SpaceX leases Colossus 1 data center to Anthropic after GPU latency issues
In brief
- SpaceX leased Colossus 1's full capacity to Anthropic, providing 220,000+ Nvidia processors.
- Latency issues across geographically dispersed sites prevented integration with Grok.
- The deal creates a new revenue stream as SpaceX expands AI infrastructure.
- SpaceX retains the right to reclaim compute from Anthropic if demand tightens.
The Infrastructure Challenge
SpaceX faced latency issues when trying to connect the Memphis site with two other locations more than 10 miles away. Aging network infrastructure and hardware differences across the facilities worsened latency delays. This matters because large AI training clusters need thousands of chips to stay synchronized, and slower connections or older hardware can force faster systems to wait, reducing overall cluster performance.
Colossus 1 contains a mix of Nvidia chip generations, including Hopper and Blackwell systems, along with older accelerators, unlike Colossus 2 and Colossus 3, which were built more uniformly around Nvidia's Blackwell chips. The heterogeneous architecture and geographic spread made it impractical to keep the Memphis facility tightly coupled to Grok's training pipeline.
A New Revenue Model
Rather than continue wrestling with integration, SpaceX decided Colossus 1 would be more valuable as a leased compute asset for outside customers. Anthropic became the tenant, securing a massive pool of GPU capacity at a critical moment when AI compute remains scarce and expensive.
The deal gives SpaceX a major new revenue stream as it pitches investors on its shift into AI infrastructure. The company has also signed a computing deal with Google, further cementing its narrative around AI infrastructure as a core business line. SpaceX CFO Bret Johnsen said the company is still developing internal AI services, including Grok, signaling that the Anthropic lease doesn't mark an exit from AI development.
Flexibility remains built into the arrangement. Musk has said SpaceX could reclaim compute from Anthropic if internal demand becomes tight, after giving advance notice. The lease structure reflects both parties' recognition that GPU scarcity may tighten further, and that strategic compute assets could shift priority depending on each company's near-term training roadmap.
