Anthropic is leasing megawatts, and xAI is testing a new AI infrastructure business
A dramatic infrastructure-scale view of Colossus 1 as a compute power plant, with Anthropic demand flowing into xAI-owned capacity and a visible 300 MW energy scale.📷 AI-generated image / TECH&SPACE
- ★Anthropic is reportedly paying xAI $1.25 billion per month for 300 megawatts of AI compute capacity.
- ★The contract runs through May 2029, but either side can terminate it with 90 days’ notice.
- ★xAI is testing a neocloud model where surplus infrastructure becomes a separate revenue stream.
The agreement between Anthropic and xAI may look like a commercial footnote in the larger model race, but the numbers say otherwise. According to TechCrunch, Anthropic will pay xAI $1.25 billion per month for compute capacity. The contract covers 300 megawatts from the Colossus 1 data center near Memphis, Tennessee, runs through May 2029, and could generate more than $40 billion in revenue for xAI.
This is not a routine server rental. It is the price of access to infrastructure at a moment when leading AI labs are competing not only on model architecture, but on who can secure power, GPUs, cooling, networking, and physical data-center space fast enough. Anthropic builds its own models and products such as Claude, but that work requires continuous access to enormous clusters. If capacity can be leased faster than it can be built, compute leasing becomes a strategic weapon.
For xAI, the deal is even more important because it changes the company’s profile. xAI is not only the lab behind Grok; it is also the owner of infrastructure that can be monetized when it is not fully used for its own models. The research brief says the agreement lets xAI monetize unused compute capacity and describes a dual monetization strategy with multiple paths to returns on invested capital. In plain terms: if Grok scales, the infrastructure feeds xAI’s own product; if capacity remains free, it can be sold to a rival.
The 300-megawatt Colossus 1 contract turns xAI into an infrastructure supplier, not just the maker of Grok.
A close operational view of a contract-control desk where monthly compute allocation, 90-day exit window and Claude/Grok workload split are visible as interface layers over server telemetry.📷 AI-generated image / TECH&SPACE
That is where the broader neocloud pattern appears. AI companies that build their own data centers aggressively enough can begin to behave like specialized cloud suppliers for other AI players. This is different from the classic cloud story because the buyer is not just asking for elastic servers. It is buying a packaged block of power and compute for training or serving large models. A 300-megawatt contract is therefore almost as much an energy story as a software story.
The 90-day termination clause keeps the deal from looking like a fully locked bet through 2029. Either side retains an exit if the economics change, if better infrastructure becomes available, or if training and inference needs shift. Even with that flexibility, a monthly price of $1.25 billion shows how high the premium has become for reliable capacity that already exists.
For the market, the real question is whether contracts like this become normal. If they do, AI labs will no longer be neatly divided into model builders and infrastructure sellers. The largest players will try to do both, because in an era of expensive compute, idle capacity is too costly to leave unused. Anthropic gets access to a large compute base. xAI gets revenue that could reshape its balance sheet. The rest of the industry gets a blunt signal: in AI, having a strong model is no longer enough. You also need somewhere to run it.
