A company that sends rockets into space may now be best known for its data centers.
Last year, executives suggested Elon Musk rent out idle compute capacity. He refused. His reasoning was very Musk-like: once you run a cloud business, outsiders no longer see you as an artificial intelligence lab. By this year, he changed his tune. The same compute capacity began collecting rent on a monthly basis — and the ones lining up to pay are its two strongest rivals.
Rivals Queuing Up to Pay Rent
Start with the numbers, all drawn from filings and prospectuses. The Alphabet deal came first. In June 2026, a SpaceX filing stated that Alphabet secured roughly 110,000 NVIDIA GPUs along with supporting CPUs, memory, and components, covering October 2026 through June 2029, at about $920 million per month. The Anthropic deal was larger. At the end of September, its prospectus, read by Reuters, showed that by 2029, Anthropic's compute payments to SpaceX would cap at roughly $84.5 billion — nearly double earlier estimates. There is another deal that has not yet started billing. SpaceX's CFO mentioned at an investment conference in September that a new agreement would begin billing on December 1, 2026, at about $1.11 billion per month, or roughly $13.3 billion annualized, with the client undisclosed. Add these together, and SpaceX's compute business has been reported at an annualized run rate of about $41 billion. In less than half a year, it grew from a business the boss rejected into one worth tens of billions.
Why Now
The key detail is buried in an inconspicuous review. According to The Information, citing people familiar with the matter, SpaceX's own AI cluster had GPU utilization below 40% at one point last year — and it remained low this spring. In other words, a considerable portion of those expensively purchased cards were sitting idle on racks. On one side is idle in-house compute; on the other is an entire industry anxious about not being able to get cards. GPU spot prices have more than doubled, and memory, power, and delivery lead times are all backed up. When compute itself is a scarce commodity, renting it out stops being a matter of losing face and becomes a business with very fast cash flow. According to Musk's recent comments on social media, the company plans to bring roughly 420,000 more NVIDIA GPUs online in November, expanding its Memphis campus — these cards were originally set to go live by the end of October, and reports say the timeline may slip.
What Has Really Changed
Beyond the buzz of this business, one thing is more worth remembering. Alphabet has its own custom TPUs. A major player holding dedicated chips still has to budget for 110,000 GPUs to rent external compute. This shows the center of competition is shifting from whose model is stronger to who can deliver stable compute on time. Models can be caught up to, papers can be reproduced, but racks, power, grid connection permits, and delivery lead times cannot be rushed. Hence an inverted scene: rivals who should be fighting to the death first become lessor and lessee at the compute layer. One research note put it coldly — selling compute is tantamount to renting out the scarcest means of production to a competitor: short-term cash in, long-term ammunition handed to the opponent.
The Policy Says "Refundable"
But this rent is not as stable as it appears. Most contracts carry a clause allowing termination with 90 days' notice. The $84.5 billion is a ceiling, not revenue already booked; the new $1.11 billion order does not start billing until December 1, and the client is unnamed. So the more accurate description is this: it is not locked-in long-term revenue, but a large-scale letter of intent that can be cancelled at any time. The market reads it as bullish, yet one rating agency reached the opposite conclusion — SpaceX's valuation is less than half its IPO target. For a company, turning idle compute into cash flow is smart; but treating rivals paying rent as a moat may overestimate the depth of that trench. Right now cards are scarce; once they are no longer scarce, that 90-day clause will start speaking.
Who Ultimately Foots the Bill
In the end, the thing to watch in this compute business is not who signed how big an order. It is where this money flows. One at $920 million per month, one capped at $84.5 billion, one at $1.11 billion per month — they all go onto AI companies' cost sheets. And over the past year we have seen this transmission chain: compute rises, model companies' bills rise, and ultimately subscription prices climb layer by layer. Enterprise clients can negotiate annual frameworks and get discounts; what is left for ordinary users is often a quietly revised price list. So when watching this spectacle, it is worth asking one more question: when compute becomes hard currency, is the final stop of the bill shareholders' financial statements, or my subscription page? Renting out what is idle is not wrong in itself — it is even rather clever. But the true sign of a technology's maturity is not how much cost it converts into revenue, but that it finally lets users afford it and understand it. AI's accounting is shifting from burning money to counting money. That is a good thing, provided that not every cent counted in ends up silently absorbed by the one who pays. Have you looked at your AI subscription bill lately — how much more is it than six months ago?