Days after the largest stock-market debut in history, the most surprising line in SpaceX's financials has nothing to do with rockets. According to disclosures in the company's IPO filings and reporting that followed, Elon Musk's space and satellite empire is now collecting roughly $2.17 billion a month by renting out idle AI compute capacity — much of it to the very companies racing against Musk in artificial intelligence.
The figure is a sum of two reported contracts. Anthropic, maker of the Claude models, agreed to pay about $1.25 billion a month to lease the entirety of Colossus 1, the Memphis supercomputer that xAI originally built to train Grok, according to SpaceX's IPO filing and reporting by DataCenterDynamics and TechCrunch. On June 5, Google added a second deal: roughly $920 million a month for access to about 110,000 Nvidia GPUs and supporting infrastructure, a contract running from October 2026 through mid-2029 that SpaceX values at more than $30 billion at full rate, per CNBC and TechCrunch.
Add them together and SpaceX has, almost overnight, become one of the largest landlords of AI compute on the planet — leasing to two of OpenAI's and xAI's chief rivals.
How a rocket company ended up with surplus GPUs
The path runs through xAI. On February 2, 2026, Musk folded his AI startup into SpaceX in an all-stock deal that valued the combined entity well over a trillion dollars, absorbing xAI's Colossus clusters in Memphis, Tennessee, and Southaven, Mississippi. xAI had spent aggressively — more than $12 billion of fresh capital and a sprawling buildout — to assemble what is now one of the world's largest GPU fleets, racing to keep Grok competitive.
The problem, laid bare in the filings, was utilization. Reporting on the Anthropic agreement pegged xAI's Grok training at roughly 11% model FLOPs utilization, meaning the overwhelming majority of that hardware sat idle. SpaceX's own numbers underscore the strain: its AI segment generated about $3.2 billion in 2025 revenue against a $6.4 billion operating loss, with $12.7 billion of the company's $20.7 billion in 2025 capital spending going to AI infrastructure, including Colossus.
Faced with billions in depreciating silicon and a public listing on the horizon, SpaceX did what any landlord with empty floors does: it found tenants. That the tenants are competitors is the twist. Musk had publicly disparaged Anthropic for years before the lease, and Google operates its own rival Gemini models and TPU hardware. Either side can reportedly walk from the Anthropic contract on 90 days' notice.
The economics of a compute crunch
The deals only make sense against the backdrop of an industry-wide scramble for capacity. AI labs are bottlenecked not by ideas but by power, chips, and the data-center shells to house them — and the lead time to build new ones stretches into years. In that market, whoever holds finished, energized capacity holds pricing power.
That is precisely SpaceX's accidental advantage. Anthropic's lease reportedly grants exclusive access to Colossus 1's 200,000-plus Nvidia GPUs and roughly 300 megawatts of power, with availability in under a month — a turnaround that would be impossible for a lab starting from a greenfield site. Anthropic intends much of that capacity for Claude inference rather than training, suggesting the deal is about serving surging customer demand right now, not future research.
"Why Google's SpaceX deal signals the rise of the AI compute landlord," ran one Business Standard headline, capturing the broader shift: compute is becoming a rentable utility, and the most valuable position in the value chain may not be building the best model but owning the warehouses where everyone else's models run. For SpaceX, the revenue also de-risks an enormous capital bet — turning a money-losing AI division into a recurring cash machine just as investors scrutinize the newly public stock.
There are reasons for caution. The $2.17 billion monthly figure is a reported sum of contracted run-rates, not audited recurring revenue, and the largest piece — the Google deal — does not begin in full until late 2026 and ramps up before then. Termination clauses, the volatility of GPU pricing, and the possibility that tenants build their own capacity all hang over the durability of the income.
SpaceX, the unexpected infrastructure player
Strip away the surprise and a pattern emerges. SpaceX has long been in the business of building expensive, capital-intensive infrastructure ahead of demand — launch capacity with Falcon and Starship, global connectivity with Starlink — and then renting it out. AI compute slots neatly into that model. Some reporting has even floated orbital and space-based data-center ambitions, though those economics remain unproven.
What is concrete is that SPCX debuted on the Nasdaq on June 12 at $135 a share, in an offering that reportedly raised tens of billions at a valuation north of $1.7 trillion — and that the AI-rental story is now part of the investment case. A division that looked like a costly distraction in early 2025 is being repackaged as a high-margin infrastructure franchise.
What to watch next
Three things will determine whether the $2 billion-a-month headline holds up. First, the Google contract's actual ramp: does revenue materialize on schedule starting in October, and at the reported rate? Second, tenant stickiness — whether Anthropic and Google renew, expand, or exercise their exit clauses as their own buildouts come online. And third, how SpaceX accounts for and discloses this revenue in its first quarters as a public company, which will tell investors how much of the headline figure is durable cash and how much is contracted optimism. If the numbers hold, the most consequential thing SpaceX builds this decade may not fly at all.