Anthropic walked on stage at its Code with Claude conference in San Francisco on May 6, 2026, and announced something almost nobody saw coming. The company signed a deal to use every single GPU sitting inside SpaceX’s Colossus 1 data center in Memphis. Over 220,000 NVIDIA accelerators. More than 300 megawatts of power. And it all gets pointed at Claude.
The result hit users the same day. Pro and Max plan token limits doubled. The peak-hour throttling that paid users had been quietly hating for months disappeared. The free plan, predictably, got nothing.
What the new limits actually look like
Pro users at $20 a month went from roughly 44,000 tokens per five-hour window to about 88,000. Max 5x customers at $100 a month moved up to around 176,000. The top-end Max 20x plan at $200 a month landed at about 440,000 tokens per five-hour window. Anthropic kept the limits framed as approximate because the numbers shift with model choice and tool usage, but the doubling is consistent across tiers.
The bigger story for heavy users is the throttling change. Before this week, Pro and Max accounts hit a “soft cap” between 11 a.m. and 4 p.m. Pacific time. Claude would still answer requests, just slower and with stricter quotas. That layer is gone. Anthropic confirmed it has been removed entirely for paid plans.
Colossus 1 is the muscle behind the move
Colossus 1 sits on the outskirts of Memphis, Tennessee. xAI built it in roughly nine months in 2024, which industry watchers at the time called the fastest hyperscale buildout in modern memory. The site runs a mix of NVIDIA H100, H200, and the newer GB200 accelerators. SpaceX took an ownership stake in the cluster during xAI’s restructuring earlier this year, which is how it ended up as the leasable asset Anthropic just locked down.
Per the joint statement, Anthropic gets exclusive access to all current Colossus 1 capacity within the month. That is roughly equivalent, by one independent estimate cited by The New Stack, to the entire compute budget Anthropic ran on in 2024.
Pro and Max users will feel it first
If you live inside Claude Code, the change is loud. Long agentic runs that used to die at the five-hour boundary now finish. Multi-step refactors, large repo audits, long-context document work, all of these were limit-constrained. The cap was the ceiling. The ceiling just moved.
For builders on the API, the jump was even steeper. Anthropic raised Tier 1 input tokens per minute from 30,000 to 500,000, a 16x increase for entry-tier developers. Engineers I follow on X were testing it within hours and posting screenshots of jobs that previously needed exponential backoff retries running clean.
The free plan got nothing
This is the line Anthropic put in the announcement and then did not explain. Free plan users continue at the same daily message limit. No bump. No reduced throttling. The signal is straightforward. The new compute is paid for, and the people paying get the upside.
For context, Anthropic is reportedly approaching an initial public offering this June at a valuation north of $300 billion. The pre-IPO posture of monetizing the paid tiers harder, while keeping the free funnel slim, fits the playbook.
Orbital compute is the part nobody is talking about
Buried in the second half of the joint statement, the two companies said they will jointly explore “multiple gigawatts of orbital AI compute capacity.” That is not a typo. Anthropic wants to put inference clusters in low earth orbit. SpaceX has the launch cadence and the Starshield platform to make it conceivable.
The pitch is power and cooling. A solar-powered cluster in orbit gets uninterrupted sun, radiates heat into a 3-kelvin sink, and skips every grid-permit fight on the ground. Anthropic has not committed a timeline, but the wording in the announcement suggests early prototypes within 24 months.
What it means for the AI compute race
This deal joins a stack of capacity agreements Anthropic has been signing all year. Up to five gigawatts each from Amazon and Google through Broadcom. A $30 billion Azure commitment with Microsoft and NVIDIA. A separate $50 billion American infrastructure push with Fluidstack. Stack them and Anthropic is now contracted for more GPUs than any AI lab has ever lined up at once.
The market took the news straight. Anthropic’s secondary-market shares ticked up. NVIDIA closed green. xAI, which still uses Colossus 2 for its own training, said the lease does not affect Grok’s roadmap. The one company that did not get a friendly mention was OpenAI, which spent most of the week on a court bench in Sam Altman’s testimony in the Musk lawsuit.
Why This Matters
Token limits sound like an arcane product detail until you live inside them. For a developer running Claude Code on a real codebase, a 44,000-token five-hour window meant choosing what to ship today. For a researcher running long context analysis, it meant chopping documents. Doubling the limit changes the unit economics of what you can finish in a single session.
Zoom out and the deal is also a signal about how the AI race is being fought now. The bottleneck stopped being model architecture two years ago. It is compute. Specifically, it is the megawatts of cooled, powered, GPU-stuffed real estate that any given lab can actually access. Anthropic just bought itself a bigger room.
USABlaze Takeaway
If you pay for Claude, log in today and check your limits. The bump is real and it is live. If you build on the Claude API, your Tier 1 rate just got a 16x lift. If you are still on the free plan and you have been waiting for relief, this is not your week.
The bigger bet is the orbital compute play. It sounds like sci-fi until you remember Starlink launches the same way. Watch the next Anthropic earnings call. If they put a satellite render on the slide, the entire AI infrastructure conversation shifts.
Sources: Anthropic, Bloomberg, CNBC, Data Center Dynamics, The New Stack.
Editorial: Reported and edited by the USABlaze staff. Have a tip? Email editor@usablaze.com.