Anthropic Doubles Claude Limits With Colossus

Anthropic announced on May 6 that it has secured the entire compute capacity of SpaceX's Colossus 1 data center. Located in Memphis, the facility is one of the world's largest artificial intelligence supercomputers, equipped with over 220,000 Nvidia GPUs and more than 300 megawatts of power.

Following the agreement, Anthropic doubled the five-hour usage limits for Claude Pro, Max, Team, and Enterprise, and eliminated the peak-hour throttling previously applied to Pro and Max tiers. The move directly addresses long-standing user complaints that Claude's rate limits and usage caps were excessively restrictive and expensive.

Colossus Cluster Resolves Compute Bottleneck

To trace the source of this compute capacity, one must look to SpaceX's acquisition of xAI. As xAI migrated its training workloads to the newer Colossus 2 facility, the Memphis-based Colossus 1 cluster was freed entirely, allowing Anthropic to step in as the sole tenant.

However, because the cluster features a heterogeneous mix of H100, H200, and Blackwell chips, it is highly inefficient for training large-scale models. Consequently, Anthropic is deploying Colossus 1 specifically for inference—serving real-time responses to Claude users—rather than training. This directly expands the core processing throughput that had driven the previous rate limits and capacity constraints.

The expansion immediately updated the service's limit schedules. In addition to doubling the five-hour caps and removing peak-hour throttling, Anthropic raised API limits for Opus. From May 13 through July 13, Claude Code weekly limits will also receive a 50% increase. The adjustments took effect automatically across all subscription plans upon the official announcement.

March Limit Shortages Lead to April Apology

The aggressive limit expansion reflects the intensity of prior user dissatisfaction. In March, Claude Code users reported that the five-hour usage cap could be exhausted within just a few coding iterations. A single prompt could consume more than half of a session's allotment, and rate limits frequently triggered even when user dashboards indicated low activity.

Throttling was particularly severe during weekday morning peak hours. Forbes reported that across all subscription tiers—from the $20 Pro plan to the $200 Max tier—users complained of paying premium rates for deteriorating service quality. Anthropic formally acknowledged the peak-hour adjustments on March 26.

Remediation efforts began in April. In its April 23 postmortem, Anthropic attributed the quality degradation and rapid limit exhaustion to system experiments and software bugs, subsequently resetting subscription limits as a goodwill gesture. That these challenges stemmed from an inability to meet surging demand was already evident during the widespread service outage earlier this year.

A $1.25 Billion Monthly Lease to Rebuild Trust

Accordingly, this dramatic compute capacity expansion is not a simple gesture of goodwill, but a highly capital-intensive intervention. Industry reports indicate that Anthropic pays SpaceX approximately $1.25 billion per month to lease Colossus 1, a commitment that would exceed $40 billion if maintained through May 2029. However, Elon Musk has described the initial contract as a six-month lease equipped with an option to extend.

Regardless of the lease structure, user sentiment has rebounded sharply. Since the limits doubled, user feedback has shifted positively, with many noting that the tool is finally viable for heavy workflows. Sentiment was further bolstered when Opus 4.8, released later that month, maintained its flat pricing model.

Nevertheless, operational risks persist. On June 1, both Sonnet 4.6 and Opus 4.7 suffered brief, intermittent errors, and the long-term sustainability of such high rental costs remains an open question. While the status page reports all systems operational as of June 2, the industry's focus now turns to how Anthropic will capitalize on the operational runway bought by this rented supercomputer.