3-point summary: (1) Elon Musk announced SpaceX will use Nvidia GPUs exclusively — no other brand at all (2) The reason given is short and blunt — “because it’s the best,” no caveats attached (3) The plan is to send the (optimized) Vera Rubin NVL72 into orbit next year
Worth stressing upfront: this is a statement + plan, not a product that’s actually shipping right now. The Vera Rubin NVL72 is an Nvidia data-center-tier product line whose full pricing and specs haven’t been made public yet.
To get a sense of what tier of GPU Nvidia can build, look at the consumer-side RTX 5060 as a reference point: the GB206 chip on a 5nm process, 3,840 cores, 8GB GDDR7 VRAM, 448 GB/s bandwidth — and that’s just Nvidia’s entry-level tier. The class of hardware headed to space is an entirely different league.
What does the Vera Rubin NVL72 actually look like
First things first: Nvidia hasn’t disclosed the full specs of the Vera Rubin NVL72 publicly. Most of what’s circulating in the news is render/concept imagery of the successor to the GB200 NVL72 — nothing more.
For scale, compare against the RTX 5060, whose specs are fully public: the standalone GB206 chip packs 21.9 billion transistors on a 181 mm² die, built on TSMC’s 5nm process. If the Vera Rubin NVL72 is a rack linking dozens of GPUs together in an NVL72 configuration, it’s reasonable to assume the hardware complexity and cooling requirements are far greater than that — which is exactly why there’s no “real” photo to show yet.
The problem that pushed Musk to bet his entire GPU fleet on one vendor
The real problem for AI-era data centers isn’t just “buying enough GPUs” — it’s that power and cooling capacity on Earth can’t keep pace with AI’s growth. Musk has already talked about power constraints before, both for xAI’s training cluster and for Starlink, which keeps having to scale up.
As each GPU generation’s TDP climbs higher (even a consumer-tier card like the RTX 5060 draws 145W on its own), stacking them into data-center-scale racks means the accumulated heat becomes enormous. Cooling on Earth depends on power and water, both of which have clear physical limits.
This is where the idea of “moving compute into space” starts to make sense — space offers effectively unlimited solar energy, and the cold vacuum aids heat rejection in a way that’s fundamentally different from anything on the ground.
The Vera Rubin NVL72 is the generation after Blackwell on Nvidia’s roadmap, and it’s the one Musk has chosen to tie the future of his entire compute empire to — from xAI’s Colossus cluster to Starlink, which has to process a staggering volume of satellite data.
What’s notable is that Tesla has long pursued its own Dojo chip, but now appears to be leaning fully on Nvidia instead. The “exclusive” announcement, then, isn’t just marketing — it’s Musk conceding that racing to build in-house silicon is slower than simply buying the best thing already on the market.
Sending the NVL72 into space isn’t a standalone experiment, then — it’s the next step in a strategy already in motion: the same chip on Earth and off it, collapsing the entire Musk-ecosystem’s complexity down to a single platform.
Comparing specs: Vera Rubin NVL72 vs. the previous-gen Grace Blackwell NVL72
Nvidia hasn’t disclosed complete real-world specs for the Vera Rubin NVL72 yet. What can be said right now is the direction of the upgrade from the existing Grace Blackwell generation — not definitive numbers.
| Factor | Vera Rubin NVL72 | Grace Blackwell NVL72 |
|---|---|---|
| Chip generation | Newer (next-gen) | Blackwell |
| NVLink bandwidth | Higher, per Nvidia's announcement | Baseline generation |
| Power draw per rack | Official figures not yet released | Official figures not yet released |
| Cooling system | Must be redesigned for space use | Designed for Earth-based data centers |
The key point is that a standard data-center cooling system simply can’t work in space — this is a spot that requires a full re-engineering, not just bolting the existing hardware onto a rocket.
The advertised features vs. the reality of actually using them
The NVL72’s high-speed NVLink is designed to let GPUs talk to each other fast within a massive cluster like the Colossus setup xAI uses to train models on Earth — but move that into orbit and the problem changes completely, because power now has to come from solar panels instead of an effectively unlimited grid connection like a data center gets.
