▶ YouTube Summaries

← All summaries
Theo#▶ Watch

Anthropic is leasing SpaceX's Colossus 1 supercluster

2026-05-07 Thu ⏱ 35 min @t3dotgg

Anthropic is leasing SpaceX's Colossus 1 supercluster (~300 MW, ~220k H100s) — despite years of mutual loathing with Elon — because they badly underestimated demand. Theo maps the compute/data/research triangle and explains why every recent Anthropic move makes sense through that lens.

Anthropic's compute crisis

  • Dario admitted at Code with Claude that Anthropic planned for 10× yearly growth but Q1 annualized at ~80×. Recent product churn (trying to pull Claude Code from Pro, peak-hour throttling) was about freeing capacity, not extracting more revenue.
  • Their compute is fragmented across Trainium, Google TPUs and Nvidia. Researchers want CUDA, so they push inference onto TPU/Tn to free GPU for training — but it still wasn't enough.

What this means for users

  • Peak-hour limit reductions gone; Claude Code 5h limits doubled. Helps burst users; weekly cap unchanged so heavy daily users see little improvement.
  • API tier limits massively raised, especially input tokens (Tier 1: 30k → much more; Tier 4: 2M → 10M).
  • Theo's conspiracy: Azure-hosted Claude latency curves mirror Anthropic's own hosting too closely — likely passthrough, not real Azure inference. AWS and GCP numbers are independent and faster.

XAI ↔ Cursor deal

  • Three pillars of an AI lab: research, data, compute.
  • XAI: lots of compute (Colossus 1 was sitting idle because nobody uses Grok), no good code data, thin research.
  • Cursor: best code-interaction data anywhere (every model's history through their editor), some research, no compute.
  • Anthropic: world-class research and good data, no compute.
  • OpenAI: has all three and is now coming to AWS, killing Anthropic's enterprise wedge — which is why Anthropic swallowed pride and took Elon's GPUs. SpaceX's $10B / $60B Cursor option is essentially "$10B for the data, or $60B for the company."

The "shift left" data trick

  • Multi-turn corrections in chat sessions become RL training data by literally moving follow-up user messages "to the left" of the assistant turn — which is why one-shot agent behavior keeps improving. This is why Anthropic fights so dirty to keep its model histories out of competitors' hands (banning Windsurf, banning XAI through Cursor, the Chinese-distillation article).