Stargate、行政令、TikTok、DOGE、公开市场估值 | BG2 与 Bill Gurley、Brad Gerstner 对谈

Elon is 100% right: nobody has $500 billion. In fact, nobody has $100 billion to contribute to this on day 1.

Hey, man. Great to see you.

I mean, never a dull moment. This goes well beyond being a dull moment. I think we have a year's worth of activity in about 48 hours.

This week, I had an amazing weekend in D.C. There was a tremendous amount of optimism. I've been saying to you and our friend group that I remember during Trump 1 in 2016, every morning you woke up kind of holding your breath at what had changed since the night before. It was really the first administration that lived on Twitter, and here we are again.

In some ways, over the last year, I feel like not a lot was going on in the world. I wasn't stressing out every morning when I woke up, and now this administration has kicked off with a bang—a major, major bang at the end of the day yesterday with Stargate, which we're going to talk a lot about today.

1. Stargate

This morning, I have to say, even for those of us who are fascinated by the back-and-forth in Silicon Valley, the back-and-forth among perhaps the most seminal figures in Silicon Valley—Elon Musk, Larry Ellison, Satya Nadella, Sam Altman, and others—over what is and is not going on with Stargate is truly a situation where, in some ways, facts are more intriguing than fiction.

Why don't we start off by having you level-set for folks? There was a huge announcement at the White House yesterday with Larry Ellison, Sam Altman, Masayoshi Son, and President Trump. Talk to us about what was announced.

Yeah, and maybe I'll do it in terms of how it rolled over me. This was the day after the inauguration and the high-profile performance Trump had with the executive-order signing. There were rumors—I don't know where they started; maybe I saw it on TV—that there was going to be a big press conference, and it revolved around this thing called Stargate.

As a reminder to everyone, Sam Altman at OpenAI had been talking about the concept of Stargate for a while. He had suggested that maybe he and Microsoft were going to do it. I think that's important, and we can dive into that later, because it's the same name that had been thrown around. I don't think there was ever any verification from Microsoft that they were in on the Stargate project, but clearly this is something Sam has been thinking about for some time.

It was supposed to happen at 4:00. I think they finally started at 5:30 or something, so it was 90 minutes late. Then this door opens, Trump comes in, and you have Masayoshi Son representing SoftBank, Larry Ellison representing Oracle, and Sam Altman representing OpenAI.

It looked a little half-assed. It looked like it had been thrown together kind of quickly, and, in my interpretation, one of the reasons it might have been thrown together quickly is that Trump was looking to show that he was having quick momentum right out of the gate in terms of having an impact. The things he said would suggest that that was part of the reasoning for doing this now and so quickly.

If you just go to the post that OpenAI put out, it's probably got the most detail. This thing is lacking quite a bit of detail, which is why there's so much speculation. It says the Stargate project is a new company, although even that is in question. Is this going to be a new corporate entity?

The company intends to invest $500 billion over the next 4 years. Microsoft is the highest right now at $80 billion a year, so this would be $125 billion a year if they met that commitment or target. They said, “We will begin deploying $100 billion immediately.”

There's a lot of talk about American leadership and about this being in America. It says the initial equity funders in Stargate are SoftBank, OpenAI, Oracle, and MGX. Maybe you can tell our listeners who MGX is in a minute.

It says SoftBank and OpenAI are the lead partners, with SoftBank having financial responsibility and OpenAI having operational responsibility. Son will be chairman. It lists the key additional technology partners as Microsoft, NVIDIA, Oracle, and OpenAI.

That's the gist of it. It sounds to me like maybe an even better-funded version of CoreWeave, but there's a lot I don't know. I'm asking questions like a lot of other people. How do you see it?

Well, I think that for well over a year, OpenAI and many others have been talking about the need for massively more compute. A lot of the talk in Washington this weekend, with various cabinet ministers and others, was that this administration was going to dramatically accelerate power generation in the U.S., data-center construction in the U.S., and compute in order that we win at AI.

There are no surprises there. The fact that they were able to announce this the day after the inauguration, that it was at this scale and magnitude—remember, this comes on the heels of a lot of talk in the back half of last year that Stargate was dead, that the need for this much compute was overstated, and that maybe models were hitting some scaling limits or smaller models could get the job done.

There was a lot of skepticism, frankly, about the demand for compute. There was a wall of worry in the back half of last year about what was going to happen with compute. This announcement was very different and much bigger than the original $100 billion that was discussed when you and I first talked about Stargate in the spring of 2024.

When they made this announcement, the first thing I started asking myself was how they actually roll this out, because you're right: there was a lot of white space left in it. I immediately, being the analyst that I am, got the team on the phone and said, “We need to build a bottoms-up model.”

Everybody is talking about $500 billion. This morning, we have debates between Elon, who says, “Funding not secured”—they don't have the money—and, on the other side, people saying, “Of course, we're already doing this in Abilene.”

We know that there was construction and work going on in Abilene last year with Oracle. Larry said at the press conference that they have 10 buildings already built in Abilene and another 10 under construction. We also know that OpenAI is already running workloads out of Abilene, Texas, which is where this megaproject is going to be built.

I bet you that's the first time you've said “Abilene” 3 times in a row in under a minute, by the way. I talked a lot about Abilene last year on the pod and with our team. Do you even know where Abilene is?

I do—West Texas.

One of the things I hope you and I add to the conversation, writ large, is that there's a lot of talking heads and battling going on this morning, but what might this look like in reality? We'll share some projections that our team has made, which are obviously back-of-the-envelope because we don't have the precise data.

Let me ask you a quick question. If this is already “underway” and, according to the OpenAI press release, this is a new company, are we presuming that the newco is acquiring some assets that were owned by somebody else? Whose balance sheet was the initial project sitting on?

Yeah, I honestly have no details on that, and if I did, I probably couldn't share them, Bill. Just as a reminder to everybody, from an investment perspective, Altimeter is invested in a lot of these companies. Our best perspective here is that we know there was construction and work going on in Abilene last year with Oracle. I assumed that was kind of a direct relationship between the company and Oracle, but I have no insights about how this thing is particularly going to be structured.

