Anthropic has made a significant move in the race for AI infrastructure. The company has signed an agreement with SpaceX to utilize the entire computing capacity of the Colossus 1 data center, a facility that will provide access to over 300 MW and more than 220,000 NVIDIA GPUs within a month. The immediate impact for users is clear: Claude Code and the Claude Opus API will have much higher usage limits.
The news perfectly summarizes the current state of the sector. For months, competition among AI labs has focused on more capable models, larger context windows, and better benchmarks. But the real bottleneck increasingly lies elsewhere: energy, GPUs, data centers, networks, cooling, and the capacity to serve millions of requests without degrading the experience. In other words, generative AI no longer competes solely with intelligence—it competes with megawatts.
Claude Code will have more room to operate
Anthropic has announced three immediate changes for its most intensive users. First, it doubles the usage limits for Claude Code in five-hour windows for Pro, Max, Team, and Enterprise plans per seat. Second, it removes the peak-hour limit reductions for Claude Code on Pro and Max plans. Third, it substantially increases API limits for Claude Opus models.
The new API limits demonstrate how much capacity Anthropic needed to add. In Tier 1, maximum input tokens per minute jump from 30,000 to 500,000, while output tokens increase from 8,000 to 80,000. In Tier 2, input goes from 450,000 to 2 million tokens per minute, and output from 90,000 to 200,000. Tier 3 sees input limits rising from 800,000 to 5 million tokens, and output from 160,000 to 400,000. Finally, Tier 4 reaches up to 10 million input tokens per minute and 800,000 output.
For occasional users, these numbers might seem distant. But for teams working with agents, large repositories, document analysis, code generation, contract review, or business automation, they are game-changing. A programming agent doesn’t ask a single question—it reads files, proposes changes, reviews errors, executes steps, consultas context again, and generates lengthy responses. Each action consumes tokens and inference capacity.
That’s why limits have become a key part of the product experience. When a developer runs out of margin mid-session, the tool stops being a smooth assistant and becomes an operational constraint. Anthropic aims to reduce this feeling precisely as Claude Code gains weight among technical profiles.
SpaceX enters Claude’s infrastructure
The agreement with SpaceX has both a technical and a strategic reading. Technically, it’s straightforward: Anthropic gains immediate access to a large pool of GPU capacity. Strategically, it’s more interesting: a company known for rockets, satellites, and connectivity becomes a relevant infrastructure provider for one of the most advanced AI labs.
Reuters locates Colossus 1 in Memphis, Tennessee, noting that the center has more than 220,000 NVIDIA processors, including H100, H200, and GB200 models. This operation helps Anthropic alleviate a clear pressure: demand for Claude Pro, Claude Max, Claude Code, and the API is growing faster than the available capacity to meet it comfortably.
Anthropic does not rely on a single architecture. The company reports training and running Claude on AWS Trainium, Google TPUs, and NVIDIA GPUs. This diversification is vital in a market where any model provider depending on a single chip type or partner risks being trapped by availability, price, or third-party commercial priorities.
Nevertheless, the volume of Colossus 1 makes this announcement more than just a service upgrade. Over 300 MW is a scale typically associated with large data centers dedicated to AI. It’s not about adding a few machines to handle peaks but deploying a whole block of infrastructure to lift limits, reduce restrictions, and prepare for more intensive workloads.
The race is now measured in gigawatts
The SpaceX deal doesn’t stand alone. Anthropic presents it alongside other large-scale compute commitments: up to 5 GW with Amazon (with nearly 1 GW of new capacity expected by end-2026); a 5 GW deal with Google and Broadcom set to start service in 2027; a partnership with Microsoft and NVIDIA including $30 billion in Azure capacity; and a $50 billion investment in US AI infrastructure alongside Fluidstack.
The scale is impressive because it shows where the market is heading. Not long ago, a data center with tens of megawatts seemed enormous. Now, AI has moved the conversation to hundreds of megawatts and multi-gigawatt deals. This evolution affects not just labs but entire power grids, GPU availability, data center construction, permits, cooling, transformer supply, and land-use planning.
Inference consumption is increasingly significant. While training foundational models remains extremely costly, serving those models at scale to millions of users and agents consumes substantial capacity as well. If agents run long tasks, maintain context, consult tools, and generate extensive responses, operational costs soar. The increase in Claude’s limits is good news for users but also confirms that real-world AI usage is growing at a pace demanding more infrastructure.
Anthropic is addressing potential criticisms. The company states it commits to covering any electricity price increases caused by its US data centers. It also considers extending this commitment to other countries where it deploys new capacity. This point will become more sensitive as social acceptance of AI data centers hinges on communities perceiving clear benefits or just additional strain on electrical grids.
Regional infrastructure for regulated industries
The expansion won’t be limited to the US. Anthropic notes that its enterprise clients—especially in regulated sectors like finance, healthcare, and government—require regional infrastructure to meet data residency and compliance standards. The collaboration with Amazon includes additional inference capacity in Asia and Europe.
This detail is critical in the enterprise market. Companies may be interested in Claude but cannot send sensitive data outside their region. Banks, insurers, hospitals, public agencies, and firms with strategic information need guarantees about where and how their data is processed, who operates the infrastructure, and under what legal framework the service is offered.
Anthropic affirms it will be selective in where it adds capacity, prioritizing democracies with robust legal and regulatory frameworks capable of supporting investments at this scale, along with secure supply chains for hardware, networks, and facilities. AI infrastructure is becoming an issue of technology, economics, and geopolitics simultaneously.
The expressed interest in collaborating with SpaceX on deploying several gigawatts of orbital AI capacity adds a speculative, futuristic note. It’s wise to approach this cautiously: orbiting compute faces huge challenges in cost, maintenance, energy, radiation, cooling, communications, and hardware replacement. Still, the fact that a company like Anthropic mentions such plans highlights how labs are exploring alternatives beyond traditional terrestrial data centers.
The immediate outcome is clearer: Claude will have higher limits, fewer restrictions within Claude Code, and greater capacity for customers with high token volumes. For Anthropic, this move makes sense—if it hopes to compete with OpenAI, Google, xAI, Meta, and others, it must ensure its top models aren’t bottlenecked by GPU availability.
The next phase of AI won’t be won solely by having the smartest model. It will depend on how quickly, reliably, and broadly that model can be deployed to millions of users and businesses. Anthropic has just secured a significant part of that margin.
Frequently Asked Questions
What has Anthropic announced with SpaceX?
They have signed an agreement to use the full computational capacity of SpaceX’s Colossus 1 data center, with over 300 MW and more than 220,000 NVIDIA GPUs available within a month.
What improvements does this bring for Claude Code users?
Claude Code doubles its usage limits in five-hour windows for Pro, Max, Team, and Enterprise plans per seat, and removes peak-hour limit reductions for Pro and Max plans.
What changes are there for the Claude Opus API?
Anthropic significantly increases token limits per minute. For example, in Tier 4, input tokens go from 2 million to 10 million per minute, and output tokens from 400,000 to 800,000.
Why do AI labs need so many gigawatts?
Because training and serving advanced models require vast amounts of GPU power, energy, cooling, and network capacity. Agents and tools like Claude Code generate many chained calls and consume more tokens than simple conversational use.
source: Noticias inteligencia artificial