Anthropic is renting Colossus-1, a Memphis data center owned by SpaceX/xAI with 200,000 Nvidia Hopper GPUs and 30,000 Blackwell GPUs. The deal lets Anthropic raise token limits for Claude and gives xAI revenue while it focuses on the larger Colossus-2 facility. It signals xAI lacks a competitive frontier model to justify keeping the compute.
Chapters
What does an Electrolux oven and Anthropic's Claude code have in common? This is not a trick question, because believe it or not, if you own an Electrolux or Frigidaire branded oven and you're using Claude code, there's a chance both originate from the very same factory. That's because the Colossus 1 data center in Memphis, Tennessee was built inside a former home appliance plant. At its peak, the factory was producing hundreds of thousands of ovens, employing over 1,200 people before Electrolux closed the plant in 2019 and shifted production to Springfield. A few years later, Elon Musk bought the now empty Homer Blinds factory and retrofitted it into one of the largest AI data centers. But as fun as this story is, we are not here to talk about ovens. We are here to talk about Anthropic buying the entire Colossus-1 compute capacity And what this deal means for Anthropic and xAI alike. There's a reason every stock somewhat related to semiconductors still only knows one way, up. It's because even after a couple of years of high-intensity compute buildout, compute for AI is still a limited resource. That's why AI models are more and more restrictive with their token use, And some even started to downgrade their models to save precious compute. This especially affects the latest frontier models that achieve the best results, because they often use the greatest number of tokens to do so. So it comes as no surprise that Anthropic is looking for more compute to power their latest AI models, such as Claude Code and Claude Opus. And with the Colossus One data center, they have just found that. The deal with SpaceX, Since February, the parent company of xAI allows Anthropic to use all of the compute capacity that Colossus-1 offers. But how much is that? The official blog post talks about over 220,000 Nvidia GPUs, but you wouldn't watch a SemiAnalysis video if you didn't want to know a bit more than that. According to our information, Colossus-1 mostly contains Nvidia's last-gen Hopper GPUs, approximately 200,000 of them. but it was upgraded with an additional 30,000 Nvidia Blackwell GP200. We can also confirm the roundabout 300 megawatts power draw. And while 200,000 Hopper and 30,000 Blackwell GPUs sounds like much, Colossus 1 barely made it into our top 10 data center list in January of this year. And because it hasn't been upgraded since then, but plenty of other sites have, it's very likely not a top 10 site anymore. Which doesn't mean it's not offering a ton of AI compute, because it still is, just not top 10 material anymore. By the way, let me know in the comments if you want an update to the top 10 AI data center video. For Anthropic, the Colossus One deal with xAI has an obvious and immediate benefit. It allows Anthropic to gain access to a lot of additional AI compute in a compute-constrained environment. And we can see instant results. In the same blog post that announced the compute deal with SpaceX, Anthropic detailed how the additional compute supplied by Colossus-1 directly allows them to increase Claude Code and Opus limits. And by increasing token limits, Anthropic improves the accessibility and competitiveness of their frontier models. Because in a compute-constrained environment, model availability can matter almost as much as model quality. If you can't supply your model to customers, it doesn't matter how good it is, because no one can use it. So, Anthropic gets AI compute to run their latest frontier models. But what about xAI? Aside from the obvious monetary value, this deal offers more insights into the focus of xAI. First, it indicates that the Colossus 2 data center, which is literally just a few miles down the road, has clearly become the number one focus for their in-house compute buildout. At the end of last year, Colossus 2 already contained over 110,000 Nvidia Blackwell GP200 at around 350 to 400 megawatts. Today it's much closer to 1 gigawatt. It might just be that Colossus 1 doesn't really matter anymore for XAI. But it also shows that XAI doesn't have a competitive frontier model at the moment, Because even though Colossus 2 is much bigger than its little brother, it's still a compute-starved environment. And if xAI had a competitive frontier model, they could surely make use of an extra 300 megawatts of AI compute. But they don't. The contract with Anthropic also indicates that xAI probably doesn't expect to have a competitive model in the foreseeable future, because again, if they did, they probably wouldn't rent out precious AI compute resources. Anthropic gaining access to Colossus 1 is a very interesting deal. On the surface, both companies benefit. Anthropic gains access to existing infrastructure, which immediately increases their compute capacity and directly impacts their Claude delivery. And xAI/SpaceX surely does get a decent amount of money for a data center they apparently have no internal use for. This means xAI can focus on Colossus 2, but it also indicates that xAI doesn't expect to compete with other frontier models in the near future. In any case, AI servers in a former home appliance plant are now running large language models. And I think that's amazing because no matter if ovens or Claude code, both can be used for cooking up something beautiful.
