Anthropic has struck a major compute-services agreement with SpaceX that gives the Claude maker access to all of the capacity at Colossus 1, a Memphis, Tennessee, data center operated by SpaceXAI. Anthropic says the arrangement will add more than 300 megawatts of capacity and more than 220,000 NVIDIA GPUs within the month, making it one of the largest disclosed infrastructure commitments behind its Claude products. [1]
The immediate effect is consumer-facing: Anthropic has already raised Claude Code and API limits, framing the added infrastructure as a response to demand from developers and paid subscribers. The deal also puts a high-profile AI customer inside Elon Musk’s expanding compute operation and includes exploratory discussions about eventually building multigigawatt AI-compute systems in orbit. That orbital component remains an ambition rather than a committed project. [1][2]
Colossus capacity will support Claude availability
Anthropic said it will use the full computing capacity of Colossus 1, which SpaceXAI describes as a dense deployment of NVIDIA H100, H200 and GB200 accelerators. The systems can support model training, fine-tuning, inference and other high-performance workloads, although neither company disclosed how Anthropic’s allocation will be divided among those tasks. [1][2]
The companies also did not disclose financial terms, the agreement’s duration, the form of Anthropic’s access to the hardware, or details of networking, storage and power allocation. The stated figure of more than 300 megawatts establishes the scale of the commitment, but the announcement does not specify whether that number refers solely to IT equipment or to a broader measure of facility power.
Anthropic’s first priority is relieving capacity pressure on existing services. The company doubled Claude Code’s five-hour rate limits for Pro, Max, Team and seat-based Enterprise plans, removed peak-hour reductions for Claude Code Pro and Max accounts, and increased API rate limits for Claude Opus models. It said the newly available SpaceX capacity would directly improve availability for Claude Pro and Claude Max subscribers. [1]
At Anthropic’s developer conference, head of product Ami Vora said the company would use all of Colossus 1’s capacity. The message is notable because it connects a large-scale infrastructure agreement directly to product limits, rather than presenting the GPUs only as a future training resource. [3]
A broader race for power, chips and data centers
The agreement illustrates how competition in AI infrastructure has expanded beyond securing accelerators. Leading model developers now need power contracts, data-center construction, cooling, networking and long-term access to operating capacity. Those constraints have become especially visible as AI coding tools and other inference-heavy services attract sustained demand.
For Anthropic, Colossus adds a substantial NVIDIA-based resource alongside an increasingly diverse infrastructure portfolio. The company says Claude runs across AWS Trainium, Google TPUs and NVIDIA GPUs. It has also cited arrangements including up to 5 gigawatts with Amazon, a 5-gigawatt Google and Broadcom agreement beginning in 2027, $30 billion in Azure capacity through Microsoft and NVIDIA, and a planned $50 billion U.S. AI-infrastructure investment with Fluidstack. [1]
That diversity can reduce dependence on any single chip supplier or cloud operator, while giving Anthropic access to different hardware platforms for different workloads. The SpaceX deal is distinctive because it provides a single large block of NVIDIA capacity at a facility associated with Musk’s AI operations.
For SpaceXAI, the agreement provides a marquee external customer as it promotes its computing infrastructure beyond its own AI efforts. Contemporary reporting has described SpaceX and xAI as having merged earlier this year, with the combined operation referred to as SpaceXAI. Musk said SpaceX had moved its own AI training to Colossus 2 and would make capacity available to companies pursuing similar goals. [4][5]
The partnership also marks a conspicuous change in tone. Musk had criticized Anthropic earlier this year over its policies and alleged model bias, but said after meeting Anthropic executives that he was impressed by their efforts to ensure Claude was developed for humanity’s benefit. [5]
Orbital compute is an exploratory proposal, not a deployment plan
Alongside the Memphis agreement, Anthropic said it had expressed interest in working with SpaceXAI to develop multiple gigawatts of orbital AI-compute capacity. SpaceXAI argues that terrestrial data centers face growing limits around power availability, land and cooling, while space-based systems could theoretically draw continuous solar power and bypass some grid constraints. [1][2]
However, the companies have not announced a satellite design, launch date, operating altitude, investment amount, power architecture, deployment timetable or a finalized construction program. The current disclosure describes potential collaboration, not an operational space data center or a jointly developed AI model.
