On May 6, 2026, xAI (merged into SpaceX and renamed SpaceXAI) announced a compute agreement with Anthropic: Anthropic will use the entire capacity of Colossus 1. The Memphis, Tennessee supercomputer is SpaceX’s largest facility to date, with more than 220,000 NVIDIA GPUs (including H100, H200, and next-generation GB200 accelerators); CNBC reports over 300 megawatts, available for training, fine-tuning, inference, and high-performance computing.
The most theatrical part is the relationship between the two parties. Elon Musk had publicly attacked Anthropic, then reversed course after meeting its senior team, posting that he was “impressed” and “No one set off my evil detector.” The result is a scene rare in AI history: one frontier lab renting a rival’s landmark data center, whole.
Why Anthropic Had No Choice
The answer sits in its own numbers. In its April announcement, Anthropic disclosed annualized revenue above $30 billion — up from roughly $9 billion at the end of 2025. By April, Claude demand had already put “inevitable strain on our infrastructure,” hurting “reliability and performance” at peak hours. The official framing of the deal is blunt: the added compute will “directly improve capacity for Claude Pro and Claude Max subscribers.”
In other words, this is not about hoarding training compute for a next-generation model — it is about fixing a revenue-generating problem where paying users can’t get through.
Multi-Supplier: Never One Basket
Zoom out and Colossus is the newest tile in a larger mosaic:
- April 2026: a multi-year Google and Broadcom deal for multiple gigawatts of TPU capacity, coming online from 2027
- Amazon’s multibillion-dollar investment agreement, reported at up to $25 billion
- A $200 billion-class commitment to Google Cloud cited in CNBC’s video coverage
- The entire capacity of Colossus 1 (May 2026)
Add the reports that Anthropic is negotiating a raise at roughly a $900 billion valuation, and the picture is clear: Anthropic treats compute acquisition as a battlefield separate from the model race, deliberately split across three supply lines — NVIDIA GPUs, Google TPUs, and third-party data centers (SpaceXAI).
Orbital Compute: From Sci-Fi to Engineering Program
The deal also hides a longer-range signal: Anthropic expressed interest in co-developing “multiple gigawatts of orbital AI compute capacity” with SpaceXAI. SpaceXAI’s argument is that compute demand “is outpacing what terrestrial power, land, and cooling can deliver on the timelines that matter,” and that its launch cadence and mass-to-orbit economics can “make orbital compute a near-term engineering program rather than a research concept.”
When that becomes real is anyone’s guess. But the premise — power and land are the hard limits on AI expansion — is enough for every lab to take the space option seriously.
What It Means for Developers and Product Teams
Three practical effects. First, expect Claude Pro and Max capacity limits to keep loosening; the paid tier is the direct beneficiary of this expansion. Second, compute has become a tradable commodity: when a Colossus can be rented in full, frontier infrastructure idle capacity is entering the market, and startups have more paths to serious compute. Third, multi-sourcing is now consensus — an architecture dependent on a single chip vendor or a single cloud is an explicit operational risk in 2026.
Sources
- New Compute Partnership with Anthropic — xAI
- Anthropic, SpaceX announce compute deal — CNBC
- Anthropic expands partnership with Google and Broadcom for compute
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.