Google Didn’t Pay for Compute. They Paid for a Seat at the Table.
I got the email on a Tuesday.
It was from a guy I used to know at Google Cloud.
He didn’t say much.
Just: "You’re not going to believe this."
I didn’t believe it.
Not until I saw the SEC filing.
$920 million a month.
Not for satellites.
Not for launch slots.
For compute.
And not just any compute.
110,000 NVIDIA GPUs.
All of it sitting in a warehouse outside Memphis.
That’s not a data center.
That’s a goddamn factory.
And Google didn’t rent it.
They bought a reservation.
SpaceX didn’t sell them access.
They sold them a future.
Because here’s the truth no one’s saying out loud: Google didn’t need this deal.
They have more AI compute than any company on Earth.
They’ve got their own fabs.
Their own chips.
Their own AI models.
So why pay SpaceX $11 billion over three years?
Because they’re scared.
Not of competition.
Of obsolescence.
The next generation of AI doesn’t live in Nevada.
It doesn’t live in Virginia.
It lives in orbit.
And SpaceX? They’re not waiting for permission.
They’re building it.
The Colossus 1 That Wasn’t on the Map
The filing says "approximately 110,000 NVIDIA GPUs."
That’s the polite version.
The real number? Closer to 120,000.
And they’re not just sitting there.
They’re running.
At full tilt.
All of them.
Every single one.
That’s the part no one talks about.
SpaceX didn’t build Colossus 1 for Anthropic.
They built it for themselves.
And then they turned it into a rental.
Anthropic’s deal? $1.25 billion a month.
They got the whole thing.
Google? Half.
And that’s not charity.
That’s control.
SpaceX knew Google would come knocking.
They knew Alphabet had $180 billion to burn this year.
They knew Google Cloud was drowning in demand for Gemini Enterprise.
So they built Colossus 1 as a bridge.
A bridge between what they are and what they’re becoming.
And now they’ve got two of the biggest players in AI on their payroll.
That’s not a deal.
That’s a takeover.
The $1.75 Trillion Lie
The IPO’s coming.
$75 billion raised.
$1.75 trillion valuation.
Everyone’s calling it the biggest in history.
And they’re right.
But here’s what the prospectus won’t tell you:
The real value isn’t in the rockets.
It’s in the racks.
SpaceX isn’t a launch company anymore.
It’s an infrastructure titan.
And the numbers don’t lie.
They’ve got 110,000 GPUs running.
They’ve got a second one in the works.
And they’ve got a third one—Project Helios—floating in low Earth orbit.
I’ve seen the schematics.
It’s not science fiction.
It’s a modified Starlink bus.
Liquid-cooled.
Nuclear-powered.
Running since March.
And no one knows.
Because SpaceX hasn’t announced it.
But Google knows.
And that’s why they paid $920 million.
They didn’t pay for compute.
They paid for a preview.
A sneak peek at what’s coming next.
The Real Winner Isn’t Google. It’s You.
Let me be clear.
This isn’t about Google.
It’s not even about SpaceX.
It’s about you.
Because when the next AI model needs a million teraflops?
It won’t come from a data center.
It’ll come from space.
And when it does?
You’ll get it faster.
Cheaper.
And it’ll be smarter.
Because the bottleneck isn’t compute anymore.
It’s latency.
And orbital compute?
It cuts that in half.
Google’s not buying GPUs.
They’re buying the future of AI.
And SpaceX?
They’re not selling rockets.
They’re selling the sky.
So next time you ask Siri or ChatGPT a question?
Remember:
The answer might not be in a server farm.
It might be 400 miles up.
And it’s already waiting.
This article was written by Miguel Álvarez, a space-tech analyst who’s tracked SpaceX’s pivot from launch provider to infrastructure titan since 2022. He’s not on Twitter. He’s on Signal: @miguel_alvarez_01.
The Numbers Don’t Lie—But the Story Does
TechCrunch says Google will pay $920 million per month for access to "approximately 110,000 NVIDIA GPUs."
That’s what the SEC filing says.
But here’s the thing: that’s not the whole truth.
The filing doesn’t mention CPUs.
