On May 6, Anthropic announced something that would have seemed surreal six months ago: it signed an agreement to consume all available computing capacity at SpaceX’s Colossus 1 data center in Memphis — the same infrastructure that Elon Musk, one of Anthropic’s loudest critics, built to power xAI’s Grok.
It’s not personal in the GPU wars. It’s just business.
What the agreement includes
Colossus 1 gives Anthropic access to over 300 megawatts of capacity — more than 220,000 NVIDIA GPUs across H100, H200, and GB200 accelerators. Full capacity will be available within a month, and Anthropic confirmed that it will directly improve capacity for Claude Pro and Claude Max subscribers.
As part of the agreement, Anthropic also “expressed interest” in partnering with SpaceX to develop multiple gigawatts of orbital computing capacity. That detail is significant — satellite-based inference infrastructure remains largely theoretical, but SpaceX submitted a proposal to the FCC in January for deploying a constellation of one million satellites as an orbital data center.
Why Colossus 1 was available
Context matters. xAI already migrated its training workloads to Colossus 2, which went operational in January 2026. Colossus 1, while still massive, was partially underutilized. The facility was also the center of controversy — civil rights groups in Memphis protested against the use of natural gas turbines to power the data center, arguing they were operating without proper federal permits and worsening local air quality.
For xAI, leasing capacity to a direct competitor is pure pragmatism: better to monetize underutilized infrastructure than leave it idle. For Anthropic, it’s GPU-hour arbitrage — access to the fastest-deployed AI supercomputer in history, without waiting years for proprietary infrastructure.
The limit changes affecting you today
Anthropic was direct about the impact for users. Effective May 6:
Claude Code users:
- The five-hour rate limit doubles for Pro, Max, Team, and Enterprise plans
- Peak-hour limit reductions are eliminated for Pro and Max accounts
API users:
- Rate limits for Claude Opus models increase substantially (see updated table in Anthropic’s official announcement)
If you’ve been hitting limits in the middle of a session — especially during peak hours — these changes are material. The elimination of peak-hour penalties is especially relevant for LatAm developers, where those “peak hours” used to coincide with the most productive part of our workday.
The number that explains it all: 80x growth
Anthropic cited 80x growth in usage during Q1 2026 as the driver of this entire compute move. It’s not a typo. A demand curve like that leaves no room for incremental planning — you have to move fast, even if it means leasing infrastructure from Elon Musk.
To put their current position in perspective: Anthropic has compute agreements with Amazon (up to 5 GW, nearly 1 GW by end of 2026), Google and Broadcom (5 GW, available from 2027), Microsoft and NVIDIA ($30 billion in Azure capacity), and Fluidstack ($50 billion in infrastructure) — with SpaceX now added as the deal that closes the short-term capacity gap.
What this means strategically
From a CIO perspective, this is a company genuinely constrained by its own growth on the supply side. The SpaceX agreement is not a long-term infrastructure bet — it’s a bridge. The real bet is on Amazon, Google, and orbital compute. Colossus 1 buys Anthropic the runway needed to reach that point without degrading the experience of its subscribers during what is clearly an inflection point in demand.
For those evaluating AI platforms in enterprise contexts: a vendor that proactively secures computing capacity and publicly announces specific limit increases tied to real infrastructure is being transparent. Compare that to platforms where rate limits are opaque and throttling happens without warning.
The geopolitical footnote is also worth following. Anthropic was explicit that it will expand capacity internationally by partnering with democratic countries that have secure AI supply chains — a direct signal that compute sovereignty and data residency requirements are becoming first-class variables in enterprise AI purchasing.
The bottom line
If you use Claude Pro, Max, or the API, your limits just improved in concrete ways. The agreement behind that change is one of the most surprising business stories of the year. And if you’re thinking about AI infrastructure as a long-term strategic variable — for your team, your clients, or your own products — the compute war is moving faster than most enterprise roadmaps account for.
Is your team already feeling the limitations of the AI platforms you’re using? How are you managing usage spikes and rate limit caps in your current projects?