What's more interesting to me is the size of the equity check. There's this question, based on some of the tweets this morning. Gavin Baker—and we'll show that as well—said this $500 billion is totally make-believe. That might make you think that somebody has to show up on day 1 with $500 billion in order for this to get off the ground. They're almost saying that it's a ruse, that the announcement was just a bunch of B.S.

But if we break it down, let's assume they started on this 3 months ago and were able to secure 250,000 GPUs from NVIDIA this year. If you break that down, that's about $13 billion of CapEx that they could spend in 2025. In the press release you quoted, they said, “We're going to start to spend $100 billion in 2025.”

With $13 billion, you can stand up maybe 250,000 GPUs this year. Of the $13 billion of CapEx required, if you make an assumption about a 25/75 or 30/70 equity-to-debt split, that probably requires something like a $3 billion equity check.

I'm assuming there are debt providers. We've read about these deals already in the past. MGX, which is the sovereign wealth fund—a big, big fund out of the United Arab Emirates—and BlackRock are also said to be involved in the debt financing. We know there are a lot of debt providers competing to provide debt to these data centers, so the equity check for 2025 would be relatively small.

Let's assume that they want to ramp that up, put the pedal to the floor, and ramp it up in Abilene for 2026. Let's assume they could secure 2 million GPUs from NVIDIA. Now you're talking about $100 billion in total capital that would be needed. Seventy to 80% of that goes to GPUs, and 25 to 30% goes to data-center land and power.

The equity check required to do that, Bill, is about another $25 billion in 2026. Just to put it in perspective, to stand up 2 million GPUs in 2026, you need about 2.7 gigawatts of power. This is not inconsequential at all. It would be the biggest deployment of GPUs anywhere in the world, even if they just stopped there.

One thing worth noting that I failed to mention: Trump did say that one of the things the Trump administration was bringing to the table was a push to clear regulation and allow more power generation, faster. We've heard from people like Satya Nadella who say they're power-limited.

One thing that may be new in this situation is that this entity may have an advantage in getting hold of power faster, although I would imagine there would be an incentive to do that for all the players.

I would tell you, being in Washington this weekend and speaking with folks coming into the administration, like Doug Burgum, that if there's anything I believe to be true, it's “drill, baby, drill.” Tapping U.S. energy reserves and removing obstacles and regulation in order to light all this up in America rather than elsewhere is probably priority No. 1, and probably the quickest thing you could do to generate power and get it up to speed as fast as possible.

Putting in an LNG plant like that is very fast.

That's exactly what's going on, I think, out there in Abilene. You and I should go for a visit. You can't get SMRs, nuclear fission, or anything else that can do this, but that's why they, I think, started building there.

But just back to the analysis, Bill: to get to the end of 2026 with 2.5 million GPUs, the equity check required is something like $25–$30 billion. That just calibrates how much somebody really has to show up with. In order to get this stood up, you're talking well under $50 billion in years 1 and 2. You're talking about $70–$80 billion over a period of 4 years, through the end of 2028.

The $500 billion is a sexy headline. It's important that these things are going to be built here, but the equity check required to stand this up—and if they just achieved what I outlined through the end of 2028—requires 7.2 gigawatts of power out of Abilene. To me, you're kind of on the outer limits of what's possible.

But they talked about multiple locations too, so they would move. Now you're assuming a 3-to-1 debt-to-equity ratio, which I guess is something similar to what CoreWeave has done. Equinix has, over the years, run closer to 1-to-1, and that would be assuming a substantial amount of risk.

No one's more prone to risk than Masayoshi Son and SoftBank, so that is very plausible as their assumption.

Remember, over the days and weeks ahead, we're going to refine this back-of-the-envelope analysis. I just wanted to calibrate that nobody has to come up with $500 billion right now. Nobody has to come up with $100 billion right now. Nobody needs to come up with $50 billion right now.

We know there are lenders willing to lend, but that's an important question, and we're going to try to get to the bottom of it: What would BlackRock and MGX require in terms of that debt-to-equity split?

2. Dylan Patel and Semianalysis take on Stargate

As we know, everybody had a huge incentive to get to this big number. Masa wanted to get to a bigger number. When Trump and the others were down in Mar-a-Lago, Trump took it from $100 billion to $200 billion, and now they got to $500 billion, which served everybody around the table really well.

We broke it down in our numbers to try to reverse-engineer our way to that $500 billion in CapEx. Even in our numbers, we demonstrated that you need a much, much smaller equity check. Nobody has to show up with $100 billion or $500 billion on day 1.

There's another way to look at this as well. We're not even talking about $500 billion in CapEx, but instead, as Dylan Patel—who was on our pod in December—has done an analysis that says maybe they were talking about the total cost of operation, not CapEx at all.

He has a piece of analysis that is quite interesting, which shows that even if you aggressively build in Abilene, the fact is you're only going to be able to spend something like $100 billion in total between now and the end of 2027. I think he estimates something like 800,000 total GPUs that would be purchased over that period of time.

He does note that those GPUs, which are much more powerful—Blackwell and eventually next-generation GPUs—will be the equivalent of millions of H100s. That's just if we want to compare what he's forecasting in Abilene with the 200,000-GPU cluster we're talking about in Memphis.

In that case as well, it's very consistent with us: the amount of equity that would be required on day 1 is a fraction of the amount we're talking about. If you're trying to square the circle of how Elon and Sam can both be right, Elon is 100% right: nobody has $500 billion. In fact, nobody has $100 billion to contribute to this on day 1.

But both our analysis and Dylan's analysis show that that's not required in order to begin scaling this up. Sam is also right: they are, in fact, building in Abilene, and you can scale up to much, much bigger clusters without this headline number being required to be delivered on day 1.

But, Bill, let me ask you a different question. Let's assume that we're going to spend a lot more money than we thought we were going to spend 30 days ago, or certainly 3 months ago. Maybe talk a little bit about where this tells us we're headed, or how this impacts the competitive landscape out there.

