Google Is Renting GPUs From a Rocket Company for $920 Million a Month

This image was generated with help from an AI program

Read that sentence again. The largest single owner of AI compute on Earth just had to rent emergency capacity from SpaceX, and the money is moving in a circle. After a week of writing about agents everywhere, this is the story about the ground they all stand on, and how little of it there is.

For seven days I’ve been writing about the AI agent boom from every angle I could reach. A chip that runs agents on your desk. Software that governs them. A breach that showed how they fail. The billions raised to fund them. A cartoon app that makes us comfortable with them. And the tollbooth Apple built to tax them. Six stories about a revolution arriving in real time.

And then, on Friday, a regulatory filing landed that quietly reframes all six. Because it answers the question I’d been circling without quite naming: underneath all these agents, the chips, the channels, the apps, the capital, what’s the thing that’s actually scarce?

The answer is compute. And here’s how I know it’s the real bottleneck: Google just agreed to pay SpaceX $920 million a month to rent it.

Sit with that. Google. Renting GPUs. From a rocket company. To the tune of nearly a billion dollars a month.

What the filing actually says

Per a SEC filing on Friday, Google will pay SpaceX roughly $920 million per month, from October 2026 through mid-2029, for access to about 110,000 NVIDIA GPUs plus the surrounding CPUs, memory, and infrastructure. Google was candid about why: surging, faster-than-expected demand for its Gemini Enterprise agent platform, requiring what it framed as short-term bridge capacity. The companies are long-time partners; Google called it a timely stopgap.

This story was written with assistance from an AI program.

It’s the second monster deal SpaceX has struck in weeks. In late May, Anthropic agreed to pay SpaceX about $1.25 billion per month through 2029 to rent essentially all the available compute at the Colossus 1 data center near Memphis, the facility xAI originally built for its own AI before xAI became part of SpaceX. Google’s deal is roughly half that scale.

Both contracts come loaded with escape hatches, either side can walk with 90 days’ notice after the end of 2026, and Google’s terms ramp up gradually with reduced fees and penalty clauses if SpaceX can’t deliver the promised GPUs on schedule. And the timing is the loudest part: SpaceX is expected to begin trading on the Nasdaq within a week, targeting a raise of around $75 billion at a roughly $1.75 trillion valuation, the largest IPO in history.

Insight #1: If Google is short on compute, compute is the whole game

Here’s the detail that should stop you. By some estimates, Google is the single largest owner of AI compute on the planet, it designs its own custom chips and operates data centers at a scale almost no one can match. This is not a startup scrounging for GPUs. This is the most compute-rich company in existence.

And it still had to rent.

When the company with the most compute on Earth has to lease emergency capacity from a rocket company, “compute shortage” stops being a phrase and becomes the defining constraint of the entire industry.

Everything I wrote about this week, agents on your desk, in your inbox, in iMessage, reading your Google account, sits on top of a finite pile of silicon and the power to run it. The agents are multiplying faster than the substrate beneath them can grow. Demand for Gemini Enterprise outran Google’s own staggering capacity. That’s the real story of the AI boom in mid-2026: it isn’t gated by ideas, or models, or even money. It’s gated by physical access to GPUs and the electricity to feed them. Compute is the oil of this era, and we are in a shortage.

For everyone downstream, and that includes me, and probably you, this validates a worry I raised earlier in the week. When the giants are fighting over capacity at a billion dollars a month, that scarcity flows downhill. It shows up as rate limits, as capacity waitlists, as price changes. Recall that Anthropic was meaningfully constrained until it locked in its SpaceX deal, it raised usage limits the very day it was announced. If you’re building on AI, this is your warning that cheap, unlimited inference is not a law of nature. It’s a temporary condition that the people above you are paying enormous sums to secure for themselves first.

Insight #2: The money is moving in a circle, and you should look at it directly

Now follow the dollars, because the structure is remarkable, and a little vertiginous.

