SpaceX rents Colossus 1 to Anthropic after latency and hardware mismatch derail Grok plans

Technical hurdles

SpaceX had planned to train its most cutting-edge Grok AI models using a cluster of three data centre campuses. The firm encountered latency issues when connecting its Colossus 1 facility in Memphis with two other sites located more than 10 miles away, people familiar with the matter told Bloomberg. Aging network infrastructure compounded the problem. Training large AI models requires ultra-fast inter-site connections, and the older, lower-bandwidth links created delays that slowed the entire cluster. SpaceX determined the facility would generate more value as a rental property than sitting underutilised.

Hardware mismatch

The chip configuration inside Colossus 1 made the situation worse. The data centre houses a mix of Nvidia chip generations, including Hopper and Blackwell systems alongside older accelerators. Colossus 2 and 3 were built more uniformly around Nvidia's Blackwell chips. In a distributed training cluster, workloads must stay synchronised, so older processors create bottlenecks by forcing faster accelerators to wait. The cluster performs closer to its slowest component.

Commercial windfall

Anthropic is now paying $1.25 billion per month to use the facility that SpaceX’s own engineers could not fully utilise. Combined with a separate $920 million monthly Google deal for compute capacity, SpaceX collects approximately $2.17 billion per month from infrastructure originally built for internal use. Musk has described the Anthropic arrangement as a 180-day lease with a 90-day mutual cancellation right, preserving the option to reclaim capacity.

“If compute gets super tight I said we might need it back at some point.”

Anthropic1,250

Google920

Anthropic

1250 million USD/month

Google

920 million USD/month

Data from the article text

IPO narrative challenged

The revelation complicates the story SpaceX presented during its IPO roadshow. The company repeatedly spotlighted that Colossus 1 was constructed in just 122 days, beating industry averages. Speed was a selling point. Bloomberg’s reporting suggests that speed came at a cost: the facility was not built uniformly enough to serve as part of a larger training cluster.

Grok’s future

SpaceX CFO Bret Johnsen said the company has not given up on internal AI services, including Grok. The door remains open for SpaceX to reclaim the capacity if needed, though Grok’s trajectory makes that uncertain. SpaceX continues to plan satellite-based AI servers, separate from the ground-based cluster issues.