3. Stargate’s Impact on AI Competitive Landscape

Well, I think in order to do that, you've got to back up and look at the motivations of each of these parties. I think there are numerous things going on that all led to this big event, and that's not to suggest that it's not going to happen. It sounds like it is going to happen; it's just that there are a lot of different motivations.

As we talked about, Sam has been talking about this for a while. He has a belief that he needs access to a very large data center. It appears, from this conjecture on my part—and I'm not an investor like you, so if you can't respond, don't feel that you need to—but it appears that 2 things might have played a role in this happening. One is Elon and xAI.

Elon and Gavin Baker, who is a big backer of xAI, have pointed out that they have a competitive advantage over OpenAI because they have their own infrastructure. This isn't exactly OpenAI's own infrastructure, but it probably feels that way, so that may have provoked this a little bit.

The other thing is the much-discussed relationship between Microsoft and OpenAI. This easily could be one of the pieces in the back-and-forth between those 2 parties. We got a quick press release out of Microsoft immediately after this was announced, and it's hard to dissect what might be going on there, but I think that played a role in this happening.

My own interpretation is that Microsoft didn't want to build this big OpenAI playground, or one as big as maybe Sam wanted to build, so he went somewhere else. Everybody, on the surface, is behaving like they're okay with that. I think the other parties—Masa is always looking to be involved in the biggest movement that's happening. That happened in previous waves we've been involved with, so it's no surprise that he shows up here.

For Larry Ellison and Oracle, if you look at their relative position as a hyperscaler, I think people think of Amazon and Google before they think of Oracle. This gives them a platform to brag a little bit that they're at the front of the line.

I think those are the things happening in the background that take us to this place. It's important to note that Oracle was, I think, at the start helping xAI in Memphis, but Memphis decided to go it alone and build out its own infrastructure. As we discussed with Jensen on the pod, listen, Elon is an N of 1: He stood it up faster, it worked better, and he built a bigger cluster.

We're going to see. He's going to launch Grok 3 here. I think they're going to have the output from the first training run on the biggest cluster in the world, and we're excited to see what that yields.

From my own perspective, set aside for a second—I think you nailed it. Those are largely the motivations of the individual parties. Again, just as a country, we just went through this amazing inauguration. We're going to talk a little bit about likely Stanley Druckenmiller later, but I think this level of competition, the fact that this scaling is occurring in the U.S., and the fact that this investment is occurring in the U.S. is fantastic for the United States of America.

This increases the probability that we're going to win the race to AI. We needed more power. We need this 5 to 10 gigawatts to come online, and we need to have it now. Does that lead us to AGI or ASI or whatever? It certainly puts us in a very strong competitive position.

You're right, I think it is that competition on the field, Bill. I think it is the fact that Elon stood up a bigger cluster, that is a competitive advantage, and he did it fast. Now you have an alternative here, which brings me back to what this means for the competitive landscape.

Bill, think about it: You and I just went through this slide that we'll share again on the hyperscaler capex expected for 2025. You've added another player on top of it. The person who is probably most obviously impacted by this is Jensen and NVIDIA, and of course, the stock is up 5% today. These aren't secrets. These are obvious facts, right?

But it's yet another large customer for NVIDIA gear, and the size and scope of what they want to do—even if an NVIDIA competitor showed up—I don't think that competitor could create enough production in the amount of time you need. You even hinted that Stargate may be limited by whatever NVIDIA is willing to give them.

We've been chip-constrained for 2 years, and now you have a new player on the field that may be raising its hand and saying it needs 2 million GPUs. Remember, just a couple of years ago, we weren't making 2 million GPUs. I think the forecast this year is for something like 6 million GPUs.

If I'm sitting in the Googleplex today, what does it say to me? I've got to spend more on capex. I've got to make sure I secure my GPUs. If I'm at Microsoft, I've got to secure my GPUs and make sure I'm spending enough. If I'm at Meta today, what do I need? I need to make sure that I'm not losing out to Stargate, OpenAI, and xAI in the order book for GPUs.

I heard just as recently as a month ago that in 2026 there wasn't going to be demand for GPUs and whatnot. Dylan Patel came on this podcast and called all of that garbage. He said the demand is off the charts. People aren't making 12-month bets; they're making 3-year bets. I think, again, this is just further validation of that.

Here's the way I think it changes the competitive landscape: The bigger the stakes get—and Sam just pushed a big pile into the middle—there are only certain companies that can be in that game. xAI, because of the genius of Elon, the momentum they have, and the operating businesses that stand behind them, is 100% going to be in that game and maybe eventually leading that game.

Amazon is facing a real test here. Is Anthropic going to be able to show that they can raise the money that's going to be required? They have less than $1 billion in revenue. They have great models, I would argue, but they don't have the consumer or enterprise traction that the other folks do. Is Amazon willing to put up that sort of money? Amazon has never been that aggressive with capex relative to the other folks.

Meta is going to be in the game, and Google is going to be in the game. I suspect that this puts more pressure on their capex to come up. I think you're right that, because of that and the fact that Jensen said there are 35 other AI factories around the world that are not hyperscalers, you've now got Oracle being a major player in the game because of Stargate.

Again, I think the net-net benefit here is that we're going to get a lot better AI from all of these players. But I think you are going to be chip-constrained, and you are going to be power-constrained, for the next several years. If you're subscale on any level, I think those players ultimately have to get folded into a much larger entity.

We saw, reportedly, Anthropic raise a couple of billion from Lightspeed and maybe another billion—I saw an announcement out of Google today—so they are raising money. They do have the capital, but every time they buy in, Bill, for a little bit more at the poker table, somebody else goes over the top and the demands of the pot size just get bigger and bigger.

Yeah, one of my big takeaways is that it's a sport of kings. It's the amount of money that's being thrown around. I think, with all the announcements, the $1 billion from Google into Anthropic almost fell through the cracks because it just doesn't feel like as big a number.

It's part of why I think there's so much hyperbole. There's a game in the background of, can I bid even more? Can I bid even bigger? Can I talk even bigger, potentially to try and scare some people out of the game?

