Anthropic Leases All 220,000 GPUs at SpaceX Colossus 1 to Meet Surging Claude Agent Demand

Anthropic signed a deal on May 6 to lease the entire Colossus 1 data center from SpaceX, gaining access to more than 220,000 NVIDIA GPUs and 300 megawatts of compute capacity. The same day, the company doubled Claude Code’s five-hour rate limits, removed peak-hour throttling for Pro and Max subscribers, and raised API rate limits on Claude Opus models. The message was blunt: Anthropic’s agent products were hitting an infrastructure ceiling.

The deal is notable for who sits on each side. Elon Musk previously called Anthropic “misanthropic” and labeled Claude “evil.” Now his company is Anthropic’s GPU landlord. Musk said on X that meetings with senior Anthropic staff had changed his view, calling them “competent and focused on responsible development,” according to TechWire Asia.

Why SpaceX Is Renting Out Its Supercomputer

Colossus 1 was built in Memphis, Tennessee to train Grok, xAI’s language model. According to Tom’s Hardware, the cluster’s mixed GPU architecture (multiple generations of NVIDIA hardware) made it inefficient for the large-scale training runs frontier models require. SpaceXAI has shifted Grok training to Colossus 2, a Blackwell-only cluster designed for that workload.

That left 300 megawatts of inference-capable compute sitting underused. SpaceX, preparing for an IPO, found a customer.

As TechCrunch’s Equity podcast put it: “They are a neocloud now, in the sense that they had to do something with all this compute that they were building, because it certainly seems like they were not going to need it for Grok.” The hosts noted that Grok sees limited enterprise adoption and that xAI employees were reportedly using competitor models internally before SpaceX’s acquisition triggered a leadership shakeup.

xAI has since been dissolved as a separate entity. SpaceXAI is now the brand.

The Agent Inference Problem

The compute bottleneck Anthropic is solving is specifically an agent problem. Claude Code sessions are long-running, multi-turn workflows where a single user can consume hundreds of thousands of tokens per hour. Autonomous agents built on Claude’s API through tools like OpenClaw or Claude Managed Agents compound the problem: they run continuously, retry on failure, and scale horizontally.

Anthropic co-founder Tom Brown said on X that Claude inference would begin moving onto Colossus “in the coming days,” according to TechWire Asia. The company’s own announcement framed the capacity expansion around agentic use cases: higher Claude Code limits, higher Opus API rate limits, and removal of peak-hour throttling.

This is a pattern across the industry. The International Energy Agency expects data center electricity consumption to roughly double from 485 TWh in 2025 to 950 TWh in 2030, with AI-focused data centers growing even faster, according to TechWire Asia.

The Capital Moat Widens

Anthropic now has compute partnerships totaling staggering capacity: up to 5 GW with Amazon (nearly 1 GW by end of 2026), 5 GW with Google and Broadcom (coming online 2027), $30 billion of Azure capacity through Microsoft and NVIDIA, and a $50 billion infrastructure investment with Fluidstack, per Anthropic’s own disclosure. The SpaceX deal adds another 300 MW on top.

This level of capital intensity creates a structural barrier. Smaller model providers and open-source projects can match Claude on benchmarks. They cannot match this inference footprint. For teams building production agent systems, the question of which model provider can sustain their workload at scale is becoming as important as which model scores highest on SWE-bench.

Venture capitalist Chamath Palihapitiya, despite warning that Anthropic “risks becoming the Friendster of the AI era” after Claude refused a stock screening prompt, inadvertently illustrated the competitive pressure Anthropic faces: not just on model capability, but on the perception that safety guardrails constrain utility. Anthropic’s response has been to solve the infrastructure side first. Claude can afford to be selective about which tasks it accepts if it has enough compute to serve the ones it does accept without throttling.

What This Tells Agent Builders

The Colossus deal reframes a question agent developers have been treating as settled. Most architecture decisions for autonomous systems focus on model selection, tool integration, and orchestration logic. The assumption is that inference capacity is someone else’s problem.

Anthropic’s behavior suggests otherwise. A company that just raised at a $380 billion valuation is scrambling to lease a rival’s supercomputer because its own infrastructure cannot keep up with agent demand. That scarcity flows downstream. Rate limits, throttling, and capacity constraints directly affect agent reliability. A system that works in testing can degrade in production when the provider’s GPUs are saturated.

The deal also introduces a new dependency. Anthropic is now running inference on hardware owned by a SpaceX entity preparing for a public offering. If SpaceX’s priorities shift post-IPO, or if the environmental lawsuit over Colossus 1’s unpermitted gas turbines changes the facility’s operating status, the compute underpinning Claude agents could be affected.

For now, Anthropic has 220,000 more GPUs. Claude Code limits are doubled. The orbital compute partnership mentioned in the announcement (“multiple gigawatts of orbital AI compute capacity”) remains aspirational. The terrestrial problem is real today.