Anthropic has just signed a $1.25 billion per month contract until May 2029, totaling over $40 billion over the duration, to reserve a GPU park whose operator described the effective utilization rate as "embarrassingly low" just weeks earlier, at 11%. Considering Anthropic's annualized revenues in April 2026 - over $30 billion, this single compute contract absorbs about half of the ARR of the Claude publisher, roughly 6.3 times OpenAI's annual commitment on the $11.9 billion deal with CoreWeave over five years. The contract's objective: exclusive access to the entirety of Colossus 1, the supercomputer in Memphis, Tennessee - over 220,000 Nvidia GPUs (H100, H200, and GB200), 300 megawatts of power, and availability in less than a month, as documented by Business Insider on May 20, 2026, from SpaceX's S-1 and confirmed by an Anthropic spokesperson. An extension to Colossus 2 is planned. Immediate effects announced on Claude Code: doubling of bandwidth limits over 5 hours for Pro, Max, Team, and Enterprise subscriptions, and removal of peak hour caps for Pro and Max plans.
A park built by xAI, absorbed by SpaceX, now leased to a competitor
Colossus 1 was initially designed and operated by xAI, Elon Musk's artificial intelligence company, before its acquisition by SpaceX during a stock merger in February 2026. ActuIA had followed since 2024 the quest for computing power by Musk for xAI which put pressure on NVIDIA, a supercomputer designed for Grok training loads on the scale of the Nvidia H100/H200/GB200 park. Two contextual elements structure the present contract. On one hand, SpaceX confidentially filed an S-1 with the SEC on April 1, 2026, targeting an IPO valuation between $1,750 and $2,000 billion, with a roadshow planned for the week of June 8, according to CoinDesk. On the other hand, Elon Musk stated on his X network that he reserves the right to reclaim computing capacity if Anthropic's AI "engages in actions harmful to humanity" (loosely translated); the exact contractual status of this clause has not been confirmed jointly by both parties. The operational context is also documented: an internal xAI memo signed by its president Michael Nicolls and obtained by Business Insider describes a Model FLOPs Utilization (MFU) rate of 11% on the roughly 550,000 Nvidia GPUs operated as "embarrassingly low", where production-grade LLM training is between 35 and 45% MFU, Meta reaching 43% on Llama 3 training, Google 46% on PaLM, and ByteDance's MegaScale posting 55.2% MFU on 12,288 GPUs during training of a 175 billion parameter model, this latter result documented in the arXiv paper published by ByteDance in February 2024.
Pricing mechanics: a capacity reservation, not a per-hour GPU rate
Three caveats are necessary before any price comparison. Firstly, the contract is not structured as a per-GPU usage price but as a reservation of 300 megawatts of guaranteed capacity: the monthly amount remunerates the availability of the infrastructure, regardless of the compute actually delivered. Related to the park of 220,000 GPUs over 720 monthly hours, the implicit cost per GPU per hour is approximately $7.78 according to ActuIA's estimation. To situate this magnitude, IntuitionLabs notes that H100 on-demand rates at AWS and Google Cloud are around $3 to $4 per GPU per hour, while new-clouds like Lambda Labs, RunPod, Vast.ai, or Cudo Compute offer rates down to $1.49–2.99, AWS having reduced its H100 prices by about 44% in June 2025. Direct comparison remains tricky due to the nature of commitment: the H100 on-demand (hourly billing) is an hourly adjustment variable, the Colossus contract is a 36-month capacity commitment - two pricing regimes relating to different objects.
Secondly, the 11% MFU rate relates to Grok training conducted by xAI, while Anthropic primarily intends these GPUs for Claude inference. The two regimes are not comparable: training is compute-bound (demanding in floating-point operations), inference is mainly memory-bandwidth-bound (limited by memory bandwidth), and the MFU metric is not a direct indicator of an inference workload's efficiency. Thirdly, SpaceX recorded, according to its S-1, an operational loss on the AI compute segment in the first quarter of 2026, whose magnitude (about $2.5 billion over the quarter, for segment revenue of about $800 million) places Anthropic's monthly rate below the operator's current profitability threshold: the agreed price reflects a strategic underpricing decision, not a stabilized market signal. The annualized extrapolation from this single quarter requires caution, as the S-1 does not publish a consolidated historical series.
The contract's weight for Anthropic and comparable OpenAI/CoreWeave
The Colossus commitment represents about $15 billion per year over three years, compared to Anthropic's revenue trajectory. According to the Los Angeles Times, Anthropic's annualized revenue (ARR) more than doubled since summer 2025, surpassing $9 billion by the end of 2025, then $14 billion in February 2026, $19 billion in March, and over $30 billion in April 2026. In light of this April ARR, the sole Colossus contract mobilizes about half of Anthropic's annualized revenues - a ratio without recent equivalent for an isolated infrastructure agreement.
The most immediate sectoral comparison is the OpenAI/CoreWeave agreement. According to The Next Platform, OpenAI signed a five-year contract for $11.9 billion with CoreWeave, which has about 250,000 GPUs spread across 32 data centers. That is $2.38 billion per year, compared to $15 billion for Anthropic/SpaceX on a comparably sized park: the Claude publisher's annual commitment is about 6.3 times higher than OpenAI's on the public neo-cloud perimeter. A financial reading of the contract must also consider its insertion into the SpaceX IPO calendar. This combination - underutilized capacity monetized at a high rate - mechanically reinforces the S-1's revenue narrative, regardless of the park's effective yield.
What Colossus delivers to Anthropic, what orbit has yet to deliver
On the immediate operational perimeter, Anthropic assigns the Colossus capacity primarily to Claude inference, a workload distinct from training by its dominant pressure on memory bandwidth and its user latency constraints. The announced extension to Colossus 2 inscribes the device in a trajectory of scaling up on the inference segment, as the Claude publisher saw its user throughputs strained multiple times during Claude Code traffic peaks in 2025.
The second part of the announcement, however, does not have the same status. Anthropic has expressed interest in developing with SpaceX several gigawatts of AI computing capacity in orbit, but no contract has been signed at this stage. MIT Technology Review documents four technical obstacles to a space data center on this scale. In the space vacuum, heat can only be dissipated by radiation, a much less efficient process than convection; electronic components face single-event upsets (logic disturbances induced by ionizing particles), cumulative ionization damage, and displacement damage, according to the MIT magazine, which adds three converging factors: the temperature in orbit would not drop below 80°C (beyond the safe operational range of commercial electronics), space-hardened components are significantly more expensive and lag several years behind the state of the art, and orbital assembly at this scale would require robotic systems that do not exist. As of May 21, 2026, the project falls within the declarative register: no orbital contract announcement has been made public.