Data center bottlenecks could make hyperscalers kingmakers

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

1. 55:06

   she's going to let people build them out until 2028 but it's unclear what happens after that point. You're obviously connected, I think, Chamath, to a lot of the major players in the AI space. What's going to happen with data centers? Can the Democrats afford to pop the AI bubble by then putting cold water and slowing down data centers? Because that will tank the economy. I think we all understand that. So is that in their best interest? And is this posturing and positioning to get better deals for their local communities, as is their job? There's an incredible video where the SpaceX team went to Cochise County in Arizona to present their plans for this new six gigawatt data center. And if I had to characterize that video, it's on X if you want to find it. But it's these two very bright, earnest people from SpaceX trying to explain

2. 55:57

   how they're going to reduce electricity costs by 40. percent. They don't use water. They'll reduce property taxes by 20 percent. They will be able to fund enough revenues to pay off a bond in the county. And it's just a litany of things, yeah? Meanwhile, the audience is just screaming. And what you see, the two people were very poised. Riley and Katie, those are their names, good SpaceX employees. They should be proud of how they conducted themselves. But they were getting run over, if you listen to this video. And my takeaway is, Jason, that this is an emotional issue that is not tethered to the facts, as Friedberg says. It represents a totem around the oligarchy and the economic fissure that exists in the US economy. But it also represents that the deals that have happened are a little gratuitous,

3. 56:49

   Meaning, should it really be the case that the most powerful seven or eight companies need a tax break to invest $30, $50, $80 billion to then generate $300, $400, $500 billion? Probably not. Now, they're doing smart deals because that's what made sense. But I think what the incoming governors are going to say is these things are off the table. And instead, we want to have directly measurable economic impacts before we even entertain this.

4. 57:16

   I think what it means is we're going to see a slowdown. And I think what that does is it benefits the hyperscalers because they have existing capacity. I think it's very bad for the neoclouds because they're renters. And it's pretty bad for the foundational models because they are so short this capacity and they need it to grow. So to the extent that these frontier models are so much better than everybody else's product, they are the ones that need this capacity. It sets the hyperscalers to be a king maker. And they have this interesting game theory now. Can you just say who the hyperscalers are, just so people understand the difference between them? Yeah. Amazon, Meta, Microsoft, Google, SpaceX. And what they'll be able to do is make a decision. Path number one is continue to rent in the way that you're doing it today. So the status quo, rent to some customers, large corporates,

5. 58:11

   right? Rent to the neoclouds, rent to the foundational models. That's what they're today. Everybody does that. But then there's the path less traveled, which all of a sudden becomes really interesting. You could decide to take all of the incremental compute and advantage your own products. So Google has a ton of capacity coming online. They could say, I'm going to give it to Gemini. What does that do? It actually allows Gemini to catch up in quality relative to the foundational models. Why? Because if you constrain the amount of assets that can actually generate compute, and a smaller surface area of it is available to the leading companies, those leading companies will have to pay more, the user experience will decay, and then other products can have time to catch up. And now that is important, because Anthropic and OpenAI are

6. 59:03

   in a very fast recursive loop, they're on a 60 to 90 day loop in improving their models. So now all of a sudden, without an infinite amount of supply coming online, you can slow these folks down. That's the path less traveled. Then there is the crazy path. Crazy path. you should look for this one very narrow thing you and i talked about this before on the pod but spacex and anthropics compute deal has a mutual 90 day out and propic can term in 90 days spacex can term in 90 days terminate if spacex says i'm going to terminate this compute in 90 days that is the king making move why it's a disproportionate amount of the compute that anthropic needs it allows spacex to reprice that compute at the spot rate, which is probably double the market rate, and it's probably not just double, it could probably triple. So whatever revenue you see on SpaceX's balance sheet from Ant could two to three x. And Ant would have no choice except to immediately capitulate and take the higher price.

7. 1:00:25

   it would also show all of the other hyperscalers that they are very far away from having economic yield that is optimized and so then it forces them to consider should i be giving it to myself should i be selling it to the other people it then forces the neo clouds to take all of this money that they've raised whether it's debt or equity and say i have two choices which is to buy as much compute today as I can at whatever price it's available, because the other path is assured destruction and bankruptcy for me because I'm just out of the game.

8. 1:01:02

   And so what do they do? They will take the last gasp and do whatever they have. So I think all the game theory, Jason, literally comes down to this one deal and whether Elon decides to rip it because if he rips it, watch out. Yeah. And the other two factors here,