Core of the Agreement
On May 6, Anthropic and SpaceX signed a compute agreement, granting Claude access to the Colossus 1 supercomputer.
These numbers speak to the sheer scale:
- 220,000+ NVIDIA GPUs
- 300MW of power consumption
- One of the world's largest AI training facilities
Anthropic has expressed interest in collaborating at the "multi-gigawatt" level. What does this mean? It means that training for Claude's next-generation models could run directly on SpaceX's infrastructure.
Why SpaceX?
Musk's AI Empire Puzzle
This deal cannot be viewed in isolation. Around the same time, Elon Musk confirmed that xAI will be dissolved as an independent entity and fully integrated into SpaceX to form SpaceXAI.
So the deal chain looks like this:
- xAI merges into SpaceX → SpaceXAI is established
- SpaceXAI owns the Colossus supercomputer
- Anthropic signs an agreement to access Colossus
- Claude's training capabilities receive a massive boost
Musk is effectively controlling nearly every layer of the AI industry chain: data (X platform), compute (Colossus), infrastructure (SpaceX + Starlink), physical AI (Tesla Optimus), models (Grok), and distribution (X + Starlink). And Anthropic has chosen to become a "tenant" in this ecosystem.
Anthropic's Compute Anxiety
Within a single week in early May, Anthropic rapidly secured multiple major deals:
- Google Cloud: $200 billion contract
- SpaceX: Colossus access agreement
- Additionally, rate limits for Claude Code doubled across all paid plans
This highlights a core issue: Anthropic's compute demands have surpassed its ability to build out capacity in-house. Even with a $200 billion contract with Google Cloud, there's still a need for additional compute sources. Colossus perfectly fills this gap.
Industry Signals
The Embodiment of "Compute is Power"
What does 220,000 GPUs mean in context? When OpenAI originally trained GPT-4, it used roughly 25,000 A100 GPUs. The scale of Colossus 1 is nearly 10 times that number.
When a single company (or ecosystem) controls the largest compute capacity in the industry, it effectively holds the "faucet" for AI development. Whoever gets the compute gains the ability to train larger models—this represents a new industry power structure.
The Shift from "In-House" to "Shared"
Historically, AI companies' strategies revolved around building their own compute clusters (Google TPUs, AWS Trainium, OpenAI's self-procured GPUs). The Anthropic-SpaceX agreement marks a new model: compute sharing.
SpaceXAI provides the compute, while Anthropic brings model capabilities—each getting what they need. If this model holds, it could become the standard collaborative framework for the future AI industry.
Contrast with Pentagon Contracts
Ironically, on the day after Anthropic signed the Colossus agreement (May 5), the Pentagon released its list of AI collaboration partners—and Anthropic was left out.
On one hand, it secures access to the industry's largest compute capacity; on the other, it's shut out of the world's largest government procurement market. Anthropic's situation is full of contradictions: surging ahead commercially, but struggling in the government sector.
What's Next
SpaceX has also submitted plans to build a $55 billion Terafab chip fabrication plant in Texas. If completed, SpaceXAI will upgrade from a "compute user" to a "compute producer"—eliminating the need to even purchase GPUs externally.
By then, the scale of Colossus could multiply several times over. Anthropic, being on this ship, will only find its position becoming increasingly crucial—and increasingly dependent.