The liquid cooling systems used on Earth rely on air to help carry heat away from the system, but there’s no air in a vacuum to convect heat at all — everything has to rely on radiators dissipating heat purely by radiation, which is an entirely different engineering discipline.
Compute density per rack also has real consequences for how many satellites or launches are needed — the more compute you can pack per unit of weight, the fewer launch cycles required.
But all of this is still just direction Musk has talked about. Actual figures for the Vera Rubin NVL72 in orbit haven’t been officially disclosed.
If not Nvidia, what alternatives could even compete
Honestly, in today’s AI accelerator market the only options that come close are AMD’s Instinct MI400 and Google’s TPU v7 (or AWS’s Trainium3) — but both remain largely closed ecosystems tied mainly to their own cloud platforms, not sold as openly as Nvidia’s chips.
So when Musk says “the best,” it’s not just about raw performance — it also covers supply-chain readiness and the CUDA software stack that the SpaceX/xAI teams are already deeply familiar with.
| Factor | Nvidia (Vera Rubin) | AMD / Google TPU |
|---|---|---|
| Ecosystem | Open, CUDA dominates the market | Tied to the owner's cloud |
| Supply readiness | Nvidia has announced a clear roadmap for next year | No comparable public data yet |
| Single-vendor risk | High (single vendor) | More diversified |
Actual spec data for the MI400 and TPU v7 hasn’t been officially disclosed at a level comparable to the Vera Rubin, so it’s still not possible to make a definitive call on performance-per-watt right now.
The pros and cons of going all-in on Nvidia alone
Musk’s announcement that he’ll use Nvidia “because it’s the best” sounds simple, but underneath it is a decision to tie the entire future of SpaceX’s compute to a single supplier. If Nvidia’s deliveries slip or prices rise, SpaceX has essentially no backup option — unlike Google, which has its own TPUs to fall back on.
Pros
- +Gets market-leading hardware performance with no real competitor to compare against
- +A deep partnership relationship at the co-design level, as seen with the Vera Rubin NVL72
- +Reduces integration-team complexity since it's all a single stack
Cons
- −Single-vendor lock-in — high risk if Nvidia hits supply problems or delays
- −Almost no pricing leverage, since there's no alternative to counterbalance it
- −The space plan depends entirely on Nvidia's roadmap alone
The cost nobody’s talking about: lifting a data center into orbit
The price of a GPU on Earth is nowhere near the real cost of the same hardware in orbit — launch cost per kilogram is the biggest variable of all. And since the Vera Rubin NVL72 is a full-cabinet rack system, the total weight including wiring and cooling infrastructure can’t be light.
The next problem is cooling: on Earth you use fans plus water, but in a vacuum there’s no air to carry heat away, so the entire radiator system has to be redesigned from scratch. On top of that, you need cosmic-radiation shielding to protect the chips from damage, plus solar panels and backup batteries that are far heavier and more expensive than a power system on Earth.
The heaviest problem of all: it can’t be repaired. A chip that fails in orbit is gone for good — there’s no technician who can go up and swap out a card the way you would in a normal data center.
As for Musk’s “next year” timeline — that’s something to verify with your own eyes before believing it. Starship itself has already slipped multiple times.
What to watch next — who follows suit, and what has to be proven before next year
Google and Amazon already have their own chips (TPU, Trainium), which would make it easier for them to get into space-based data centers than Microsoft, which — like SpaceX — depends on Nvidia. It remains to be seen whether they’ll announce parallel plans of their own.
The checklist that needs to actually happen before the “next year” claim can be believed: repeated successful launches of heavy payloads into orbit, a radiation-shielding prototype that’s been proven in real conditions rather than just in a lab, and a cooling system that actually works without any air to assist it.
More important than any of that is whether Nvidia itself is ready to ship a GPU generation certified for use beyond Earth — because specs designed for ground-based data centers and specs that can withstand cosmic radiation are two completely different things.