPT, and that's the—

Actually, it's in Google Docs, but you're right. The knowledge of it—

The knowledge of it, the interaction, and the token generation. So my only point is this: the battle for the consumer is ultimately where the value occurs, Sonny, not who runs what hardware. And so Google's dominance—it's been the greatest business in the history of capitalism for 20 years because they owned the consumer. They owned the verb in something that was extraordinarily high-margin.

And all I would say is the first real threat in 20 years came about because, in the ChatGPT moment, it has continued to accelerate. I think ChatGPT will cross 1 billion weeklies, maybe this year. Probably this year, I would guess.

And so that, to me, has always been the case for OpenAI. And when you look at the rest of these frontier labs, the case you've made throughout this pod about 7 of these models in China being able to open-source, distill off one another, drive up intelligence, and drive down cost, what that tells me is the model layer is being increasingly commoditized and that there's not going to be a lot of intrinsic value in that intelligence layer, that operating layer.

Sunny Madra

You're going to have to build applications that guys like Bill Gurley and you and I are using every day. And that's where the battle is, on the consumer side. You're going to have the exact same battle when it comes to coding agents and general enterprise applications. And I've said there, I think it's going to be more of a heterogeneous world. I think there are going to be lots of players that compete. I think the margins will be lower.

In that world, it may very well be that the tide is going up so much here, right? The whole world is transforming so much around this that you're still going to have lots of players who do incredibly well.

I have to say I'm surprised. If you would have told any of us a year ago or 18 months ago, right, that the combined enterprise value of OpenAI and Anthropic together would be over half a trillion dollars—and you throw xAI in there—it'd be a trillion dollars across the 3 of them, roughly. It's bigger and faster, and the compute demand is higher than any of us, I think, anticipated.

4. China, Tariffs, and Reordering of Global Trade

Sunny, hopefully we can keep you on here for a bit. We're just going to wrap up with a topic that I think has really dominated the conversation in the markets over the course of last year, and that's been about tariffs and the reordering of global trade. Bill, I know you had some questions and some thoughts about it, and I'm happy to dig in and talk about it as well.

Well, I would really just love to hear from you, Brad. The markets got very nervous when the—what was it, Liberation Day?—when the unpredictability of how big some of the numbers were and what that might mean, and whether we were walking away from the notion of comparative advantage. I think the markets got spooked, and you turn around and look at where the markets are today, and we've really gone through an evolution of how Wall Street is interpreting both the initial launch of the tariffs and the reality of where they're landing.

So how would you describe that? And why do you think the markets are getting very comfortable with where they're landing? I mean, not only comfortable, we're at all-time highs—

And on April 2, I was going on CNBC saying the Nuclear Navarro was going to be a disaster, and I'm out, right? And that's the amazing thing about this administration: there's really a team of rivals within the White House. You had Bessent and Lutnick, who were basically outlining this plan for, let's call it, 10% to 15% to 20% tariffs across the board that would amount to about $300 billion in total tariffs, up from $75 billion in 2024.

But you had Navarro, who was basically saying, "We're going to replace the Internal Revenue Service. We're going to get rid of the income tax, and we're going to have $2 trillion of tariffs." Okay? And I was very clear, and I think the market was very clear. We all voted with our wallets and said, until the president tells us whether it's door 1 or door 2, we're out.

The market shot first and asked questions later.

And that's where you saw that huge drawdown in the market. The Nasdaq was down 21% right at its trough this year. Now the Nasdaq's up over 10%. It's a 30% move in about 60 or 70 days, which is extraordinary even by the historical patterns that we've seen over the course of the last 5 years. But let me back up here for a second.

I think the consensus view of all economists—90% of economists—is that tariffs are going to be bad. They're going to be a tax that gets paid by the U.S. consumer. There was a small group led by Bessent and Kevin Hassett at the National Economic Council that said, "No, it's actually going to be different this time."

And the reason it's going to be different this time is their theory, they argued, was that the world had become dependent upon exporting to the United States, so that the total trade deficit to the United States of goods and services was about $915 billion last year—a $1.2 trillion goods deficit. And that basically meant that China was selling a lot more to the United States than they were buying of U.S. goods.

And so what Bessent and Hassett postulated was that these countries have no choice. If we impose a tariff on them, so long as it's not draconian—70%, 80%, what Navarro was talking about—if we impose a 15% tariff on them, they have to eat it. The producers have to eat it, because otherwise they're going to end up laying off millions of people in Vietnam, in China, in these countries. And politically, they can't afford to lay these folks off. So that was their theory of the case.

The consensus economists said, "No way is that true. You're going to see massive inflation." But what have you seen? You have not seen the inflation percolate through. I will caveat this by saying "yet."

