$45B Colossus Deal Secures 220K GPUs for Claude Inference

SpaceX has filed its S-1 registration statement for an upcoming IPO, detailing a massive infrastructure agreement where Anthropic will pay $1.25 billion per month for access to the Colossus data center. The transaction spans three years from May 2026 to May 2029, totaling roughly $45 billion. The deal provides Anthropic with immediate access to one of the largest single compute clusters in the world.

The hardware footprint includes more than 220,000 NVIDIA GPUs, distributed across H100, H200, and GB200 accelerators. The arrangement guarantees Anthropic over 300 megawatts (MW) of dedicated power capacity to run these clusters.

Infrastructure Scale and Workload Migration

The agreement grants Anthropic the entirety of the Colossus 1 facility in Memphis, Tennessee, along with additional secured capacity in the newer Colossus 2 campus. The availability of this infrastructure stems from a recent corporate restructuring. In February 2026, SpaceX acquired Elon Musk’s AI startup xAI, consolidating operations under a new subsidiary called SpaceXAI.

Following the acquisition, SpaceXAI migrated its internal model training pipelines, including workloads for Grok 5, exclusively to the Colossus 2 facility. This shift left Colossus 1 operating at just 11 percent utilization. By leasing the underutilized facility to Anthropic, SpaceX converts a high-depreciation hardware asset into a massive revenue stream. The $15 billion annual payment equals nearly 83 percent of SpaceX’s entire $18.67 billion revenue from 2025.

Inference Economics and Availability

Anthropic is utilizing the Memphis cluster strictly for inference rather than foundation model training. The compute handles the high-throughput requirements of Claude Pro, Claude Max, and Claude Code. As a direct result of the capacity injection, Anthropic has already doubled usage limits for developers scaling Claude Code environments.

The contract includes discounted billing rates during an initial hardware ramp-up period spanning May and June 2026. Either party can exit the agreement using a flexible 90-day notice termination clause, providing an escape hatch if compute requirements or hardware lifecycles shift dramatically before the 2029 expiration.

Orbital Compute Ambitions

Beyond the terrestrial hardware in Memphis, the two companies are exploring extraterrestrial data center deployments. Prior to the S-1 filing, Anthropic confirmed it is partnering with SpaceX to develop multiple gigawatts of orbital AI compute capacity.

Deploying solar-powered data centers in low Earth orbit aims to bypass the severe power, land, and cooling constraints currently bottlenecking ground-based AI infrastructure. SpaceX noted in its prospectus that it views space-based compute as a new market with a $1 trillion valuation ceiling.

If you are building complex multi-agent systems that require continuous context routing, this infrastructure expansion directly affects your production environment. The Colossus deployment ensures Anthropic has the raw hardware scale to support high-volume API polling and aggressive autonomous loops without severely throttling rate limits.