Anthropic secured access to SpaceX's Memphis data center, Colossus-1, which has over 220,000 NVIDIA GPUs and 300 MW of power. The deal will increase Claude token-per-minute limits from tens of thousands to millions for Pro and Max users. The partnership signals how control over energy and GPU supply is reshaping AI's power dynamics.
Anthropic raises Claude code usage limits, credits new deal with SpaceX. Here's what happened and here's what nobody is talking about. Elon Musk is really excited to partner with Anthropic via SpaceX to get more compute, essentially. At its Code with Claude developer conference on Wednesday this week, Anthropic announced a deal with SpaceX to utilize the entire compute capacity of SpaceX's data center in Memphis, Tennessee. And on stage at the conference, CEO Dario said the deal was intended to increase usage limits for Anthropic's Pro and Max plan subscribers. That's us. And for other AI builders out there, you must be happy too when you saw the announcement that we will get more usage limits. And so in this, you could see, uh, this is a table that kind of outlines the changes for Opus from tier 1 to tier 4 usage limits. As of, uh, May 6th, a couple days ago. So maximum input tokens per minute at Tier 1 went from 30,000 to 500,000, and then maximum output tokens per minute 8,000 to 80,000. And then on Tier 4, it's 2 million to 10 million and 400,000 to 800,000 respectively. Anthropic claims that the deal gives the company access to more than 300 megawatts of new compute capacity. SpaceX focused its announcement on the capability of the Colossus-1 supercomputer that I think that's at the center of the deal. Quote, Colossus One features over 220,000 NVIDIA GPUs, including dense deployments of H100, H200, and the next-gen GBE200 accelerators. Anthropic also expressed interest in working with SpaceX to build up multiple gigawatts of orbital compute capacity. I think that's like the kind of underlying story there, which is, oh crap, like we're going to start getting compute from orbit, which is just like from a nerd perspective, that's just super sick. So like, we're just living in the future and this kind of stuff is becoming normal to talk about. Jonathan, from your CTO perspective, kind of strategically and tactically, like thoughts on this move? Obviously we're happy about it. Uh, from a builder perspective, usage limits go up. I'm going to be cloud coding a lot more.
You know, as a builder, you immediately get this kind of like, oh, that's great. Like I've run into that problem and maybe I won't have that problem anymore. And like has affected my usage where it's one of those things nudging me closer to Codex and using Codex more frequently. On the other hand, like, you look at the background of what's going on here and you wonder, Musk is basically giving SpaceX's leftovers to Anthropic in this moment, you know, and how long is that deal going to last? He says, you know, we're using Colossus 1, so you can use our leftover. You can use Colossus 2. Uh, we have some extra capacity. And I think there's a broader thing going on here where there are constraints to how these AI services are delivered. There's energy and there's compute. Which includes the hardware. And we're going to see shortages of both of these ongoing over the, over the years. And the people that sort of hold the most GPUs, that hold the most energy to deliver them, uh, are being kind of sort of power brokers. And I think that's where Elon has found himself or explicitly positioned himself. That's really interesting. And I think people that are saying, you know, like that OpenAI and Musk are playing the long game of looking at like, how do we secure power? How do we secure a semiconductor supply chain? How do we get the data center capacity? They're very well positioned, whereas Anthropic took a slower approach. And so they're more reliant on people that have those resources in order to, like, continue to provision at scale. And we're just seeing one of the first deals. I think this is being an ongoing thing and probably a source of a lot of geopolitics kind of drama as time goes on.
Say more on the geopolitics drama for those that are not geopolitics nerds. I feel like Jonathan knows everything that's going on in the world, like, a lot of the times, at least when it relates to existential risk of humanity. Tell the audience more about why this power broker game is really just one domino, the first to fall in your prediction of, hey, there's going to be some geopolitical implications from this, where from, from my view, it's simply put, we cannot get enough compute to meet demand. And, uh, these frontier labs are going to look for more compute.
Well, I guess you can look at like, where are the potential shortages, you know, in power and semiconductors, even raw materials that go into building semiconductors.. And one interesting bifurcation is, you know, the energy strategy of the United States versus China, for instance. China's producing more energy than us. And so they have like more overall electricity capacity, and it's gonna take time for the US grid to evolve and, and like respond to the demand of the energy that these data centers have. It's kind of like building a building in San Francisco. You know, it's very slow. There's a lot of permitting. There's a lot of people who have to say like, this is okay. It's a long process. We can't do this overnight.. And, um, it's one of the reasons that, uh, Musk is in particular is talking about space data centers because then you have an access to solar power, uh, that not sort of, sort of like where the regulation isn't there yet. And, uh, you can build, uh, potentially massive data centers that are just, uh, floating around in space. And it's kind of like interesting in theory, but that's also a really risky sort of project vector that could take years to, to unfold. So what you get is like there's an energy constraint. We can only unblock it slowly. Then you move to the semiconductor constraint and GPUs. And like, you know, everybody knows that NVIDIA GPUs are very difficult to get right now and capacity is sort of like booked up over the coming years. One of the constraints is manufacturing and TSMC in Taiwan is the singular choke point to all semiconductors being built. And not only is it, is that like kind of a monopolistic type of situation,, but it's also under threat by, you know, geopolitics from China being interested in taking over Taiwan. So there's like significant threat and there's a slow process there of unbundling what's going on in TSMC and bringing it into the United States, into Arizona fabs. And there's also processes like Elon Musk's TerraFab, you know, which would plan to like build at scale. These things will all take time. So I think it's like there's a known set of constraints. All of them are slow to unbundle. Um, and that doesn't even get into like the raw material constraints that could, could hit us because of global supply chains being disrupted., and us just not having enough, like, cadmium, for instance, to, like, support the semiconductor production process. So this is gonna be like a supply chain whack-a-mole, and, um, you know, I think we're gonna see this and feel this as consumers, uh, as time goes on.
