A deep dive into the Anthropic & xAI agreement

Highlights

• Anthropic is renting the entire 300-megawatt Colossus 1 data center from SpaceX to increase Claude usage limits for developers.

• The Colossus 1 deal provides immediate access to over 200,000 NVIDIA GPUs to alleviate Anthropic's severe compute constraints.

• Demand for AI tokens is scaling exponentially rather than linearly due to the rise of longer-running agentic coding and data science sessions.

• Anthropic has secured future agreements for 10 gigawatts of power split between Amazon and Google, with significant capacity arriving in 2026 and 2027.

• SpaceX is planning to launch orbital data centers as part of its strategy to offer multiple gigawatts of space-based AI compute capacity.

• xAI is shifting its internal training and inference to the more powerful Colossus 2 cluster and the Macro Hardware data center.

Timeline

1. 00:00 - 01:10

   Immediate Expansion of Claude Usage Limits

   Developers frequently hit restrictive usage limits when using agentic tools compared to competitors like OpenAI's Codex. The acquisition of the entire Colossus 1 capacity provides the necessary infrastructure for inference and training to relax these restrictions. This move directly addresses the immediate bottleneck in serving high-demand developer tools.

   • Claude users will receive higher usage limits to accommodate the heavy token consumption of Claude Code.
   • Anthropic is renting the full 300-megawatt capacity of the Colossus 1 data center to support these limits.
   • The Colossus 1 facility was built by Elon Musk in a record 122 days during 2024.
2. 01:11 - 02:37

   The Shift Toward Exponential Token Consumption

   Current AI models are no longer used for simple one-off queries but for long-running sessions where the agent performs complex tasks. These extended sessions generate massive amounts of code and data, consuming tokens at a rate that exceeds original capacity planning. This shift in user behavior necessitates a rapid and massive expansion of hardware resources.

   • AI companies are struggling with token demand that is growing exponentially rather than linearly.
   • Modern agentic harnesses for coding and data science run for longer durations and burn significantly more tokens.
   • Compute constraints remain a primary challenge for all major AI labs despite massive infrastructure investments.
3. 02:38 - 03:51

   Long-term Infrastructure Roadmap and Partnerships

   While the 300 megawatts from SpaceX is available now, the vast majority of Anthropic's planned capacity is not yet online. The total 10 gigawatts pledged by Amazon and Google represent future-dated infrastructure. This highlights the gap between current compute needs and the multi-year lead times required to build massive data centers.

   • The SpaceX deal adds 200,000 NVIDIA GPUs to Anthropic's current operational pool.
   • Anthropic has committed to 5 gigawatts of capacity with Amazon, with 1 gigawatt expected by the end of 2026.
   • An additional 5 gigawatts of compute from Google is scheduled to begin coming online in 2027.
4. 03:52 - 06:13

   Strategic Logic for xAI and SpaceX

   Leasing underutilized capacity to a competitor allows xAI to generate revenue while focusing its internal efforts on newer, more powerful clusters. This arrangement suggests that Anthropic's user demand currently outweighs that of Grok. The revenue generated helps offset the massive capital expenditures required for xAI's rapid infrastructure build-out.

   • xAI is leasing out Colossus 1 because its own demand for Grok does not currently require the full capacity.
   • The more powerful Colossus 2 cluster and the Macro Hardware data center now serve xAI's primary needs.
   • xAI is on a trajectory to reach nearly 2 gigawatts of operational power by 2026.
5. 06:14 - 12:18

   Market Competition and Orbital AI Horizons

   The transition from subscription-based to usage-based billing in the industry, led by GitHub Copilot, forces Anthropic to grow its market share aggressively. The potential move to space-based data centers aims to solve terrestrial energy and cooling constraints, though the technology remains unproven for large-scale AI. This partnership provides SpaceX with a high-profile potential customer to show investors ahead of its IPO.

   • Anthropic is under pressure to match the generous usage limits and subscription models offered by GitHub Copilot and OpenAI.
   • SpaceX and Anthropic have expressed interest in developing multiple gigawatts of orbital AI compute capacity.
   • The mention of orbital compute serves as a strategic narrative for SpaceX's upcoming initial public offering (IPO).