Elon Musk announced on May 28 that SpaceX has agreed to lease its Colossus AI training clusters to Anthropic for just six months. The deal includes a mutual 90-day cancellation notice, meaning either side can walk away with relatively little friction.

The deal behind the deal

The lease is a continuation of a partnership that kicked off on May 6, when Anthropic secured exclusive access to Colossus 1. For those unfamiliar, Colossus 1 is one of the largest AI supercomputers on the planet, housed in Memphis, Tennessee.

The numbers are staggering. Colossus 1 packs over 220,000 NVIDIA GPUs and draws more than 300 MW of power capacity. The facility became operational in 2024 and scaled rapidly to become a leading force in AI computing. Anthropic plans to use the capacity to expand its Claude Pro and Claude Max subscriber limits.

SpaceX’s S-1 filing indicated potential payments of $1.25 billion per month from the arrangement, which would total nearly $15 billion annually through May 2029. But Musk was quick to clarify that figure doesn’t represent a firm long-term commitment from SpaceX.

The decision to keep the lease at 180 days was made at SpaceX’s request. The company wants to maintain flexibility as its own compute demands evolve.

An unlikely partnership

The deal is notable for reasons beyond its financial terms. Musk has previously been publicly critical of Anthropic, the company founded by former OpenAI researchers Dario and Daniela Amodei. His willingness to lease SpaceX’s crown jewel AI infrastructure to a company he’s openly sparred with signals that pragmatism is winning over personal grudges, at least for now.

Anthropic has also expressed interest in future collaborations, including orbital compute partnerships — the possibility of running AI workloads on hardware in space.

What this means for investors

Anthropic benefits in the near term by getting access to one of the most powerful training clusters ever built. More compute means faster training runs, which means faster iteration on Claude’s next generation of models.

But the short lease also creates uncertainty. If SpaceX decides its own needs take priority, Anthropic could find itself scrambling for alternative compute in six months — a real risk when its product roadmap depends on access to a specific cluster of 220,000 GPUs.

For investors in NVIDIA, this deal is another data point confirming demand for their hardware. Over 220,000 GPUs in a single facility, with a customer willing to pay more than a billion dollars a month for access, is a concrete demand signal.

Watch the 90-day mark closely. If neither party triggers the cancellation clause, it suggests the partnership is working well enough to extend. If SpaceX pulls the plug early, it likely means internal compute needs are growing faster than expected.