I also think, though—and we've talked about this in the past—that causes an increase of risk as well. There's some point at which you deploy capex, which is slow to bring on, where you could overrun demand. If you have a 3-to-1 debt-to-equity ratio, that could be a really painful downside.

Clearly, no one right now is thinking that way. Clearly, all these people are very positively minded. I tip my hat to Sam Altman's aggressive, ambitious gamesmanship. This is clearly something he wanted to do, and he found the right window.

It's certainly interesting to watch the triangle of Sam, Elon, and Trump here. Some people said, "Is Elon going to tell Trump what to do?" Clearly, you've got some dynamics where not everybody is on the same page.

This comes back to the question I asked about where this tells us we're headed. What I mean by that question, Bill, is really: Where are we in the stage of model development? We know that pretraining is asymptoting. You've been at the forefront of discussing this, but we also know, as Jensen showed in this slide, that we now have 3 waves of scaling.

We have post-training, and now this inference-time compute. Lost again in the shuffle today, OpenAI announced Operator, which is built on the back of inference-time compute. That's really more around planning and actions.

When I look at 2025 and what this will enable, there is clearly a lot more enthusiasm about the progress being made by these models, whether that enthusiasm is at Google, at xAI and what we're going to see out of Grok 3, or what they're seeing out of the o-series at OpenAI.

Part of the reason I think folks have the confidence to commit to this level of equity and debt is because of the use cases and the demand that they're seeing from both consumers and enterprises for more of this, as well as the progress that's being made on these models.

4. DeepSeek

Maybe just talk for a second about these breakthroughs happening in China around DeepSeek. Over the weekend, people were blown away by the benchmarks achieved by this small, open-source, inexpensively trained Chinese model.

I can assure you of this: One of the top priorities on the mind of our new AI czar, David Sacks, and lots of folks in the administration is national security. Their number-one objective is, how do we speed up the United States? You saw the Stargate announcement; that's all about speeding up the United States. But the other one is, how do we not make it easy for China to compete?

Here's a situation where, even with degraded chips—they don't have cutting-edge chips out of NVIDIA—they seemingly trained something that's very competitive. What's your read on that, Bill?

My studying of what other people are saying, because I try to absorb as much of this as I can, is that the Chinese—DeepSeek in particular—have figured out a way. The word people are using is “distill.” They've figured out a way to basically shrink their model, theoretically possibly by connecting to an API of one of these other foundational models that we've talked about and using that as a guide, and even make these things more efficient.

They can pack the same amount of horsepower, if you will, into a much smaller engine and get to the exact same competitive benchmarks on the output. In one way, you could say, “Oh, my God, this is this great commoditization of the foundational models.” I think that's potentially valid.

The other thing you could say is, “Wow, we're going to get so much more performance per token price,” and that can increase the market because it gives so many more tools to startups that they can play with. Apparently, one of the tests even had this thing doing chain-of-thought reasoning, and it was successful at that.

When you consider that this was a 10- to 100-times increase in token use, if you get down to these lower price points, now you can do even more for less. So it could unlock quite a bit.

For me, the most interesting part of it is—I think I've said this before—I actually think the policy of the American government to try to keep China out of the AI game is futile. I've always felt that way. I don't think it will work. In this case, I actually think it backfired.

5. Special Guest Rene Haas (CEO of ARM Holdings)

There's a phrase that people use: “Constraints drive creativity.” Steve Jobs used to say this quite a bit.

Whoa, whoa, Bill—bookmark that. René Haas, CEO of Arm, just jumped into our conversation. Hey, René.

René Haas

Yeah, pretty amazing stuff, right?

I appreciate you jumping in here, René. I know we've got maybe a 15-minute cameo. We'll have you back when we can spend a couple of hours together because you're such a thought leader, having spent so much time at NVIDIA and now running Arm.

As most people know, Masa, I think, owns 90% of Arm, so you're very close partners with Masa. Obviously, you've been party to the conversations. You're named as a technical partner, along with OpenAI, Microsoft, and Oracle, in yesterday's really earthquake-level announcement out of the White House.

René, help us understand a little bit how you see this playing out. There's a lot of talk about this $500 billion number—whether this money exists. Bill and I just went through this analysis, and from my perspective, this ramps up and it's going to be a combination of equity and debt. Nobody has to show up with $500 billion on day 1.

I'm forecasting maybe 250,000 GPUs this year, maybe 1 million or 2 million next year. What's your role in this, and how are you thinking about how this scales up over the next few years?

René Haas

I have a few roles in this, and I know you guys know Masa and me well. I know you definitely do, Bill. I spend a lot of time with him. Part of it is that he owns 90% of Arm, so I have a lot of investor meetings with him, with him as my chief investor, and we talk about strategy.

He himself has been pretty big on this idea of the singularity for quite some time. It's something he's had a vision for in the long game. The ChatGPT moment was a bit of an accelerant for him.

There's a game to be played here relative to capital, relative to compute, and relative to power, and he wants to play a big part in it. There were a lot of discussions, whether it was Sam wanting to buy fabs or different assets that were trying to look at power in different areas of the planet.

I think something had an opportunity to come together to solve this giant problem of how you get access to so much energy that's needed, we think, to drive AGI and ASI at numbers that are even well beyond the balance sheets of giant companies like Microsoft, Google, Meta, and AWS and Amazon.

It all kind of came together at the right time and the right place. But like everything in life with these types of things, it didn't happen overnight. There were a lot of discussions and conversations taking place for weeks and, actually, many months.

I think this all came together with a confluence of ambitious partners like Sam, ambitious partners like Larry Ellison and Oracle, and, quite frankly, a new administration that was ready to take action in a very fast way.

That's what you saw come together yesterday. One of the most amazing things about it—I know you, Brad, and I had talked about this in the past—is to imagine that 28 hours after the president took office, he's announcing a project around AI data centers and buildout with 3 large players in the tech industry. It's kind of amazing. That speaks to the importance of how the new administration views all this.

It's truly an Apollo-scale project. I guess, as you think about Arm, why don't you explain to us again what Arm is actually delivering into Stargate? I know you're embedded in the GB200, but maybe just share with everybody else the role you play.

