Anthropic leased xAI's entire 220,000-GPU Colossus 1 supercomputer from SpaceX for inference. The mixed-architecture design was too inefficient to train Grok, per Tom's Hardware.
Key facts
- 220,000 GPUs in Colossus 1 cluster.
- Mixed H100/H200/MI300X architecture caused training bottlenecks.
- Colossus 2 will be unified Blackwell-only.
- Anthropic raised $11.5B+ total funding.
- xAI considering IPO with Colossus 2 as asset.
Anthropic has leased xAI's entire 220,000-GPU Colossus 1 supercomputer from SpaceX to run Claude inference workloads, according to Tom's Hardware. The deal addresses Anthropic's growing compute bottlenecks as Claude usage scales—Claude Code alone appeared in 693 prior articles on gentic.news—but the arrangement reveals deeper structural flaws in the cluster's design.
Why Colossus 1 Failed at Training
The mixed-architecture design—combining NVIDIA H100s and H200s with AMD MI300X accelerators—produced inter-GPU communication bottlenecks that made it unsuitable for training large frontier models like Grok. [According to Tom's Hardware], the inefficiency was so severe that xAI couldn't effectively train its own model on the cluster. Instead, the 220,000 GPUs now handle inference workloads for Anthropic's Claude models, including Claude Opus 4.6 and Claude Sonnet 4.6.
Colossus 2: A Clean-Slate Bet
Musk is now preparing Colossus 2, a unified Blackwell-only cluster designed specifically for training frontier models and supporting a potential xAI IPO. [The source reports] that the new cluster will avoid the heterogeneity issues that plagued Colossus 1. The move aligns with xAI's broader infrastructure ambitions, including orbital data centers, as noted in the knowledge graph.
Strategic Implications
For Anthropic, the Colossus 1 deal signals willingness to bypass traditional cloud providers for raw compute capacity. The company, which has raised over $11.5B and is projected to surpass OpenAI in ARR by mid-2026, now has a dedicated inference farm outside AWS/GCP. But the arrangement also ties Anthropic's inference infrastructure to a competitor's hardware—a risk if xAI reprioritizes Colossus 1 for its own needs.
The Colossus 1 cluster is reportedly powered by mobile gas turbines and diesel generators, raising sustainability questions that may conflict with Anthropic's stated safety and ethics commitments.
What to watch
Watch for xAI's Colossus 2 deployment timeline and whether Anthropic extends the Colossus 1 lease beyond inference workloads. Also track any IPO filing from xAI that discloses Colossus 2 as a key asset.