Google is paying SpaceX ~$920 million a month. Google is also a long-time investor in SpaceX, with a stake reportedly worth more than $100 billion after the IPO. So Google’s compute payments flow into a company Google partly owns, inflating the value of Google’s own investment. SpaceX, meanwhile, acquired xAI, whose data centers were built for xAI’s own AI ambitions, and is now renting that compute out to xAI’s direct rivals, Google and Anthropic. Anthropic, paying $1.25 billion a month, is itself filing to go public. NVIDIA sells the GPUs that sit at the center of every one of these deals. And all of it is being announced in the days and weeks before the two largest tech IPOs ever attempted.

I’m not going to tell you this is a bubble, and I’m not a financial advisor, none of this is investment advice. But I will tell you what it is structurally: a small cluster of giants paying each other enormous, recurring sums, where many of the same players are simultaneously each other’s customers, suppliers, and shareholders, and where these contracts conveniently de-risk historic IPOs by demonstrating locked-in revenue right before the shares price.

That can be completely healthy. Locked-in compute contracts are exactly what you’d expect from real, surging demand, and Google’s “we ran out of capacity” explanation is entirely plausible given the agent gold rush. But it can also be the kind of self-reinforcing loop that looks like unstoppable momentum right up until the demand assumption underneath it wobbles. When I warned earlier this week that some AI revenue might turn out to be companies paying each other in a circle, this is the concrete shape of that concern. I’m not predicting a reckoning. I’m saying: look at the circle clearly, and price your own decisions with your eyes open.

Insight #3: Compute is so scarce the next data center might be in space

And then there’s the kicker, the detail that tells you exactly how severe this crunch is. Google and SpaceX are reportedly in talks to build data centers in orbit.

When the proposed solution to your supply problem is leaving the planet, you are not dealing with a minor capacity hiccup. Terrestrial AI compute is colliding with hard physical limits: power, cooling, land, grid connections. The fact that orbital data centers are a serious boardroom conversation, rather than a science-fiction punchline, is the clearest possible signal of how hard the ceiling is being pushed. SpaceX’s post-IPO future, it turns out, may be less about Mars and more about becoming the landlord of compute in low Earth orbit.

What I’d take from this if you build things

The practical lessons compound everything from the week:

Treat compute access as a real risk, not an assumption. If Google can be capacity-constrained, your provider can be too. Build with fallbacks. Don’t architect a business whose unit economics assume infinite, cheap, always-available inference, because the people who own the GPUs are spending billions to make sure they get them first.

Expect the cost floor to rise, not fall. A shortage with a near-trillion-dollar capex race on top of it does not point toward cheaper tokens. Price your product with margin to absorb a meter that’s pointed up.

The on-device option just got more interesting. I opened the week with NVIDIA’s RTX Spark and treated local agents as a promising-but-unproven curiosity. In light of a global compute crunch, running models on hardware you own stops looking like a novelty and starts looking like a hedge: against rate limits, against price hikes, against being last in line behind the giants. Owning your compute, even modestly, is a form of insurance now.

The week, from the ground up

Step back and the whole arc inverts into focus. NVIDIA showed what agents do. Microsoft showed how we’ll govern them. Meta showed how they break. Alphabet showed who’s funding them. Google’s Dreambeans showed how we’ll be made comfortable with them. Apple showed who controls the door. And now SpaceX and Google have shown the thing underneath all of it: the ground, a finite, fiercely contested pile of GPUs and power that even the richest company in compute can’t get enough of, paid for in a circle of intertwined billions, with the next expansion frontier pointed at the sky.

The agent revolution is real. But this week ended by revealing its foundation, and the foundation is narrower and stranger than the keynotes suggest: a handful of giants renting silicon to each other at a billion dollars a month, racing physics to the edge of the atmosphere, and quietly deciding, long before any of it reaches you, who gets to compute at all.

Watch the compute. It’s the constraint that everything else in this story is secretly about.

If you build on AI: are you planning for compute scarcity and rising costs, or assuming the spigot stays open and cheap? And when you trace the money in deals like this, do you see healthy demand or a circle worth worrying about? I’d genuinely like to hear how you read it. (Still not financial advice, just the conversation of the year.)