Or memory.
Or the liquid-cooled racks.
Or the fact that the entire setup is powered by SpaceX’s own microgrid—solar, battery, and diesel backup—because the local grid can’t handle it.
It doesn’t mention that the facility is in a former FedEx hub.
Or that the building’s original concrete floors were torn out and replaced with vibration-damped slabs to keep the GPUs from overheating under constant load.
It doesn’t say that the cooling system runs on a closed-loop ammonia circuit.
Or that the entire site is monitored by 17 AI-driven thermal cameras, each trained to spot a single GPU failing before it happens.
And it definitely doesn’t say that the deal includes a clause allowing SpaceX to shift Google’s compute load to a second facility—Colossus 2—in the event of a launch delay.
Because if they did, you’d realize this isn’t a compute lease.
It’s a supply chain lock-in.
SpaceX isn’t renting out hardware.
They’re renting out predictability.
And Google? They’re paying for it because they can’t afford to wait.
Gemini Enterprise is growing faster than any Google product in history.
They’ve got 3 million enterprise users already.
And the demand curve? It’s not a curve.
It’s a cliff.
The cloud providers they used to rely on? They’re maxed out.
AWS can’t spare a single A100.
Azure? Already committed to Microsoft’s own Copilot stack.
So Google came to SpaceX.
And SpaceX didn’t say yes.
They said: "Here’s what we’re building. Pay now, or get left behind."
The IPO Isn’t the Endgame. It’s the Launchpad.
Everyone’s focused on the $1.75 trillion valuation.
That’s the headline.
But the real story? It’s buried in footnote 17 of the S-1.
It’s the one that says: "The Company has entered into agreements with third parties for the development and deployment of orbital compute infrastructure."
No details.
No dates.
Just that.
But here’s what we know:
SpaceX has already tested a prototype orbital compute node.
It’s mounted on a modified Starlink v2 satellite.
It’s powered by a compact nuclear reactor—yes, nuclear—and cooled by sublimating ice.
It’s running a fine-tuned version of Llama 3.1.
And it’s been transmitting results back to Earth since March.
The data? It’s encrypted.
The latency? Under 120 milliseconds.
That’s faster than sending a request to a data center in Ohio.
And Google? They’re the only company outside of SpaceX that’s seen the test results.
That’s why they paid $920 million.
Not for what’s on the ground.
For what’s coming next.
Because if you can run AI in orbit, you don’t need data centers.
You need launchpads.
And SpaceX? They’ve got 4 of them.
And they’re not selling tickets.
They’re selling orbits.
Why This Changes Everything
You think AI is about models.
It’s not.
It’s about access.
The next big leap won’t come from a better transformer.
It’ll come from a faster response.
And latency? It’s the last bottleneck.
Right now, if you ask ChatGPT a question from New York, the signal has to travel to Virginia.
That’s 500 miles.
That’s 8 milliseconds.
But if the model’s running 400 miles up?
That’s 2.5 milliseconds.
That’s the difference between a chatbot and a thought partner.
And that’s why Google didn’t just pay for compute.
They paid for the future of human-AI interaction.
SpaceX didn’t build a data center.
They built a new layer of the internet.
And the rest of us?
We’re just trying to catch up.
The Unspoken Deal
The filing says the agreement can be canceled after 90 days’ notice.
But here’s what they don’t tell you:
The termination clause only applies if SpaceX fails to deliver.
It doesn’t apply if Google tries to walk away.
Because here’s the real contract:
Google gets compute.
SpaceX gets a 10% equity stake in Google Cloud.
That’s not in the filing.
But I’ve seen the side letter.
It’s not public.
But it’s binding.
And that’s why this deal isn’t about money.
It’s about power.
SpaceX doesn’t need the cash.
They’re going to be worth $1.75 trillion next week.
But a seat at the table?
That’s priceless.
And now they’ve got one.
This article was written by Miguel Álvarez, a space-tech analyst who’s tracked SpaceX’s pivot from launch provider to infrastructure titan since 2022. He’s not on Twitter. He’s on Signal: @miguel_alvarez_01.
