Anthropic Bets $1.25B per Month on xAI's Colossus: What the Compute Deal Means for Builders

On May 20, 2026, SpaceX's S-1 filing made the price of one of the strangest compute contracts in AI history public. Anthropic has committed to pay xAI roughly $1.25 billion per month through May 2029 for the full output of the Colossus 1 data center near Memphis, Tennessee. Over the life of the contract, that is in the neighborhood of $40 billion in revenue to xAI. A discounted rate applies for the first two months while xAI completes the ramp-up.

Anthropic also published a companion note titled higher limits and a SpaceX compute deal, and xAI's own announcement confirmed the partnership. Below is what the deal actually says, why both companies signed it, and the three things this changes for teams building on Claude.

The Deal in Numbers

Anthropic gets the entire output of Colossus 1: more than 300 megawatts of compute capacity and over 220,000 NVIDIA GPUs.

The monthly payment is roughly $1.25 billion, discounted in the first two months during ramp-up.

The contract runs through May 2029. Either party can terminate with 90 days' notice.

Anthropic separately confirmed an expansion to Colossus 2 as that facility comes online.

That 90-days-out clause is the structural detail worth re-reading. A contract this large with a clean unilateral exit is unusual and reflects both parties' uncertainty about the hardware roadmap. Blackwell Ultra is ramping in production and Rubin sampling is on the horizon. Neither side wants to be locked into a generation of accelerators that will look slow in 18 months.

Why Anthropic Signed

Anthropic was already growing through one of the most aggressive demand curves in the industry. The Wall Street Journal reported that Anthropic is on track for its first profitable quarter, with Q2 revenue projected to more than double to about $10.9 billion and an operating profit estimated at $559 million, arriving two years ahead of internal projections. Inference capacity has been the rate-limiting factor; signing up for 300 megawatts of dedicated GPUs is the most direct way to unblock that.

It is also a hedge against single-supplier risk in cloud compute. Anthropic's existing relationships span AWS and Google Cloud; adding xAI's Colossus as a third, fully dedicated source diversifies exposure if one provider has an outage, a price spike, or a regional supply problem.

Why xAI Signed

On its face, leasing your largest data center to a direct frontier competitor is counterintuitive. The financial side answers part of it. A guaranteed $1.25 billion per month for several years is a material multiple of xAI's previous run-rate revenue and turns Colossus 1 from a cost center into a near-utility revenue stream.

It also matches the larger pattern after the SpaceX-xAI merger: xAI's compute build-out is now best understood as an AI infrastructure business with a model-training tenant attached, not the other way around. NVIDIA's most recent quarter helps explain why that arbitrage is attractive right now: $81.6 billion in quarterly revenue with $75.2 billion from data center, with CEO Jensen Huang describing demand as having gone parabolic. If you own GPUs at scale, the best near-term return is to rent them.

Three Things This Changes for Claude Builders

1. Rate limits are about to ease. Anthropic explicitly framed the same-week update under higher limits. If you have been bouncing off tier limits on Claude, expect material headroom across the API in the months ahead as Colossus 1 capacity flows into serving.

2. Long-running agent workloads become more viable. Sustained throughput for multi-hour agent runs is a function of capacity, not model size. With 220,000+ GPUs of dedicated headroom, long-horizon coding and research agents on Claude get a real backend.

3. Vendor diversification just became cheaper to do right. With Anthropic, OpenAI, Google, and xAI all running parallel compute build-outs, the cost of running a multi-provider AI stack is no longer dominated by capacity scarcity. Architect your applications to be model-portable now, because the supply side is finally catching up to demand.

The Bigger Picture

A frontier AI lab paying $15 billion a year to a competitor for GPUs is the cleanest signal yet that the inference-economy phase of AI has decoupled from the model-rivalry phase. Whoever owns the silicon wins regardless of whose model is on top this month. That is also why the 90-day exit clause is in there. Both sides know that whichever next-generation accelerator wins the next 18 months will reshape the math of this contract, and both want the option to react.

Sources

TechCrunch on the contract value: https://techcrunch.com/2026/05/20/anthropic-will-pay-xai-1-25-billion-per-month-for-compute/

Anthropic announcement: https://www.anthropic.com/news/higher-limits-spacex

xAI announcement: https://x.ai/news/anthropic-compute-partnership

Axios on the contract size: https://www.axios.com/2026/05/20/anthropic-spacex-compute

Axios overview of the May 21 news cycle: https://www.axios.com/2026/05/21/ai-news-cycle-openai-anthropic-spacex