Okay. So here we are in July. The consensus economists said it would have already happened, and the National Economic Council was out with a paper last week that deconstructed core PCE. So that's the best proxy the Fed watches for inflation since the start of the year. And it showed—this was really interesting—import prices have been going up at a slower rate than domestically produced goods.

Okay? So this is the exact opposite of what you would have expected from tariffs. Of course, you would have expected imported prices would have been going up more than domestic goods. And so we'll show these charts, and we'll put the link to this paper. People ought to take a look at that.

But to me, when I look at the president's deals he's landing—the deal he just announced yesterday with the European Union: 15% on all goods coming from Europe, 0% on US goods going to Europe—it's an opening up of European markets, with Europe paying us 15%. On top of that, we're getting commitments like $750 billion—almost a trillion dollars—of energy purchases from the US.

Or look at Japan, which they announced last week. Another huge market. Again, similar: they're going to pay tariffs to the United States, with no tariffs imposed on the United States. And they're going to invest $550 billion into the US in a way the president gets to direct.

I just think we need to give the president credit where credit is due. Everybody said this was going to lead to retaliation and trade wars, and was going to be disastrous for the US. All we've seen so far is deals, deals, deals, deals.

And I have to say, if this was the CEO of one of our companies, right? Let's say we had a board meeting at the start of the year, and he outlined these plans, and we said, “Hey, we're really nervous about this. This is a high-risk, high-reward strategy. It's either going to backfire and we're going to fire you, or it's going to work really well and we give you a bonus.” If we're measuring him halfway through the year, I would say that he's in line for a bonus based upon the trillions of dollars that are going to be coming into the United States.

We now have a $300 to $350 billion recurring stream of revenues into the Treasury in the form of these tariffs, which are being paid. And I think Scott Bessent said last week that in June we had our first monthly surplus in the United States since 2015.

Right, because of the tariff revenues that are coming in.

So I will say this: I was on the fence. I knew the nuclear tariffs—the trillion or $2 trillion—I knew that was a disaster. I said if we landed the plane where Bessent wanted to come in, at $300 billion, I thought there was a decent chance that those prices could be passed on in the home countries, and so far it looks like that's the case.

Do you have any concerns? What's the—anything you're watching out for?

Well, I think the number one thing is core PCE. It did bottom last year, and it started to tick up, so we have to keep our eye on inflation. Of course, I think it's almost impossible to conceive that we would have totally reordered the entire global trading system on the first pass with zero mistakes. We're going to have some goods and some products that US consumers are going to end up paying the taxes on, and we're going to have to go back and fix some of these things.

I will say that it's turning out massively better than the consensus criticisms. Remember Larry Summers at the Cato event.

This was just a few months ago, and he was saying this was the biggest economic disaster of his career. And I don't think you can describe it that way.

The markets—the voting machine—are telling you, and these are a lot of sophisticated investors, that no retaliation, no trade wars, and all these deals getting done work for the US economy. I will tell you that the Atlanta Fed GDPNow tracker, which tracks real-time GDP, has now ticked back up to 3%. After going down a lot in April, it's bounced way back up. Economic activity appears to be going up.

And then one final thing here: remember, one of the key reasons for doing this, right, wasn't just that the EU conceded in the trade negotiations that our relationship was unfairly balanced in the direction of the EU. They said that's the starting point, so we have to rebalance it.

But on top of that, a key reason for doing this was to support the domestic production of critical national industries and to make our supply chains more resilient: chips, data centers, energy production. At Altimeter, we just led the Series A in a company—I don't even know if we've announced it, but I'll announce it here—which is an all-American producer of rare-earth magnets.

These are now viable investments because of the tariffs and the resolve of the government to re-onshore these critical supply chains. That's a huge benefit that you would be willing to pay something for, but we're getting the benefit and, on top of that, we're getting paid.

Sunny Madra

Yeah. Can I add one thing, guys? Not on the economic side, but running the supply chain at Groq—and look, we're fortunate that the majority of our supply chain is US-centric, including our chips. But we still have small, discrete components, and, Brad, exactly what you were saying is happening: the producers, the manufacturers of those things, are coming, and you're negotiating. We've had pretty significant negotiations with those folks, and that was, I think, not anticipated.

The other thing is it hasn't been static. Again, I'll go back to this administration: we've seen our tariff schedule move around probably 6 to 8 times since this all started because they keep evaluating, they understand, and they listen. I think that's also a function of how this administration operates, and we want to give them credit for that.

People can go and share with them, “Hey, these things are not available. We can't replace them right away,” and I think that's also helping with that last thing you said, with the investment you're making, as companies get established onshore to take advantage of what these tariffs are causing.

