Anthropic just…wait what

TL;DR

Anthropic partnered with SpaceX to lease 220,000 H100s because 80x unexpected growth left them critically compute-starved, while XAI needed Cursor's coding data to compete.

Key Points

• 1.Anthropic experienced 80x annualized revenue and usage growth in Q1 2025, which no one could plan for. Dario admitted they conservatively under-bought compute, expecting only 10x growth, leaving them unable to serve demand.
• 2.Anthropic's compute is split across three incompatible hardware architectures. They run on AWS Trainium, Google TPUs, and Nvidia GPUs — researchers prefer CUDA, so Trainium and TPUs are used mainly for inference to free up Nvidia for training.
• 3.The Anthropic-SpaceX deal gives Anthropic access to over 300 megawatts and 220,000 Nvidia H100s from Colossus 1. That's roughly 70% of Colossus 1's capacity, leaving XAI only ~100 megawatts for Grok inference — revealing how little Grok is actually used.
• 4.SpaceX has already moved training to Colossus 2, which is 5x larger at 1.4 million H100 equivalents and 1.5 gigawatts. This freed Colossus 1 to be leased to Anthropic, since Grok has minimal inference demand.
• 5.For Claude Code users, the 5-hour rate limits were doubled and peak-hour reductions removed. However, users hitting weekly limits (heavy daily users, parallel agents) won't benefit — only burst users hitting the 5-hour cap will see improvement.
• 6.The XAI-Cursor deal is fundamentally about data, not compute. XAI has compute but lacks quality coding interaction data; Cursor holds the most valuable corpus of human-AI coding back-and-forths across all major models, which is the exact data needed for RL training.
• 7.The $10B fee in the Cursor deal is effectively the price XAI is paying just for Cursor's data. The acquisition option at $60B covers the full company; if SpaceX declines acquisition, they've paid $10B purely for training data rights.
• 8.OpenAI is uniquely positioned with all three critical resources — research, data, and compute — while every other lab has gaps. Anthropic's biggest competitive moat (AWS Bedrock exclusivity) is now dead as OpenAI gains AWS support, and Codex is rapidly closing the coding capability gap.