Anthropic Signs SpaceX Deal for 220,000 GPUs — and Doubles Claude Code Limits Immediately

The SpaceX Deal: What Anthropic Is Actually Getting

Anthropic has signed an agreement with SpaceX giving it access to the full compute capacity of Colossus 1, SpaceX's flagship AI data center. The facility contains over 220,000 NVIDIA GPUs — including dense deployments of H100, H200, and next-generation GB200 accelerators — and delivers more than 300 megawatts of power. According to Anthropic's announcement, this capacity will come online within the month and will directly improve service for Claude Pro and Max subscribers.

Colossus 1 was built by SpaceX (via its xAI subsidiary) and is one of the largest and fastest-deployed AI supercomputing clusters in the world. SpaceX acquired xAI in January 2026, and the facility had previously been associated with Grok model training. By bringing Anthropic in as a compute partner, SpaceX is positioning Colossus as shared infrastructure for multiple frontier AI labs — a significant shift in how the cluster is used.

Anthropic head of product Ami Vora confirmed the deal at the company's developer conference in San Francisco on May 6, 2026, describing it as an agreement to use "all the capacity" of the data center.

Why This Announcement Matters: The Compute Backstory

Since late 2025, Anthropic has been publicly compute-constrained. CEO Dario Amodei described the shortage as the primary bottleneck on model deployment, and developers using Claude Code experienced it directly: rate limits tightened through early 2026, and on March 26 Anthropic introduced a peak-hour throttle that reduced 5-hour session limits during US business hours (5 AM–11 AM PT) — the exact window when most developers do their heaviest work. Around 7% of subscribers hit session walls they had never hit before.

Anthropic acknowledged on March 31 that users were hitting limits "way faster than expected" and called it the team's top priority. Today's SpaceX deal is the direct answer to that problem: new hardware, online within weeks, powering an immediate reversal of the most unpopular policy the company has introduced in the past year.

Anthropic's Full Compute Stack — May 2026

The SpaceX partnership is one piece of a much larger infrastructure buildout Anthropic has been quietly assembling. The full picture as of today:

Anthropic trains and runs Claude across AWS Trainium, Google TPUs, and NVIDIA GPUs. The SpaceX deal adds NVIDIA GPU capacity outside the hyperscaler ecosystem — meaningful diversification at a moment when GPU availability remains the binding constraint across the entire AI industry.

The Orbital Compute Angle

The deal includes something more speculative: Anthropic has expressed interest in partnering with SpaceX to develop multiple gigawatts of orbital AI compute capacity. Space-based inference would sidestep the terrestrial constraints — land, power, cooling, regulatory approvals — that are slowing every AI lab's infrastructure buildout. No timeline or technical details were disclosed, but the expression of interest signals that Anthropic and SpaceX are thinking about compute on a longer horizon than today's GPU deals.

What Changes for Users — Right Now

Three changes are live as of May 6, 2026, and all three are directly tied to the new compute capacity coming online:

Claude Code 5-hour limits doubled. Every subscriber on Pro ($20/month), Max 5x ($100/month), Max 20x ($200/month), Team, and seat-based Enterprise now gets twice the Claude Code allowance within each 5-hour session window. The window structure and weekly reset schedule remain unchanged.

Peak-hour throttle permanently removed. The March 26 reduction that caused sessions to burn faster during 5 AM–11 AM PT weekdays is gone for Pro and Max. Subscribers can now code at consistent capacity across the full day, regardless of whether they are in Pacific, Eastern, or European time zones.

Opus API rate limits substantially raised. Developers accessing Opus models via the API get significantly more tokens-per-minute headroom. The pooled Opus limit covers Opus 4.7, 4.6, 4.5, 4.1, and Opus 4 combined. Updated figures are published in the Anthropic API rate limits documentation at platform.claude.com.

The Competitive Context: SpaceX's AI Ambitions

SpaceX is preparing for an IPO expected in autumn 2026, widely anticipated to be among the largest in corporate history. The company acquired xAI in January 2026 and has a separate deal with Cursor that reportedly includes an option to acquire the coding startup for $60 billion. By hosting Anthropic on Colossus 1 alongside its own xAI workloads, SpaceX positions its data center business as a neutral compute provider to the broader AI industry — a different strategy from keeping Colossus as an exclusive xAI asset.

For Anthropic, the arrangement is unusual: the company is now running on infrastructure owned by the parent of xAI, which builds Grok — a direct Claude competitor. The pragmatism of the deal reflects how tight GPU supply remains across the industry, and how much Anthropic needed new capacity quickly.

Frequently Asked Questions

When does the Colossus 1 capacity actually come online for Claude users?

Anthropic says the 300+ megawatts of new capacity will be available within the month from May 6, 2026. The rate limit improvements announced today are effective immediately, ahead of the full hardware coming online.

Does this affect the Claude Code token-inflation bug?

The compute deal addresses capacity, not the software bug. Claude Code v2.1.100+ introduced a token-inflation issue burning roughly 40% more tokens per request than expected. As of May 6 this remains unpatched in the current v2.1.126 build. The workaround is to reinstall via npm or downgrade to v2.1.34.

Does the Enterprise plan get the rate limit doubling?

Yes. Anthropic's announcement specifically includes seat-based Enterprise plans in the 5-hour rate limit doubling, in addition to Pro, Max, and Team.

What is Colossus 1?

Colossus 1 is SpaceX's AI supercomputing cluster, built and initially operated under the xAI brand. It features over 220,000 NVIDIA GPUs — including H100, H200, and GB200 accelerators — and was constructed in record time to support large-scale model training and inference. SpaceX acquired xAI in January 2026.

What does orbital compute mean and is it realistic?

Orbital compute refers to running AI inference workloads on hardware in orbit rather than in ground-based data centers. SpaceX has the launch infrastructure and power generation expertise to make this theoretically viable at scale. Anthropic's interest was described as exploratory, with no announced timeline. The appeal is bypassing terrestrial bottlenecks — land, power, cooling — that are constraining every AI lab's expansion right now.