SpaceX S-1 Reveals Anthropic Will Pay $1.25B/Month for Colossus

SpaceX's S-1 filing, made public Wednesday as part of its long-awaited IPO paperwork, disclosed that Anthropic has agreed to pay $1.25 billion per month for access to xAI's Colossus and Colossus 2 supercomputing clusters through May 2029. The contract is worth roughly $45 billion over its full term — the largest single-customer AI infrastructure deal ever disclosed.

The disclosure resolves months of speculation about the terms of the Anthropic-SpaceX arrangement first hinted at in April, and it gives the public its first concrete look at how much frontier AI labs are now spending on raw compute.

What's in the Filing

The S-1 lays out the deal in unusually specific detail. Anthropic will pay $1.25 billion monthly for capacity at the Colossus 1 facility in Memphis, Tennessee. A discounted rate applies during a capacity ramp-up in May and June 2026, with full pricing kicking in from July.

The Memphis facility delivers more than 300 megawatts of compute capacity and houses over 220,000 NVIDIA GPUs — a mix of H100, H200, and next-generation GB200 accelerators. SpaceX is using the SpaceX corporate entity to operate Colossus on behalf of xAI, which is why the filing is under SpaceX's name.

According to the deal terms, Anthropic is ramping to 100% utilization of Colossus 1 within May 2026. The company further expanded the arrangement just as SpaceX filed its IPO paperwork, scaling up on GB200 capacity in the under-construction Colossus 2 cluster throughout June.

Either side may terminate the contract with 90 days' notice — a flexibility clause that gives Anthropic an exit if better compute becomes available and gives SpaceX an exit if it needs the capacity for xAI's own training runs.

Why the Numbers Matter

The deal implies an annualized compute spend of $15 billion just for one customer at one data center. For context, that single contract is bigger than the total annual capex of most Fortune 500 technology companies.

The economics are even more striking on the SpaceX side. The Anthropic deal alone is expected to generate between $3 billion and $4 billion in annual revenue for SpaceX (with the rest of the $15 billion likely flowing as pass-through to xAI and chip costs). According to one analyst note circulating after the filing, the arrangement represents the highest AI data center monetization density globally — roughly $10.9 billion in annualized revenue per 300MW.

Why this matters: This filing is the first hard data point on what frontier-lab compute actually costs at scale. The number is bigger than even pessimistic outside estimates, and it explains why every major AI lab is now in a continuous capital-raising posture.

How Anthropic Plans to Use It

Anthropic compute chief Tom Brown wrote earlier in May that the Colossus compute would be used primarily for inference — the cost of serving model responses to users — rather than training new frontier models.

That breakdown matters. Inference workloads scale linearly with usage, while training is more episodic. Anthropic's product business is now growing fast enough that inference capacity has become the binding constraint, not training.

The company has separately announced massive training-side deals with Google Cloud (for TPU access) and Amazon Web Services (for Trainium-based inference and training), giving Anthropic a three-cloud strategy that no other frontier lab has matched at this scale.

The xAI Connection

The deal puts Anthropic in the unusual position of paying a competitor's parent company for infrastructure. Colossus was originally built for xAI's Grok training runs, and the cluster came online in late 2024 with one of the fastest GPU deployments in the industry.

By renting almost all of Colossus 1's capacity to Anthropic, xAI is effectively monetizing its data center while shifting its own training to Colossus 2 and other facilities. It's a sharp pivot for a company whose CEO has publicly feuded with OpenAI and pitched xAI as a frontier-research counterweight.

For Elon Musk, the read is simpler: Colossus has become a profit center, and the deal helps justify SpaceX's IPO valuation, expected to land near $2 trillion next month.

Industry Reaction

Compute-economy analysts spent Wednesday afternoon recalibrating their AI infrastructure spend models. Several flagged that the $1.25B/mo figure makes Anthropic's revenue numbers look more like a thin margin than a fat one — the company is on track for $10.9 billion in Q2 revenue, but its compute bill from this deal alone runs $3.75 billion per quarter.

Other observers pointed to the 90-day termination clause as a sign that compute supply may be tighter than assumed. A clause that easy to walk away from is unusual in long-term infrastructure deals.

What's Next

SpaceX is targeting an IPO pricing in late June or early July, with the Anthropic contract front and center in its growth story. Expect the company to add additional named AI customers in its updated S-1 amendments as the roadshow progresses.

For Anthropic, the immediate question is whether the company can grow revenue fast enough to keep this compute bill rational. The Q2 revenue trajectory suggests yes — but management has warned that sustained 2026 profitability is uncertain, in part because of exactly this deal.

For the broader market, the next shoe to drop is similar disclosure from OpenAI's Stargate partners, Meta's Hyperion buildout, and Google's TPU contracts with Anthropic and others. The compute economy just got a lot more transparent.

The Bottom Line

A single AI lab is now spending more on compute every month than most Fortune 500 companies spend on capex in a year. The Anthropic-SpaceX deal isn't just a contract — it's a benchmark for what frontier AI now actually costs, and a stress test for whether the revenue can keep up.