
Tech • IA • Crypto
Anthropic has secured access to SpaceX’s Colossus 1 supercomputer, adding roughly 220,000 GPUs and reshaping AI infrastructure alliances.
Elon Musk has agreed to lease the full capacity of Colossus 1, a large-scale AI supercomputing cluster, to Anthropic. The system reportedly includes around 220,000 GPUs, dramatically increasing the compute available to run and scale Claude models. Deployment is expected within weeks, making the expansion unusually rapid for infrastructure of this size.
The additional compute is likely to reduce usage constraints and latency for users of Claude, while enabling broader global deployment. Increased capacity also supports more advanced model training and inference, potentially accelerating feature rollouts and model upgrades.
The partnership marks a notable shift in tone from Musk, who has previously criticized rivals in the AI sector. He now frames Anthropic as aligned with safer AI development practices, suggesting a pragmatic pivot toward collaboration despite competitive tensions across the industry.
The deal reinforces Anthropic’s approach of distributing dependencies across major providers, including Amazon, Google, and Microsoft. Adding SpaceX infrastructure further diversifies its compute base, reducing reliance on any single partner while increasing operational complexity.
The agreement hints at SpaceX’s expanding role in AI infrastructure, including potential future orbital data centers. This could position the company as a unique provider of off-Earth computing capacity, a concept gaining attention for energy and cooling advantages.
The move may signal a shift away from building a dominant consumer AI platform internally, toward focusing on infrastructure and specialized AI applications such as robotics, including the Optimus program. Partnering with a leading AI developer could also strengthen SpaceX’s valuation ahead of future funding rounds.
The unexpected alliance between Anthropic and SpaceX underscores a rapidly evolving AI landscape where access to compute is becoming as निर्णutive as model innovation itself.