René Haas

At the highest level, the way to think about it is that you've got a giant, as you said, Apollo-Manhattan Project—whatever terms you want to use. I guess it's probably the largest infrastructure buildout in the history of the world.

Every data center, whether it's running general-purpose compute, inference, or training, needs a base CPU to run everything, and that's our role. Whether that's what we're part of today, which is GB200—we're super happy to be partnering with NVIDIA on that product—or other areas that we haven't talked about yet in terms of productization, there's lots of opportunity for Arm because the base CPU will be Arm.

I think therein lies a huge opportunity. One of the things that people don't always appreciate with, let's take GB200 running in an AI data center, is all of the other work that needs to take place. Whether it's the hypervisors, virtual machines, or anything that the normal CPU has to do in a data center, it has to be run by something, and that's what Grace does.

When you baseline that relative to GB200 being where we are today, the opportunity going forward, in terms of these large data centers doing some level of mixed inference and training, reasoning, and reinforcement training, is huge. There are a lot of opportunities for Arm to do even more than what we're talking about today.

It's super exciting. Bill, I know you had some questions.

René, as you think about Stargate—it's a new company, a new entity—I know Masa is the chairman. Any insights for us? Who's running this thing, and how should we think about this entity relative to the other entities?

My mental framework is that it's kind of an operating shell that everything runs through, but the actual compute, the offload, is going largely, or 100%, to OpenAI; the inputs are going to be coming largely from you guys and NVIDIA, and of course all the other people needed to network and do the things in the data center.

Is there an idea that you guys are investing in an entity that, unto itself, will have power and maybe grow and serve other customers? Or is this really just about coordinating the activities of the people who are around the table?

René Haas

What I can say today, Brad, is that it's much more of the latter than the former. Could there be an opportunity for the former somewhere down the road? Potentially. But right now, it's what you just described.

The operational control will be from OpenAI, so they'll call the shots relative to all the things involving the operation. Obviously, there are existing relationships with NVIDIA, Oracle, Microsoft, and ourselves, but going forward, OpenAI is going to be in a very key role in the operation.

If you go back again to Sam and his team spending a lot of time and energy over the past 12 to 18 months seeking ways to get opportunity and access to large resources to advance the training of these large models, this is kind of where he's been with this. I think it's not inconsistent with some of the actions and behaviors you've seen over the last number of months.

Can you think of an entity like a comparable entity in the past? I'm having a hard time imagining what that actually is.

René Haas

I don't think there's a good comparison on this, Bill, because when you think about the amount of capital that's required, it's bigger than anything anyone has done. This required a very, very novel set of partners to come together, both with a big vision, a large opportunity to get access to capital, and, candidly, probably a little bit of a willingness to say, “We're going to figure this out as we try to grow it.”

This is beyond what we've done before. The only analogy—and it's not a good one in terms of how I can think about it—is maybe GlobalFoundries and Mubadala, relative to starting to see that traditional fabs needed to get extra capital. That was at a much smaller scale.

Now, with Satya talking about spending $80 billion of capex, even for Microsoft, that's just a giant number. At these numbers, no one company can do it.

René, if you think about what's built into this forecast—and you don't have the privilege of seeing what we talked about earlier—I had my team do what we do well. We're a bunch of analysts: We take what's known and try to build a bottoms-up model.

What we got to was, by the end of 2028, consuming about 7.5 gigawatts of power in Abilene, standing up about 2 million GPUs a year in that infrastructure—6.5 million GPUs in total. That would spend roughly $300 billion to $350 billion of the $500 billion that they said they would spend.

When I look at that level of power and think about where the bottlenecks or risks are here to this buildout, obviously this is operationally difficult. As I said, building 2 million GPUs is a third of what NVIDIA is expected to make this year, right? So that's not an insignificant amount of demand out of NVIDIA.

Where do you see the bottlenecks, to the extent they arise? Is it power, GPUs, or Arm?

René Haas

Those are all the right questions and stuff that we've been talking about, as you can imagine, for a number of months. One of the bigger bottlenecks was—or is, hopefully, addressed yesterday—in the sense that you have a government that's going to help you in terms of permitting and regulations.

That was not small, by the way. A lot of the aggressive things that were happening in terms of buildout were literally buying Bitcoin-mining facilities that were out of business and repurposing them. Now, when you're talking about having to build a bunch of new things, regulatory matters are a big one.

The numbers your team looked at are probably in the right ZIP code. Our numbers might be a little bit bigger, but your numbers are about right. Fab capacity does become an issue, no doubt, because when you start thinking about 3-nanometer and 2-nanometer, TSMC is the leader. It's the only game in town on some level, so I think that is a potential limitation, or at least a constraint.

HBM memory and DRAM, for sure, are another potential constraint. Then you get into a couple of things. First off, when you're talking about gigawatts of power, can you locate it? How physically close can those data centers be?

If you're trying to do training dispersed across multiple facilities, what does that look like? Can you really train a large model relying on the connectivity of the network across multiple physical sites? That is a challenge.

Human labor to connect all these cables—I know that sounds a little trivial in the context of some of the technological problems we're talking about, but it's nontrivial at this scale. I think automation is a potential solution, and I think robotics are a big potential solution here.

How do you build these? Do you build them from the ground up, or do you containerize them and put them together in a modular fashion? I think there are opportunities relative to scaling this—some known, some not so known—and I think all of that is going to be played out over the next number of years.

René, I think you nailed it, and I think you nailed it when you said, echoing likely Stanley Druckenmiller, that in his entire career he's never seen this big a reversal from an anti-business administration to a pro-business administration.

We upleveled the ambitions dramatically. Clearly, the people around the table—Larry Ellison, Masa, and others—understood that vibe shift occurring in the White House. As I said to Bill, we have certainly intramural battles between Meta, Google, and xAI.

And Elon and OpenAI, but we're all on Team America, and unequivocally, this advances us on AI at a rate and a pace that I think is great for all of us. We appreciate your leadership on that, and appreciate you jumping in here for a few minutes today. We look forward to having you back on and going deeper with you.

