← Back to feed

Anthropic has signed an agreement with SpaceX to access all compute capacity at Colossus 1, the AI supercomputing facility SpaceX built in Memphis in 90 days in mid-2024. The deal adds more than 300 megawatts of new capacity — over 220,000 NVIDIA GPUs — to Anthropic’s infrastructure stack, with availability coming within the month.

Three rate limit changes took effect today:

• Claude Code five-hour limits doubled across Pro, Max, Team, and Enterprise seat plans
• Peak hours throttling removed for Pro and Max accounts
• Claude Opus API rate limits raised across all tiers, effective immediately

The SpaceX agreement is structurally different from Anthropic’s hyperscaler deals. Colossus 1 is a physical third-party data center, not a cloud product. SpaceX built it in Memphis specifically for xAI’s training needs. Anthropic is now co-accessing the same cluster. xAI has not commented on the arrangement.

The compute stack

The Colossus 1 deal joins four other active infrastructure commitments:

• Amazon: up to 5 GW total; nearly 1 GW of new capacity by end of 2026
• Google and Broadcom: 5 GW agreement; coming online 2027
• Microsoft and NVIDIA: $30 billion in Azure capacity
• Fluidstack: $50 billion committed to American AI infrastructure

Anthropic trains and runs Claude across AWS Trainium, Google TPUs, and NVIDIA GPUs simultaneously. The three-hardware strategy is unusual at this scale. Adding SpaceX’s NVIDIA-native cluster expands the NVIDIA-side capacity without adding a new hardware family to manage.

The orbital signal

The announcement closes with a sentence worth isolating: Anthropic has “expressed interest” in partnering with SpaceX to develop “multiple gigawatts of orbital AI compute capacity.”

This is interest, not a signed deal. Satellite-based AI inference at gigawatt scale is not a commercially deployed technology. But SpaceX is the only organisation with the launch cadence, mass-to-orbit economics, and constellation operations experience to make it an engineering program rather than a research concept. No other major AI lab has named orbital compute as a near-term infrastructure direction.

The context is terrestrial scarcity. The compute required to train and run frontier models is outpacing what land, power, and cooling can deliver on relevant timelines. Anthropic’s run-rate revenue is estimated at approximately $40 billion annually as of late April 2026. Supply is the binding constraint, not demand.

Whether orbital compute progresses from interest to agreement will depend on factors Anthropic does not control — Starship launch cadence, power-in-orbit economics, latency tolerances for inference workloads. The announcement plants a flag. The engineering problem is genuinely unsolved.