SpaceX Forced to Rent Colossus 1 from Anthropic: A Flaw Exposed

SpaceX Leases Colossus 1 to Anthropic: An Unexpected Revelation

SpaceX, Elon Musk's company, has recently decided to lease part of its artificial intelligence computing capabilities to Anthropic. This choice, which comes just weeks before SpaceX's IPO, is driven by internal issues rather than a surplus of resources.

The contract with Anthropic, valued at $1.2 billion per month, appears to be a fallback solution for SpaceX. According to recent reports, this lease is related to technical difficulties faced by the company.

Connectivity Issues for Colossus 1

Anthropic is now using SpaceX's supercomputer Colossus 1 to train its artificial intelligence models. However, access to this infrastructure has been made possible due to latency issues that SpaceX has encountered. Indeed, the company has struggled to connect Colossus 1 to two other supercomputers located 16 kilometers away.

These difficulties are primarily due to aging network infrastructure. SpaceX had initially planned to train its Grok models using a cluster combining these three sites. However, bandwidth limitations and delays have slowed down the entire cluster, leading to underutilization of resources.

Colossus 1: A Heterogeneous Infrastructure

Colossus 1 stands out due to its varied composition, integrating NVIDIA GPUs from the Hopper and Blackwell generations, as well as older accelerators. In comparison, Colossus 2 and 3 are fully equipped with Blackwell GPUs, NVIDIA's latest generation.

This diversity of hardware creates bottlenecks in a distributed training cluster. The older chips slow down the entire system by forcing the faster accelerators to wait. As a result, the cluster operates at the pace of its slowest components.

In light of these challenges, SpaceX has found it more advantageous to lease Colossus 1 to Anthropic, as well as to Google, thereby optimizing the use of its infrastructure.