Google will pay SpaceX $920M a month to rent GPUs sitting in xAI's data centers
An SEC filing reveals a triangular compute deal: Google renting capacity from SpaceX, which is renting it from xAI's Colossus build-out — at a rate one engineer calculated to $11.61 per GPU-hour, roughly double the going price.
A compute deal with three names on it and an IPO at the end

CNBC reported, citing an SEC filing, that Google will pay SpaceX roughly $920 million a month for GPU compute capacity tied to xAI's data centers — the Colossus super-cluster Elon Musk's companies have been racing to stand up. The structure is the story: Google, which already profited enormously from an early SpaceX stake, is now routing compute spend back through SpaceX just as SpaceX lines up an IPO reportedly valued over $1.75 trillion.
The filing's hard terms — delivery deadline and a 90-day exit
The thread surfaced the load-bearing clauses straight from the filing. SpaceX must deliver, or the deal unwinds:
"if it fails to 'deliver access to the committed amount of GPUs by September 30, 2026,' Google can immediately end the agreement, or accept the number of GPUs provided at a reduced fee after a one-month grace period. … After this year, the agreement can be terminated by either party provided they give 90 days' notice." — a commenter quoting the SEC filing directly
At ~$920M/month, the run-rate is roughly $11 billion a year flowing to SpaceX for hardware that physically lives in xAI's facilities. One commenter noted xAI's total GPU-rental revenue may now be around $2.1B/month, raising the question of whether xAI — and by extension SpaceX's compute arm — is suddenly profitable.
The thread: the per-GPU math doesn't sit right
The most concrete skepticism was arithmetic, not vibes:
"920 million a month for 110,000 GPUs is $11.61 per GPU hour. That seems very high to me." — a commenter who ran the per-GPU-hour math
That's roughly double typical market H100/H200 rental rates — which fed a louder suspicion about why a company building its own data centers would rent at a premium:
"Circular financing at its peak for the IPO. There has to be some regulatory body to not allow such shady things" — a commenter alleging circular financing ahead of the IPO
"I guess their training runs aren't going great if they're dumping all the compute they can. Or I guess juicing the numbers for IPO" — a commenter on what dumping spare compute implies
A datacenter-networking expert asked the question the filing invites but doesn't answer:
"With ~$2.1B/month of GPU rental revenue, is xAI now profitable? Are all divisions of SpaceX now profitable?" — Wes Felter, datacenter and cloud-networking expert, ex-IBM Research
And the inversion that reframes the whole thing — read it from xAI's side:
"Another way to read this is, 'Twitter's AI dept. has been doing so badly they don't have a use for billions of dollars of hardware they bought.'" — a commenter reframing it as offloading idle hardware
Shovels, not gold
The cleanest summary in the thread came as a one-liner: "the gold rush has entered a new phase as miners pivot to selling shovels," wrote a commenter on the shovel-selling phase of the AI boom. SpaceX — a rocket company — is now a landlord for the AI build-out, billing Google $11B a year for GPUs it doesn't even house. Whether that's a genuine capacity crunch at Google or a tidy way to pump revenue into a $1.75-trillion IPO is exactly the question the filing's clean termination clauses can't dispel. The September 30 delivery deadline is the tell to watch: if the GPUs are real, they ship on time.
Get breakdowns like this every day
The top Hacker News stories, unpacked — by email, Telegram, or RSS.
Ladybird stops taking pull requests — because a big patch no longer proves effort
The independent browser project will close all open public PRs and let only maintainers commit code. The reason in one sentence: a substantial patch used to imply substantial effort, and that proxy for good faith just broke.