René Haas

Yeah, one thing I might add to this—and I don't remember, Bill, whether it was a competing podcast that you were on where you were talking about innovation, that innovation thrives because you're 3,000 miles away from a certain city in the United States.

Yes, I was there the last 3 days. Freaking cold, by the way. But to Brad's point, you do the rounds of the dinners and the galas, but I was struck by the intent for a pace of change that was very stark. It felt much more business-oriented than anything else.

Certainly, in my career, I've spent a lot of time in DC working with different administrations and different pieces of government, setting policy. A lot of times, you're having to explain what it is that you do, what your products are, what they go into, and how they work. This time, the conversations are all about, how do we remove barriers to go really fast and do it here?

I think the biggest testament I can give to it is that 28 hours after taking office, Donald Trump was there with Masayoshi Son, Sam Altman, and Larry Ellison talking about gigantic investments. It was his first press conference, for gosh's sake. Yeah, I think that's something to look at. I totally agree, by the way. How excited is Larry Ellison, and is he really all in on this?

Larry doesn't do a lot of public speaking, and he lives in Florida. He likes the warm weather. He flew up to freezing Washington to stand outside and do an interview with Fox, and then he was there. I don't want to speak too much for Larry, but he's very committed to this.

Larry's got a lot of things he's doing with his Ellison Institute around advanced health research, all enabled by AI. As you and I have talked about, I think the last mile on this stuff is all about drug research and cancer research. Larry's very passionate about it, so this is a big connection to the stuff he believes very strongly in.

As you and I have talked about, Brad, I think the last mile on this stuff is all about drug research and cancer research. Larry's very passionate about it, so this is a big connection to the stuff he believes very strongly in.

Well, of course, the podcast that you referenced was Bill's famous talk that everybody ought to watch, “2,851 Miles from Washington,” by Bill Gurley at the All-In Summit from our good friends at the All-In Pod. It was an incredible talk, and what I took away this weekend was exactly what you did, Renee, which is this is the first time I've seen an intersection and partnership between technology and Washington the way I did this weekend.

Given that the whole field of battle for national security has moved to the playing field of AI, it couldn't come at a better time. Thanks for joining us. There's a lot more to talk about. We'll see you soon.

Thanks, guys. Thanks, Renee. That was incredible.

Renee answered a bunch of questions. What was your takeaway from that? What was your key takeaway?

I think it reinforced a lot of what you said earlier and maybe helped frame the math for the unit volume as you had structured it. I'm still left with a huge question. I didn't understand the answer about the structure. You can't invest equity and you can't put debt on something at a 3:1 ratio and not expect it to be a standalone entity capable of creating equity value. A captive, one-company subsidiary JV doesn't make sense to me, but there's more to unpack there.

Yeah, I've got 3 or 4 questions. Who's going to run it? Will they be great at it? We gave a lot of credit to Elon for being operationally tight in Memphis. We know how other companies are. Will that happen here? Is OpenAI the only customer? Will it serve other people? Is it a CoreWeave look-alike, maybe even on steroids? There are some things I still don't know.

Yeah, no, I think he started to answer some of those questions, but we'll go deeper with him in time.

6. DeepSeek continued

Hey, I had interrupted you. We had bookmarked DeepSeek. I want to finish that off, and then we're going to talk a little bit about just a couple of the other impactful EOs that came out of Washington. We'll do a quick little market check. Why don't you wrap on DeepSeek?

The point I was making when Renee popped in was that constraints can drive innovation. Here you have a case where the Biden administration—I think it was led by this guy, Alan Estes—and the Select Committee on the Chinese Communist Party inside Congress are constantly looking for ways to block China with export controls. Even on leaving office, Estevez said, “Hey, we won. We made it work.” Meanwhile, Eric Schmidt is saying we're behind.

I certainly don't think we're slowing anybody down, but in this case, maybe by giving them constraints, we actually created a world where they innovated in a different direction and created these hyper-small models that they didn't need massive training infrastructure for. I think the reality is we did do that.

We've talked about this, but some of the most remarkable entrepreneurs I've ever met are in China. They're resourceful. You tell them, “Well, you can't play with those tools,” and they go figure out how to do it with even lesser tools. Maybe they put themselves at equal footing, or even better footing, because they've learned to be more innovative in a way we weren't being innovative.

7. Trumps Executive Orders & Regulatory Freeze

There's a lot more to discuss about DeepSeek, but I think you're absolutely right. There is a real risk that well-intended legislation and well-intended efforts in Washington backfire and result in the exact opposite. Everybody wants to be tough on China, but being tough on China may not lead to China being slowed down, as evidenced by DeepSeek. That may be case exhibit number 1.

This brings me to the transition to talk about what exactly happened. We had the president of the United States literally at the Capital One Arena. I've never seen anybody touch this guy. He's the greatest marketer as president I think we've ever had. He does his inauguration, seemingly doesn't need sleep, and at the Capital One Arena he's signing executive orders.

One of those executive orders halts all federal regulations that are currently being promulgated unless they've been published in the Federal Register. One of the questions I had was about a very insidious piece of regulation, again under the guise of being tough on China, that I think is really going to slow down American AI. This is the Framework for Artificial Intelligence Diffusion, and it was promulgated by the Department of Commerce.

The idea was basically that Nvidia and all these companies could only send chips to a few friendly countries, and then it creates all of these different acronyms and levels for who can get what. People think there's a lot of regulatory capture in there for a certain company in Seattle. It hurts a lot of the smaller companies, but without a doubt, I've heard across the board from almost every company in the ecosystem that this is a terrible idea.

I was hopeful that this rule, which was promulgated, would be revoked. Unfortunately, the order said “unless it's been published in the Federal Register.” We went and researched whether it had been published in the Federal Register, and we'll show here that it has. It seems to me that that EO wasn't enough to roll back this diffusion rule that was promulgated by Commerce. Howard Lutnick or others are going to have to act specifically to roll back this rule, which I don't think is meant to be adopted for some 3 or 4 months anyway. I don't think it will get ratified, but it's an example of the new look I think the administration is taking at all these things.

