Rotating Data
Explore the map
Home/Deal Room/Anthropic commits $1.25B/month to SpaceX Colossus compute

Cloud capacity · Evidence brief

Anthropic commits $1.25B/month to SpaceX Colossus compute

Anthropic + SpaceX

Near-primary sourceSEC summary2026-05-20
Deal structureAI compute capacity agreement / GPU supercomputer cloud services
GeographyMemphis, Tennessee / SpaceX Colossus and Colossus II AI data-center clusters
Infrastructure layersCloud capacity + chips / compute + data centers
Disclosed scale$1.25B per month through May 2029, subject to ramp-up discounts and 90-day termination rights; Colossus 1 adds more than 300 MW and over 220,000 NVIDIA GPUs, with Colossus 2 capacity also being added

Why it matters

The strategic read

Turns frontier-model compute scarcity into one of the largest disclosed AI infrastructure service contracts on record. Anthropic is renting massive NVIDIA GPU capacity from a direct model rival, while SpaceX monetizes Colossus as AI compute-as-a-service alongside its own Grok/xAI workloads.

Source-supported terms

What was disclosed

  1. 01

    Anthropic announced on May 6 that it signed an agreement with SpaceX to use all compute capacity at Colossus 1.

  2. 02

    Anthropic said Colossus 1 gives it access to more than 300 MW of new capacity and over 220,000 NVIDIA GPUs within the month.

  3. 03

    Reuters reported from SpaceX’s IPO filing that Anthropic agreed to pay SpaceX $1.25B per month through May 2029.

  4. 04

    The compute-capacity deals now include both SpaceX AI training data-center clusters, Colossus and Colossus II.

  5. 05

    The filing says May and June 2026 fees are reduced during the capacity ramp and either party may terminate with 90 days notice.

  6. 06

    Axios separately reported Anthropic is expanding beyond Colossus 1 to Colossus 2 GB200 capacity during June.

Disclosure boundary

What is known—and what is not

Anthropic primary announcement plus near-primary reporting from SpaceX IPO filing; full cloud services agreements, exact committed capacity by cluster, SLA terms, GPU mix by phase, networking topology, power/cooling details and termination economics are not public.