Anthropic Lands Colossus 1, Ships Managed Agents in One Day
Anthropic's most consequential 24 hours of 2026: a compute deal leasing all of SpaceX/xAI's Colossus 1 supercomputer — 300+ megawatts, ~220,000 Nvidia GPUs — immediately doubled Claude Code rate limits and unlocked three platform launches at the Code with Claude conference. Three independent source batches confirmed the story within hours of the announcement.
What the Source Actually Says
The Colossus 1 deal took effect the same day it was announced because the hardware was already running. xAI had shifted its Grok training to the newer Colossus 2 facility, freeing the original cluster to rent. The immediate rate-limit changes are substantial: Claude Code 5-hour limits doubled across Pro, Max, and Team plans; peak-hour throttling on Pro/Max was removed entirely; and Opus API throughput jumped 10–16× (tier-1 input: 30K → 500K tokens/min; tier-1 output: 8K → 80K/min). The deal adds to Anthropic's existing compute agreements — AWS 5GW, Google/Broadcom 5GW coming in 2027, and a Microsoft/Nvidia $30B Azure arrangement — all catch-up moves following a conservatively managed capex strategy that left Anthropic compute-constrained throughout late 2025. The partnership also carries a notable subtext: Elon Musk, who had previously called Anthropic "misanthropic," accused it of "stealing training data at massive scale," and retweeted anti-Anthropic content as recently as March 2026, offered a pointedly restrained endorsement — noting "no one set off my evil detector" after xAI had already vacated Colossus 1 for Colossus 2.
On the product side, Anthropic shipped three Managed Agents features at Code with Claude. Managed Agents (public beta) enables a lead agent to delegate tasks to parallel specialist sub-agents for complex jobs. Dreaming (research preview) consolidates agent memory asynchronously between sessions, surfacing recurring mistakes and shared patterns — building on Letta's prior sleep-time compute research. Outcomes closes the quality loop with a rubric-plus-grader architecture: a separate grader evaluates output in its own context window and instructs the agent to re-run until criteria are met; Anthropic reports +10 points on task success for the hardest problems, +8.4% on DocX, and +10.1% on PowerPoint in internal testing.
Strategic Take
The Colossus 1 deal relieves a compute constraint that directly caused every rate-limit reduction of the past six months — a single root cause resolved overnight. For builders: multi-agent workflows and the 1M context window are now production-viable. Dreaming's manual trigger is a cost-control concession today; watch for it to become the default as capacity scales.