Were there any EOs, Bill, that caught your eye, by the way?

8. Coordination in AI Legistlation

Just following up on that, we talked about this on the last podcast, but there are like 20 different state-by-state initiatives being pushed right now. I think the unfortunate reality is that there's so much effort and so much push to write regulation here. Maybe the intent of pushing at the state level is to provoke a federal piece of legislation, because the dumbest possible thing we could possibly have is state by state. It's just mud in the gears. Really, really stupid.

So, to put a finer point on that, you talked about SB 1047. Fortunately, it was vetoed in the state of California, but now you have the piece of legislation working its way through Texas that we talked about. I think it's roughly 25 states that, within this purview, have an SB 1047 equivalent that would be very problematic.

I'm happy to report that I spent a lot of time with an incredible congressman from the state of California, Jay, likely Obernolte. Jay was, or is, the chair of the Artificial Intelligence Task Force in the House. From the House of Representatives' perspective, he's on point developing a plan of attack.

It was great to hear Jay. I coordinated some meetings this weekend between him and a lot of CEOs of leading AI companies, and there is a plan by Congress. On the one hand, we need federal preemption. Clearly, this is a matter of interstate commerce, so we should have federal preemption. But because that's going to take some time, they also have a plan, Bill, that if any of these states were to pass something like SB 1047, they can issue a moratorium that would effectively freeze the state law for the period of time it would take them to get the preemption passed.

But that gets back to my point: that may have been the end goal all the way around. They may be happy with that outcome. Obviously, all this is going to fall on our friend Mr. David Sacks, so I'm sure we'll have more to hear from him going forward.

No doubt about it. Given what's at stake right now, can we just say how fortuitous it is to have people who actually understand this stuff, like Sacks, helping to coordinate the needs of industry and the various agencies of government? There's a lot of friction in this process, and I can tell you he is working tirelessly in information-gathering mode, trying to understand all these pieces and how to coordinate these priorities.

One of the things I think he and many others celebrated was that one of the EOs revoked by Trump was the Biden EO on AI.

Mm-hmm. One thing I would say about the Biden executive order is that there were many Republicans who said to me, “There's some good stuff in this. We need to keep certain things.” I think even Sacks and others acknowledge that there are some good things in there.

But the adopted approach was to start from a clean slate—tabula rasa. Rather than try to edit the Biden rule, just start with a blank sheet of paper. If there were good ideas in there, you can add them to an executive order on AI that could come out of this administration.

To me, when I think through the most important executive orders impacting Silicon Valley, one is on this diffusion rule. Bill, we've got to kill this. This was terrible legislation that was rushed out of Commerce so that they could say they were being tough on China.

We got the AI EO revoked. The next one for me was DOGE. This DOGE EO was frankly further-reaching than I expected, setting up a department, or effectively a task force—a SWAT team—within every federal agency. It's a 4-person SWAT team to go through the DOGE procedure.

We're now in a period of maybe 3 to 4 months where the president has said he wants a reconciliation bill on his desk. He wants to put us on a path to balancing the budget, and so DOGE is going to have to move very fast. We now see that V is apparently going to run for governor of Ohio, and so Elon is singularly at the head of DOGE. I think that will help make decisions go maybe even faster.

I was surprised at the level of coordination and reach that they already had, including around software modernization. There was an EO around software modernization across the entire government, so it was exciting to see that move forward quickly.

Yeah, and look, I think that's one of those things where most people are skeptical. I think they've been around the American government a long enough time that they don't think that type of radical change and efficiency creation is possible.

If you look at what likely Milei has done in Argentina and say, “What could go right if it were to work?” it would be so positive for the U.S. dollar and the American economy. It would be remarkably positive. I put it at high impact, still low probability in my brain, just because it's such a bureaucratic place, Washington. But it'd be great if it succeeds.

9. US National Debt and Balancing the Budget

Well, I would say that if you think about American national security, the 2 things that are top of mind for me are, number 1, not losing the race on AI. We've got to win the race on AI. The second one, to me, is our national financial security.

We cannot remain on a path to financial ruin if we don't make these changes. It's immoral to leave this level of debt to our children. We outlined a path on this pod for how we can balance the budget with $6 trillion of revenue and spending by the end of Trump's term.

I'm going to predict it here: I think in the president's State of the Union address in the first week of March, he's going to make a commitment to balance the budget in his first term in office. You don't have to do it in year 1. If we get to $2 trillion on DOGE, then you'll do better than balance the budget—you'll have a surplus by the end of his first term.

But I think a commitment to balance the budget in his first term would cause the bond market to react, which means, if the bond market gets bought and rates go down, our cost of borrowing will go down by $100 to $200 billion per year, positively impacting even more of what can be achieved here.

10. TikTok

Of course, the bond market hasn't believed—and shouldn't believe, because we've never delivered—that we actually can make these cuts. But I'm more and more convinced that the coordination between Congress, the executive branch, and DOGE is going to get something important done before we leave Washington.

I've got to make a few comments about the TikTok stuff that kind of popped up in Trump's initial press conference, and what I thought was different this time—and he then did it again after the Stargate conference—is that he's talking about putting partial ownership of TikTok on the United States government's balance sheet, which I think is new.

I think the way this happened is that he's been a negotiator his whole career, and he recognized that there was an asset that, if this law is enforced, goes away. Most of that equity value probably goes to Meta, and maybe a little bit to Snap, or to someone else if someone else pops up. He views enforcement as destruction of value that could be negotiated and captured.

I literally think that's what's going on in his head. What he may be underestimating is the willingness of China to just let it go away, or the unwillingness of someone to buy it and then give away half right away to the government. It'll be interesting. It sounds like he's only got 90 days to play this out. It's a fascinating one.

