Anthropic token use quadrupled Jan–Mar, triggering compute crunch and token throttling
- Anthropic raised Claude usage limits on May 6 after signing a new compute deal with SpaceX, reversing the peak-hour throttling that had frustrated heavy users. - The new capacity is big: more than 300 megawatts at SpaceX’s Colossus 1, or over 220,000 Nvidia GPUs, arriving within a month. - The bigger story is infrastructure lock-in — Anthropic is stacking Amazon, Google, Microsoft, Nvidia, and now SpaceX to keep serving demand.
AI demand is starting to look less like a software story and more like a power-and-chips story. That’s the real meaning of the Claude throttling drama from late March through early May. Heavy Anthropic users were suddenly chewing through five-hour limits in minutes, peak-hour access got tighter, and developers started asking the obvious question — is the model getting worse, or is the company running out of room? On May 6, Anthropic answered by doing two things at once: it raised Claude limits and announced a new compute partnership with SpaceX. (tech.yahoo.com) ### What actually changed at Anthropic? Anthropic said it is doubling Claude Code’s five-hour rate limits for Pro, Max, Team, and seat-based Enterprise plans, removing the peak-hours limit reduction for Pro and Max, and raising API rate limits for Claude Opus models. Those changes took effect May 6, 2026. Anthropic tied the move directly to new compute coming online, especially a SpaceX deal it says will substantially expand near-term capacity. (anthropic.com) ### Why were users so angry before that? Because the squeeze was visible in the product. In late March, heavy Claude users posted screenshots showing five-hour quotas disappearing in about 20 minutes. Anthropic also told subscribers that usage would burn down faster during peak periods, and some third-party tools lost access to flat-rate subscription limits. That felt less like (anthropic.com)he UI. (tech.yahoo.com) ### Why does a “token crunch” happen? Because inference is expensive now, not just training. Every long prompt, every tool call, every coding loop, every big context window burns compute after the model is already deployed. And Claude is increasingly aimed at exactly those workloads — coding, agents, computer use, long-running tasks. So if usage sp(tech.yahoo.com)lly what users were feeling. Anthropic hasn’t publicly confirmed a “tokens quadrupled Jan–Mar” figure in the material I could verify, but it has repeatedly described demand as extraordinary and exponential. (anthropic.com) ### What did SpaceX add? A lot, at least on paper. Anthropic said the SpaceX agreement gives it access to all compute capacity at the Colossus 1 data center — more than 300 megawatts and over 220,000 Nvidia GPUs within the month. Anthropic says that extra capacity will directly improve service for Claude Pro and Claude Max subscribers. That is a very concrete answer to a very concrete bottleneck. (anthropic.com) ### Is this just a SpaceX story? No — that’s the catch. SpaceX is the latest patch in a much bigger infrastructure land grab. On April 20, Anthropic said it had expanded its Amazon deal to secure up to 5 gigawatts of capacity and committed more than $100 billion over 10 years to AWS technologies. On April 6, it announced another multi-gigawatt compute expansion with Google and B(anthropic.com)ership with Microsoft and Nvidia that includes $30 billion of Azure capacity. (anthropic.com) ### Why does cloud positioning matter so much now? Because frontier-model competition is drifting downward into infrastructure. Anthropic keeps emphasizing that Claude runs across AWS Trainium, Google TPUs, and Nvidia GPUs, and that it is available through AWS, Google Cloud, and Microsoft Azure. That mix gives it leverage, redundancy, and bargaining power. But it also tells you (anthropic.com)chips, electricity, data centers, and cloud contracts. (anthropic.com) ### Is OpenAI dealing with the same pressure? Broadly, yes, though the symptoms differ. The wider market is pouring money into AI infrastructure because demand is outrunning supply. Microsoft said on April 29 that its AI business had passed a $37 billion annual revenue run rate, up 123% year over year, while cloud obligations also surged. That kind of growth is great for busines(anthropic.com)availability. (news.microsoft.com) ### Bottom line? The Claude throttling episode mattered because it exposed the hidden constraint. The winning AI companies may not just be the ones with the best models. They may be the ones that lock up enough compute to keep those models reliably on. (anthropic.com)
