Elon Musk posted on X Wednesday that SpaceX is now providing AI compute as a service at "significant scale," pointing to the company's expanded partnership with Anthropic as proof of concept. The announcement arrived alongside SpaceX's long-awaited IPO filing, which disclosed that Anthropic will pay $1.25 billion per month through May 2029 for access to the company's Colossus data center infrastructure.
"We are in discussions with other companies to do the same," Musk wrote. "Over time, especially with orbital data centers, we expect to serve AI at extremely high scale."
The deal gives Anthropic full access to Colossus 1, a supercomputer in Memphis originally built to train xAI's Grok models. The facility delivers over 300 megawatts of power across more than 220,000 NVIDIA GPUs, including H100, H200, and next-generation GB200 accelerators. To put that in perspective: a typical hyperscale data center runs on about 100 megawatts. SpaceX is offering triple that to a single client.
The Economics of Surplus Compute
What makes this arrangement worth examining is the underlying market dynamic. SpaceX, through its February acquisition of xAI valued at $1.25 trillion, inherited a GPU fleet it wasn't fully utilizing. Reports from earlier this year suggested xAI was achieving only 11% utilization on its hardware. Rather than let expensive silicon sit idle, SpaceX is monetizing the overcapacity. Anthropic, meanwhile, reported an 80-fold surge in revenue and usage in early 2026, far outpacing its infrastructure. The company had been aggressively pursuing compute through deals with Amazon, Google, Microsoft, and NVIDIA.
The SpaceX IPO filing confirms the company expects to sign more compute-as-a-service contracts. "We expect to enter into additional similar services contracts," the filing states. "We have sufficient capacity to provide compute for our own AI models, including support of our training and inference demands, and to satisfy the obligations under these agreements."
This positions SpaceX as what the industry now calls a "neocloud." Companies with massive GPU infrastructure are discovering they can lease excess capacity at premium rates to AI labs starving for compute. It's a seller's market, and the sellers are the ones who moved first on data center buildouts.
Orbital Ambitions
Musk's statement also highlights the longer play. In January, SpaceX filed with the FCC to launch up to one million satellites designed to function as orbital data centers. The filing described satellites operating between 500 and 2,000 kilometers altitude, using sun-synchronous orbits to maximize solar power collection.
"By directly harnessing near-constant solar power with little operating or maintenance cost, these satellites will achieve transformative cost and energy efficiency," the company wrote in its FCC application.
Anthropic has already expressed interest in partnering on that effort. According to SpaceX's announcement of the original deal in May, Anthropic wants to help develop "multiple gigawatts of orbital AI compute capacity."
Skeptics remain. Jonathan McDowell, an astronomer at the Harvard-Smithsonian Center for Astrophysics, has called the scale "staggering" and doubts the timeline. SpaceX's own IPO filing warns that orbital data centers "may never become commercially viable." But the company expects to begin deploying them "as early as 2028."
Leverage in a Constrained Market
The Anthropic deal illustrates a broader shift: companies that built GPU infrastructure early, for whatever purpose, now hold leverage over those that didn't. SpaceX built Colossus to train its own models. The fact that it's now leasing capacity to a direct competitor says less about any thaw between Musk and Anthropic than about how expensive compute scarcity has become.
The $15 billion annual payment from Anthropic represents nearly as much as SpaceX's entire 2025 revenue of $18 billion. AI compute rental is already transforming the company's financial profile before it even goes public.
For the AI labs still scrambling to secure capacity, the lesson is uncomfortable. The next frontier model might depend less on algorithmic breakthroughs than on which companies locked up GPU supply contracts early. And the pricing power belongs to whoever happened to overbuild.