Sunny, has your supply-chain planning and that dynamic nature started to settle down? Do you see this—are we reaching the end of the tariff negotiations such that everything can kind of settle down and operate?

Sunny Madra

It's gone from week to week, or even day to day, when it first started and we were trying to figure out what happened, to now we're looking at it monthly. So it's definitely settling. And, Brad, we still have a big thing looming with the China discussion, correct?

Yeah. You know, listen, we landed Europe, we've landed Japan, and we're going to have—the long list is going to be coming out. But when you look at our big trading partners, the EU—we do, I think, $900 billion of trade with them a year. With China, it's about $600 billion, but we have about a $300 billion goods-trade deficit with both Europe and China. They were the 2 big ones.

China is the big enchilada because it's not just trade with China, right? It's strategic, it's national security, and it's trade. And it's the AI race. We know that the rare-earth ban on Chinese magnets was devastating to US industry. We know the retaliation, where H20s were cut off in terms of Nvidia's chips going back to China.

Reading the tea leaves, I'm going to go out on a limb and say the consensus still believes that the China thing is going to be a problem or that it will be small. I think this president wants to do the biggest deal ever done with China. I don't think he's dogmatic at all. I don't think he's some big China hawk. He said at the AI Summit last week, “I'm a deal junkie. I like to do deals.” You can't be the biggest dealmaker in the world without doing a big deal with China.

All right.

Right. If you just look at today, the Chinese reciprocated in a way I think the US government was looking for. They said, “Hey, we'll postpone all of our retaliatory tariffs.” They've invited the president to China. The president has suggested he's going to go to China the first week of September, or sometime between September and November.

I think there's a very big deal that's going to get done with China that's going to reorient the relationship in a big way. Let me just tease this: the president said earlier this year something that caught all of our attention. He said, “You know, if I could wave a magic wand, I would cut the defense budget in half for the United States, for China, and for Russia.” We've never heard a US president in history utter anything like that.

That would be amazing. That is what I would call an extraordinarily flexible mindset.

And if you go into this deal negotiation with China with that sort of flexible mindset, I think it could include all of the above: rare earths, chips, maybe even military cooperation, certainly a rebalancing of trade.

I think China—listen, we entered the year with China paying 15% in tariffs. That was pre-Trump. They were paying 15% to the United States in tariffs. So I don't think we're going below 15%. I think they're going to pay that in Trump 1 and Trump 2, right?

So I think they will continue to pay at least 15%, but I think it's going to be much more structured, much more nuanced. Bessent's been very clear: China has to rebalance to domestic consumption and away from an export economy that's really sticking it to the US in terms of the trade deficit.

I think China gets that. I think they want that for their own country. I think they're willing to do that. I think the United States understands that can't happen overnight; it has to happen over a period of years. I think there's going to be a big Chinese deal done before the year's out.

So let's close with this, Brad. You've often, on the pod, been willing to speak about your own temperature for the US markets and whether you're net long or net short. You've been more enthusiastic on this podcast than I've ever seen you, both about AI, but also about this China theory that you have. I think it would cause the markets to rip if you're correct.

But you also said we're at all-time highs, and so you want to buy low and sell high. Where's your head hanging?

We did a pod, I think, around May 2. Well, we did the pod in March, and I said we're out of the market.

I remember.

Right. And you said you were early, and Liberation Day came, and we were happy to be out of the market. On May 2, I said we're all back in because the Bessent consensus has won. It's going to be 300 billion. We're going to land the plane, and I outlined a flight path. I said, you can land the plane: no inflation, you get rate cuts, and it's kind of off to the races.

We're up 30% off of that bottom in the NASDAQ since then. So 30% is a huge move, but when you telescope out, we're up 10% for the year.

Okay?

And if I had told you guys on day 1 of this year, here's what's going to happen: we're going to rebalance global trade and we're going to land the plane around 300 billion. We're going to have the economy grow at 3%, accelerating. We're going to have no inflation, heading toward rate cuts by the end of the year. That is the backdrop, and we're going to have all of this AI demand and accelerating demand for AI compute.

I would have said the market can be up at least 15% for the year. I think we've captured a lot of the return for the year. Bill, I will tell you this, though: we see tons of opportunities, and so I would say that we're also bullish on what we see happening in AI.

Sunny's going to raise a huge new round here. We're happy to be investors with Sunny as well, and it's extraordinary to watch what Sunny's—

No, seriously. He wasn't supposed to disclose that. He can edit it out.

He has it, but his face is turning all red.

Have a good week. And thank you, Sunny, for coming on with us today. Thank you so much.

Sunny Madra

Thanks for having me.

Yeah. Thanks, guys.

Sunny Madra

Bye-bye.

Take care, guys.