Space-based compute would face difficult engineering and economic tradeoffs. Josep Jornet, a Northeastern University professor of computer and electrical engineering, told the Associated Press that chips in space require robust thermal systems and that facilities at the proposed scale would need massive radiator arrays. Such structures would be large and fragile, with no prior example at the scale being discussed. [6]
Other concerns include orbital debris, radiation exposure and repairability. John Crassidis, a University at Buffalo professor and former NASA engineer, warned that huge satellite constellations could create collision risks, while Aetherflux CEO Baiju Bhatt pointed to the challenge of repairing hardware in orbit and protecting GPUs from high-energy particles. Eurospace research director Pierre Lionnet also suggested that orbital-data-center plans could help preserve preferential launch economics for SpaceX’s own infrastructure. [6]
Musk has said space could become the most economically compelling location for AI compute within roughly 30 to 36 months. That prediction was not accompanied by a detailed engineering plan or independent cost validation. [6]
Memphis facility faces environmental scrutiny
The terrestrial portion of the deal carries its own questions. WIRED reported that emissions from gas turbines at the Memphis Colossus facility have prompted complaints from local residents, and that environmental protesters demonstrated near a SpaceX investor event. [4]
Those concerns underline a tension at the center of the AI infrastructure boom: expanding compute capacity can improve availability and accelerate model development, but it also requires enormous amounts of electricity, water or other cooling resources, land and supporting equipment. SpaceXAI’s orbital vision is partly pitched as an answer to terrestrial constraints, yet it introduces a separate set of unresolved technical and environmental risks.
For now, the concrete outcome is on the ground. Anthropic gains a large new pool of GPU capacity in Memphis, and Claude users are already seeing higher limits. The orbital-compute discussion is strategically significant, but it remains far less defined than the Colossus 1 services agreement.
Editor’s Take
This is a meaningful deal because it addresses the constraint Claude users actually feel: inference capacity. Raising Claude Code and API limits is more commercially important in the near term than another vague training-compute announcement. A large dedicated block of NVIDIA capacity gives Anthropic more room to serve coding workloads, where latency, availability and rate limits directly determine whether a developer keeps a tool in the workflow.
The important caveat is that the headline scale still lacks the operational details that determine its value. We do not know the contract term, effective GPU availability, network design, storage architecture, power definition or how much of the capacity goes to training versus serving users. I would watch for sustained improvements in Claude availability and API throughput, rather than treating the 300-megawatt figure as proof that every one of those resources is immediately productive.
Orbital compute is the part where the story outruns the facts. Space has attractive solar-power and land-use arguments, but heat rejection, radiation hardening, launch economics, maintenance and debris risk are not minor implementation details. For now, the investable story is terrestrial: AI companies are becoming power-and-data-center buyers as much as model builders, and reliable capacity is rapidly becoming a product feature.
References
- Anthropic – https://www.anthropic.com/news/higher-limits-spacex
- SpaceXAI – https://x.ai/news/anthropic-compute-partnership
- Axios – https://www.axios.com/2026/05/06/anthropic-spacex-elon-musk-compute
- WIRED – https://www.wired.com/story/anthropic-spacex-compute-deal-colossus/
- Investing.com – https://www.investing.com/news/stock-market-news/anthropic-unveils-dreaming-feature-to-help-its-ai-agents-selfimprove-4664442
- Associated Press – https://apnews.com/article/elon-musk-orbital-ai-data-centers-xai-spacex-92bc8ad95593bf3b5b801ddf36427194
_024.png){kind=link}