TBPN · 7 min · Oct 27, 2026

GPU supply is tight despite the capacity coming online

Gavin Baker says compute prices and future cluster costs are rising, even as new capacity is planned. He argues the market needs experienced operators who can bring GPUs online quickly.

Gavin Baker is managing partner of Atreides Management.

Runs the Atreides Management investment firm, a technology investor and former Fidelity OTC Fund portfolio manager.

John Coogan, Host · Co-host, TBPN
Co-founded Soylent and Lucy; OpenAI bought his show TBPN in April 2026.
Jordi Hays, Host · Co-host, TBPN
Founded Branded Native; co-founded fintech Capital, which Rho bought.

About this segment

Gavin Baker says compute prices and future cluster costs are rising, even as new capacity is planned. He argues the market needs experienced operators who can bring GPUs online quickly.

From “Gavin Baker on Anthropic's S-1, GPU Prices, and SpaceX”, Oct 27, 2026. Watch the full episode

Answers “What will AI look like in the next few years?” See all 122 answers

Answers “Is there still a GPU shortage?” See all 21 answers

“But even with all of that supply coming on, it is hard for the companies in our portfolio to find compute even well into '27.”

Gavin Bakerat 14:34 · TBPN14:34

Takeaways

Compute prices have risen throughout October.
Clusters for 2027 can cost much more than current ones.
Fast, competent operators matter as much as GPU supply.

In this episode

12:46GPU supply is tight despite the capacity coming online6:56
7:56AI agents are helping investors route information faster3:51Up next
34:35Public listings could unlock more capital for AI3:49
3 more moments in the full episode

Transcript

From the transcript TBPN publishes with the episode.

12:49
Yeah. I think that's I think that's right. It's it's easy to say. It's it's hard to internalize and believe. And October has kind of been a strange month because the price to rent, essentially any kind of a GPU, has gone up every day. So the spot market for compute is getting tighter. And it's not just the spot market, you know, it is... I do love reading these primers on x from, you know, commodity and macro investors about these super complicated concepts called backwardation and contango Yeah. And what they might mean for the forward curve of compute and the future price of compute. You know, these are contango backwardation.
13:34
But, you know, Silicon Data, which I do think is, you know, we're... My firm is an investor in them, but I I do think they're the most authoritative data source. Carmen Lee, the founder, stood up at GTC and said that the compute market is now in contango. Mhmm. The, you know, the forward... It's not just that the spot market is tight. If you want a cluster a year from now, you have to pay a lot of money. Yeah. And then this is and this is maybe another, you know, discontinuity between public and private markets. But just the number of companies we have in our portfolio who are like, okay, we got a cluster of 4,000 Blackwells and, you know, and we got it in May or June '26, and we wanna bring on another cluster of 4,000 in September 2027.
14:25
Yeah. And they're shocked by how much more they have to pay in the future. And it's not like they're... Everybody is unaware of this... Everybody's aware of the supply that's coming on. But even with all of that supply coming on, it is hard for the companies in our portfolio to find compute even well into '27. Mhmm. So for, you know, all the commodity experts who are focused on backwardation and what that implied, I got news for you. We're not in backwardation anymore.
14:58
Mhmm. And then I do think as far as how much capacity is going to come online, one thing that I think is actually healthy about what's happened in these financing markets is I hope it concentrates capacity and supply in the competent operators where they can finance it through prepayments. You know, you can kind of finance these GP purchases two ways. By the way, CoreWeave is great. CoreWeave runs GPUs well. They bring a lot...
15:26
They they are very very good. They just sadly locked in a vast amount of compute and long term contracts that are now way way under market. You know? And I should
15:39
be doing So they sold their they sold their GPUs for fixed prices, and that's why they're getting hammered right now. It's not that they don't have compute. They're not bringing it on. They have the semi analysis diamond tier. Everyone knows that clusters Yeah. Work. They're good. They are reliable, but they just sold for too cheap, and they should have hold... Held on or had variable pricing.
16:00
Is is you think Elon studied that and said, like, we're not gonna make that same mistake, and that's why they're adopting this more short term contract...
16:08
I don't... Elon is a first principles thinker. I don't think he was probably studying CoreWeave or Nebius or anyone Yeah. He was probably reasoning from first principles. But the conclusion he came to is prices are going up. We... If we are willing to sell into the spot market, we can have paybacks well under a year Yeah. Get big prepayments Yep. And that's, I think, that's a better probably financing model for the market we're in. And what I also like about it... And by the way, CoreWeave, I think, is very aware of this. They're trying to get more spot exposure. They're trying to, you know, do less of these big long term off take agreements.
16:50
Mhmm. But standing up a big GPU cloud... I I should say, bringing on the power, standing up the cluster, it's like running a Formula One team. Yeah. And a bunch of people who have no real relevant experience, you know, whenever they acquire some land and some power, and then, you know, they get a zero down payment off take agreement from someone. And a lot of that is being cleaned up and I think, you know, supply is being what... This is a forcing function to allocate GPUs to people who can actually bring on a cluster, energize the GPUs, run it well. Like, you know, running a Formula One team, it's like really hard. You gotta like pack everything up, put it on a plane, go to a new city, unpack it, set it up, set the car up. You need all the people, you know, to, like, you you know, whatever.
17:51
Change the tires, to change the rear wing if you get in a collision, to, you know, do whatever needs to be done during the race, to do the strategy, to communicate with the driver. It's really, really hard if... And it's not like, you know, some of these people, they don't have the relevant experience. You know, Corwee, Crusoe, Nevis, SpaceX, the big hyperscalers, you know, these are companies that could bring on GPUs and energize them and and run them. But...
18:22
And I do think something that was announced at GTC Berlin is actually pretty important for this, and that's that DSX pod, which is gonna make it a lot easier, you know? Oh, sure. Seven racks. It comes preassembled. Yeah. You know, you just kind of literally plug it in and it works. Yeah. And right now, when you get the GPUs, you know, maybe if you're if you're SpaceX and you're really really exceptional at execution in the physical world, you can get them energized quickly. But a lot of people, it takes them six months, you know, three months, six months to get them energized, really get them working. And if we can, through, you know, this new product from Jensen, get that time down, I think that's that's also healthy because that effectively takes a lot of working capital out of the market. You know, you have as these numbers are getting bigger, if it takes six months to energize them, you know, that could be hundreds of billions of dollars of assets that are not generating revenue. So I think that's another...
19:18
That's a product innovation that's gonna help with this. But I think this is, you know, we we will see who who energizes GPUs in in in in '27. I'm I'm confident that the, you know, the companies I named, I think they're gonna, I mean, the future is always uncertain, but these are companies with a proven track record of execution. Yeah. And we'll see about others.