I mean, just to replay the weekend, Bill, TikTok went dark. They literally shut down the app, which nobody thought was going to happen. I happened to tweet that I thought it was going to be shut down for less than 36 hours. Sure enough, he took to the floor of the Capital One Arena and said, “I've brought TikTok back for the masses.” That makes kids and younger folks really excited. He's standing on the side of free speech, but yes, Trump is absolutely a dealmaker, and he thinks if the U.S. is going to turn this on, then we're entitled to something.

I think there are a couple of things. Number 1, he said multiple times, “I do well on TikTok. I like TikTok.” I think there is a part of him that just commercially knows there are 7 million folks who make a living on TikTok. He's done well on it, and he would like to keep it open.

But nobody on the planet wants a fair playing field more than Trump, and he realizes that American internet companies aren't allowed to be in China. He wants a deal, whether the deal is as you described or some other way. I think it's yet to be determined, but what we know for now is that it's alive and well.

For TikTok shareholders—and as a reminder to folks, we're investors in ByteDance, TikTok's parent company, since 2016—I just think they want resolution. TikTok U.S. is a small part of that global business. Maybe there'll be a commercial deal worked out, but to your point, it's likely to be done or not done in the next 90 days.

Well, they've had 180 days to do it, so the odds that they actually want to do it, I think, are low, because they could have done it already. Based on what I've read, I think it's out of Trump's hands. He found a way to delay it, but Congress already voted and the Supreme Court already backed it up. He can't veto what they've already done, so it is going to come to a head pretty quickly.

Yeah, it was funny. In the press conference yesterday on Stargate, he was asked a question on TikTok, and he started laying out his case—what he wants. It could have been like he was running the investment bank at Goldman Sachs, the way he laid it out. It was pretty brilliant.

He said, “Hey, Larry”—talking to Larry Ellison—“why don't we just live-negotiate the TikTok deal right here? I'm going to tell you what I think. You tell me whether or not you want to do the deal.”

At the end of the day, remember, the Oracle Cloud runs Project Texas, which is where TikTok U.S. is run, so Larry Ellison has a big stake here. He was also there yesterday talking about that. Somebody asked the president, “Would it be okay if Elon bought part of this?” He said, “Yes, I'd be fine if Elon bought part of it.”

So might Larry and Elon be the 2 buyers of 50% or more of U.S. TikTok? It's going to be just another part of— But there are 3 parties. China has to want to sell, ByteDance has to want to sell, and the U.S. has to agree. There are 6 parties all together, so this will be a tough one to land. It's going to be fascinating.

11. Tech Check

Why don't we maybe wind up with a little tech check?

Sure. The first thing I saw, Bill, is that on the actual day of the inauguration, one of our heroes, likely Druckenmiller, was on CNBC. He talked about his observation on this administration, saying we're going from the most anti-business administration of his lifetime to the most pro-business administration of his lifetime.

Think about this guy. He's seen it all, and Stan is not a person prone to hyperbole. When you look at what's happening in the market just year to date, I think the Nasdaq is up about 4% through today. That's on top of it being up a fair bit since people started expecting that Trump was going to win the election.

The fact that he made those comments, I think, you have to take as very significant. But there are 2 other things Stan said in that interview. He said, “Just because I believe that doesn't mean I'm all in on the markets.” He said things are priced reasonably high.

If you look at the S&P, we think the S&P is baking in a big earnings acceleration this year, and that's on top of already high multiples. The S&P ex-Mag 7 has a big hurdle this year. If you look at what's expected out of the Mag 7, it's a pretty big earnings deceleration, and the multiples are pretty consistent with where they've been over the last 5 years. We'll post both of those charts here.

Number 1, he said we're not all in on the market. Then he talked about the 10-year, which we said in our first show this year to watch. That's going to determine where the market goes. It went as high as 4.85%. We're now down today closer to 4.6%, and that's really about whether or not folks think reflation is reengaging here and whether we're going to be faced with higher interest rates.

It's really interesting: Lennar, the homebuilder, was down a ton yesterday; Ford, the auto builder, was down today. At the same time, these tech stocks are screaming. Why? Because interest rates are high and the economy is restrictive. Not all parts of this market are benefiting, and I think that will be concerning to the president. He said he doesn't like these interest rates. He thinks the interest rates are too high.

Stan said, “I feel great about the administration. It's super pro-business, but we've got these 2 things: high valuations and maybe interest rates that are going to go higher. That could ruin the party.”

We've gotten off to a strong start here in the early part of January, but for managers like us, we're paying close attention to those things. I'm kind of on record saying I don't think inflation is going to reignite and that I think rates will—we're probably seeing the top in rates for the next 3 or 4 months. But I don't know. I'm not a macro forecaster. We just have to look at them and take them as they come.

If rates were to go to 5% or 5.5% and we saw that happening, our exposures would come down. But I would say right now it's been a pretty incredible start to the year. Any comments on the Netflix quarter, which looks like it's up $40 billion today in market cap?

I mean, just an extraordinary company, incredibly well run. Frankly, it's a tailwind—another tailwind today—for all internet companies. Remember one of the things I said we were worried about with these companies? We said FX was going to be a big headwind because the dollar has strengthened tremendously on a year-over-year basis.

We thought for most of these companies, FX was going to be a 2% or 3% headwind. They said they took up their guidance, Bill, even accounting for the FX headwind. It's a much bigger raise in guidance on an FX-adjusted basis than people expected.

It just goes to show you—I was with a well-known CEO of an internet company this weekend, and he said, “Brad, we've doubled our revenues, and our costs keep coming down because of AI.” I've said that I think this moment that we're in, the next 5 years, is going to be a golden moment of margin expansion for technology businesses and, frankly, all businesses.

Human productivity is going to go up, which means the costs to do exactly the things you're doing in your business are going to come down. For market leaders who can hold on to that revenue and not have it competed away, their margins are going to expand. Look at what's happened at Meta and these other companies. I think that's accelerating in 2025 and 2026.

Everything's price-dependent, and things can change really rapidly, as the first week of this year showed. But it's fun to be with you. It's going to be a fun year to do this. We'll have some more great guests on in the next few weeks, doing cameos and the long form. I appreciate you doing this with me again. It keeps me sharp.

All right, man. Take care. Great seeing you, Brad.