Anthropic Leases SpaceX Colossus 1: 220K GPUs, Claude Rate Limits Doubled

Anthropic announced on May 6, 2026 that it has leased SpaceX's Colossus 1 data center in Memphis, Tennessee — the same facility SpaceX built to train Grok 3 and Grok 4. The deal gives Anthropic access to over 220,000 NVIDIA GPUs across H100, H200, and GB200 configurations, drawing more than 300 megawatts of power. The immediate user-facing effect: Claude Code's 5-hour usage limits are doubling, peak-hour throttling is being removed, and Claude Opus API rate limits are increasing for enterprise customers.

This is the largest single compute acquisition in Anthropic's history, and it comes at the same moment xAI is being dissolved into SpaceX under a reorganisation Elon Musk announced separately as SpaceXAI. Colossus 1 — which xAI no longer needs at full capacity under the merger — is now effectively Anthropic's primary inference and training cluster.

What Colossus 1 Is

Colossus 1 was completed in Memphis in September 2024 in a record 122-day build. SpaceX (through xAI) packed the facility with NVIDIA H100s before transitioning to H200s and GB200 NVL72 racks through 2025. At full draw it consumes approximately 300 megawatts — comparable to the power draw of a mid-sized city district.

The facility originally handled Grok model training and was the GPU cluster behind xAI's claim that it trained Grok 3 faster than any previous frontier model. After xAI's merger into SpaceX, the compute capacity became available for external lease. Anthropic moved quickly to sign.

The deal structure is a long-term lease rather than an acquisition. Anthropic does not own the building or the physical infrastructure — SpaceX retains ownership. Anthropic purchases compute time and capacity allocation, with contractual commitments on availability and priority access. Terms of the financial arrangement have not been disclosed.

What Changes for Developers Using Claude

The practical effects announced with the deal:

Claude Code rate limits: The 5-hour rolling usage window doubles in effective capacity. The mechanism Anthropic uses is not extending the 5-hour window itself but increasing the token budget available within it. For developers running long agentic coding sessions, this means fewer mid-session interruptions.

Peak-hour throttling removed: Anthropic previously reduced API throughput during high-demand windows (roughly 9 AM - 5 PM US Eastern). That throttling is being removed as the Colossus 1 capacity comes online. Real-time applications and workflows sensitive to latency variance will see more consistent response times.

Opus API limits for enterprise: Claude Opus tier customers on enterprise plans are seeing rate limit increases. Anthropic has not published exact new limits — they vary by plan tier — but the increase is described as significant enough to unblock use cases that previously required batching or off-peak scheduling.

The rollout is phased. As of May 6, Claude Code limits are the first to increase; API limits for Opus follow over the subsequent two to three weeks as Colossus 1 capacity is formally migrated and validated.

The xAI-SpaceX Merger Context

The timing of this deal is directly tied to xAI's restructuring. Elon Musk announced that xAI is being formally folded into SpaceX under the SpaceXAI brand, consolidating Grok, Colossus 1, and xAI's engineering team under SpaceX's corporate structure. The stated rationale is capital efficiency — maintaining a separate xAI entity with its own data centers, headcount, and fundraising overhead is no longer necessary when SpaceX can absorb the operation.

One consequence of that merger: SpaceX has more compute capacity than it needs for its own AI workloads in the near term. Leasing excess Colossus 1 capacity to Anthropic converts stranded infrastructure into revenue while Anthropic gets the compute expansion it needs without building a new data center from scratch.

From Anthropic's perspective, leasing is faster and cheaper than construction at this scale. Building a new 300 MW facility from greenfield takes 18-36 months and costs between $1.5B and $3B depending on power infrastructure. The Colossus 1 lease sidesteps that timeline entirely.

The Orbital Compute Angle

Separately, Anthropic and SpaceX are in exploratory talks about what has been described as "several gigawatts" of compute delivered via Starship-launched orbital data center hardware. This is conceptually distinct from the Colossus 1 lease — it is a research and feasibility discussion, not a signed commercial agreement.

The basic premise: a constellation of satellites carrying GPU compute, powered by solar, latency-linked to ground stations via SpaceX's Starlink backbone. For inference at extreme scale, orbital compute eliminates land acquisition, power grid constraints, and cooling infrastructure as bottlenecks. The engineering challenges — thermal management in vacuum, radiation hardening of GPUs, high-bandwidth ground uplinks — are unresolved at commercial scale.

This is a real conversation between the two companies, but treating it as imminent infrastructure would be a mistake. It is the kind of exploratory framing that precedes a 5-7 year development arc if it proceeds at all.

Why This Deal Is Significant Beyond the Rate Limits

The Colossus 1 deal signals something about the structure of the frontier AI industry: compute supply is increasingly separated from model development. Anthropic is a model development company that does not want to be in the data center business. SpaceX is now in the compute supply business as a consequence of building infrastructure for xAI.

This is the beginning of a pattern. As xAI merges into SpaceX, and as Google DeepMind, Meta, and Microsoft build surplus capacity in anticipation of future demand, there is a growing secondary market for frontier-scale compute. Anthropic is buying into that market. Cohere, Mistral, and other labs without hyperscaler backing will face increasing pressure to find similar arrangements or accept permanent compute disadvantages.

For developers: the near-term effect is better Claude availability and higher limits. The medium-term effect is that Anthropic's ability to train Claude 4.5, 5, and beyond is no longer as constrained by compute access as it was 18 months ago.

Key Takeaways

• Colossus 1 lease signed May 6, 2026: Anthropic leases SpaceX's Memphis data center — 300+ MW, 220,000+ NVIDIA H100/H200/GB200 GPUs — the largest single compute deal in Anthropic's history
• Claude Code rate limits double: 5-hour session token budget increases; peak-hour throttling removed; Claude Opus enterprise API limits increase in phased rollout over next 2-3 weeks
• xAI merger context: SpaceX is absorbing xAI into SpaceXAI brand; Colossus 1 excess capacity leased to Anthropic rather than sitting idle
• Lease not acquisition: Anthropic pays for compute access; SpaceX retains ownership; financial terms undisclosed
• Orbital compute exploratory: Anthropic and SpaceX in early talks on satellite-delivered GPU compute at "several gigawatts" — feasibility phase only, not a commercial agreement
• Industry pattern: Frontier labs without internal compute infrastructure will increasingly lease from excess capacity built by well-funded competitors; Colossus 1 is the first major instance of this pattern

For the xAI and SpaceXAI merger details, read xAI Dissolved Into SpaceXAI: Musk Merges Grok and Colossus Into SpaceX. For Anthropic's simultaneous enterprise expansion into banking AI, read Anthropic Goldman Sachs Blackstone $1.5B JV and FIS Banking Agents.