“I am highly confident that SpaceX will be launching Nvidia VR NLV72 AI computers in space next year,” Musk wrote on X, per Seeking Alpha. Read that again. Highly confident. Next year. Computers. In orbit.
My first reaction wasn’t awe. It was a question I ask about every AI product that lands in my inbox: what does this actually do for the person trying to get work done today? Because I review AI tools and agents for a living, and I’ve developed an allergy to announcements that sound enormous and change absolutely nothing about my Tuesday.
What was actually said
On its second-quarter 2026 earnings call, SpaceX laid out the plan. An exclusive partnership with Nvidia for AI infrastructure. A push toward $100 billion in annualized revenue run rate by the end of 2026. One trillion dollars by 2030. And within that, $60 billion attributed to AI by 2027.
The market’s response was the most honest part of the whole event. Nvidia gained. SPCX fell. That split tells you who investors think is holding the sure thing and who’s holding the promise.
There’s also a detail in the coverage worth sitting with. Nvidia holds a stake in SpaceX. So we have a chip company investing in a launch company that then announces an exclusive commitment to buy that chip company’s hardware, with a revenue target attached. I’m not saying anything improper is happening. I’m saying the AI industry has been building these loops for two years now, and every time one gets announced, the press release describes it as a partnership rather than what it structurally resembles, which is a customer and a supplier sharing a balance sheet.
About that product name
“Nvidia VR NLV72.” That’s the string in the quote. If you’ve followed Nvidia’s rack-scale systems at all, you know what it’s gesturing at, and you also know that isn’t the name. Maybe it’s a typo on a phone. Maybe it’s a transcription artifact. Either way, the single most-quoted sentence about this partnership contains a garbled product identifier, and it got repeated across financial media without anyone stopping to clean it up.
Small thing. But I mention it because it’s a tidy illustration of how AI infrastructure news travels right now. The number gets checked. The vision gets amplified. The specifics get waved through.
The case for orbital compute, stated fairly
I’m not here to be reflexively sour. The engineering argument for putting data centers in orbit isn’t stupid:
- Solar power is constant and free once you’re above the weather, with no grid interconnect queue to wait in
- Cooling into vacuum sidesteps the water and thermal problems strangling terrestrial builds
- Nobody in orbit files a zoning complaint about your substation
- SpaceX is, uniquely, both the launch provider and the customer, which collapses the biggest cost variable
That last point is the real one. Most companies floating orbital compute concepts have to buy rides. SpaceX builds the rides. If anyone gets to make this claim with a straight face, it’s them.
Where my skepticism lives
Not in the physics. In the revenue line.
Sixty billion dollars from AI by 2027 is a specific number with a specific deadline. For that to happen, somebody has to be paying for something. Training runs? Inference? Downlinked data products? The announcement gives us a dollar figure and a hardware vendor. It does not give us a customer, a pricing model, or a latency story.
And latency is the whole ballgame for the tools I test. Agents are chatty. They loop, they call functions, they retry, they wait on each other. Every round trip matters. Orbital compute is attractive for workloads that are enormous, batch-shaped, and indifferent to when they finish. Training checkpoints. Bulk processing of data that’s already up there. It is deeply unattractive for a coding agent waiting on a tool call.
So the honest framing is this: if orbital infrastructure works, it changes the economics of compute supply, which eventually reaches you as cheaper tokens. It does not change your agent’s architecture. Nobody’s workflow gets faster because a GPU rack is in low Earth orbit. That’s a supply-chain story wearing a science-fiction costume.
What I’d watch instead of the headline
One launch next year proves the thermal and power design survives contact with reality. That’s genuinely interesting and genuinely hard. What it doesn’t prove is that you can operate hundreds of racks up there, service failures you can’t physically reach, and bill customers for the result at a margin that justifies $60 billion.
The gap between “we launched a rack” and “we run a business” is where most infrastructure ambition goes to die. I’ll take the demo seriously when it flies. I’ll take the revenue number seriously when someone names a paying customer.
Until then, treat this like any other tool pitch that arrives with a trillion-dollar figure attached and no pricing page. Interesting engineering. Unproven business. Your agents still run on Earth.
🕒 Published: