SpaceX leased Colossus 1 to Anthropic after failing to optimize it for Grok
▼ Summary
– SpaceX rented Colossus 1 to Anthropic because latency and chip mismatch issues prevented it from effectively using the facility for its own Grok model training.
– Colossus 1 contains a mix of older Nvidia chip generations, which created bottlenecks when paired with the newer, uniform Blackwell chips in Colossus 2 and 3.
– Anthropic pays $1.25 billion per month for Colossus 1, contributing to a $2.17 billion monthly compute revenue from infrastructure SpaceX originally built for itself.
– The facility’s quick 122-day construction led to a lack of uniformity, undermining SpaceX’s IPO narrative about its efficiency.
– Grok’s declining downloads and low paid conversion make reclaiming the compute less urgent, as rental income has become a major revenue source.
SpaceX has leased its Colossus 1 data centre to Anthropic not because of excess capacity, but because the facility proved incompatible with its own AI training needs. According to a Friday report from Bloomberg, the company ran into significant latency issues when attempting to link the Memphis site with two other data centre campuses located more than 10 miles away. Aging network infrastructure only deepened the problem.
The original plan was to train its most advanced Grok models across a cluster of three interconnected facilities. Training large-scale AI systems demands ultra-fast connections between sites. When those links are older or lower bandwidth, they introduce delays that drag down the entire cluster. Faced with this bottleneck, SpaceX decided the facility would generate more revenue as a rented asset than sitting idle.
A hardware mismatch compounded the difficulty. Colossus 1 contains a mixed inventory of Nvidia chip generations, including Hopper and Blackwell systems alongside older accelerators. In contrast, Colossus 2 and 3 were built more uniformly around Nvidia’s Blackwell chips. In a distributed training setup, workloads are spread across machines that must stay tightly synchronized. Older chips create bottlenecks by forcing faster accelerators to wait, meaning the entire cluster performs closer to its slowest hardware rather than its fastest.
The outcome is striking: Anthropic is now paying $1.25 billion per month to use a facility that SpaceX’s own engineers could not fully utilize. Combined with the $920 million monthly Google deal, SpaceX is pulling in roughly $2.17 billion per month in compute revenue from infrastructure originally built for internal use.
This development complicates the narrative SpaceX presented during its IPO roadshow. Musk’s company repeatedly emphasized that Colossus 1 was built in just 122 days, outpacing industry averages. Construction speed was a key selling point. Bloomberg’s reporting suggests that speed came with trade-offs: the facility was not built uniformly enough to function as part of a larger training cluster.
SpaceX CFO Bret Johnsen has stated the company has not abandoned internal AI services, including Grok. Musk has described the Anthropic lease as a 180-day arrangement with a 90-day mutual cancellation right, preserving the option to reclaim the capacity. “If compute gets super tight I said we might need it back at some point,” he said.
Still, Grok’s declining traction makes reclaiming that compute less urgent. Downloads fell from 20 million in January to 8.3 million in April. Paid conversion rates are a fifth of ChatGPT’s, and federal adoption has stalled. The product that was supposed to justify the data centre investment is underperforming, while rental income from Anthropic and Google has become a $26 billion annualized revenue line. SpaceX built a data centre for AI training and inadvertently became an AI landlord instead.
(Source: The Next Web)
