Anthropic Secures All of Colossus 1 — Rate Limits Doubled Same Day
Anthropic just solved its most visible compute problem — not with a model update, but a supply deal. The company has secured access to the entire Colossus 1 supercomputer in Memphis, Tennessee: more than 220,000 NVIDIA GPUs and over 300 megawatts of capacity operated by SpaceX. The same day the deal was announced, Claude's 5-hour rate limits doubled across Pro, Max, and Team plans, peak-hour throttling was eliminated, and API tier limits rose 10–16× across all four tiers.
What the Source Actually Says
BridgeMind ran an immediate empirical stress test on the $200 Max plan: six Claude Opus 4.7 instances spawning roughly 50 total sub-agents in parallel — SEO audits, security audits, UI revamps, and mobile app builds running simultaneously. Result: 38% of the 5-hour window consumed for a workload that previously would have exhausted 90–100% of it. The practical shift is clear: the 5-hour limit is no longer the binding constraint. The weekly limit — which was not raised in this announcement — is now the ceiling to plan around.
Dario Amodei explained the prior throttling on stage at the Anthropic Developer Conference: the company planned for 10× Q1 2026 demand growth and encountered 80×. "That is the reason we have had difficulties with compute," he said directly. The SpaceX deal secures "all of the compute capacity at their Colossus One data center." Anthropic's announcement also notes a 5-gigawatt agreement with Amazon, a 5-gigawatt agreement with Google, and stated interest in partnering with SpaceX on "multiple gigawatts of orbital AI compute capacity" — the first formal acknowledgment of space-based AI infrastructure in a major lab's public roadmap.
AlphaSignal's newsletter (163K likes) corroborated the compute figures and flagged that Anthropic shipped four Managed Agents upgrades the same day: multi-agent orchestration, outcome graders (+10pp success on hard tasks), dreaming (background memory consolidation between runs), and webhooks.
Strategic Take
The rate-limit doubling is immediately actionable: multi-agent harnesses that were economically impractical on consumer plans are now viable workhorses. Audit your weekly caps before architecting new fan-out systems — the 5-hour window is no longer the wall. The orbital compute roadmap is a longer-term signal that SpaceX has formally entered the AI infrastructure business.
