NVIDIA's release formalizes the NVL72 basis Musk disclosed August 4. That evening, SpaceX's two public hosts served different AI1 spec sheets.
NVIDIA's August 24 release with SpaceXAI covers three things: Vera CPUs for Grok's agentic workloads, Vera Rubin as the common architecture as SpaceXAI "expands toward gigawatts of computing capacity," and the statement that the company's "planned first-generation Starmind AI satellite will be based on the optimized NVIDIA Vera Rubin NVL72 rack-scale system."
The orbital line is confirmation, not disclosure. Elon Musk had already told the August 4 earnings call that Starmind "will be essentially an optimized Vera Rubin NVL72 computer," and that SpaceX expects "to start launching these next year." What August 24 adds is NVIDIA formally attaching its own name to the plan, with one new sentence of substance: the companies are "working to adapt that foundation to the requirements of orbital computing while preserving a common NVIDIA architecture and software ecosystem." The release gives no launch date, satellite count, or orbital capacity figure.
It also lands on an awkward evening. As of August 24, SpaceX's two public web hosts serve materially different spec sheets for AI1, the satellite design on the company's Starmind pages. The page at www.spacex.com, in a build stamped August 11, describes a spacecraft with a compute payload of up to 250 kW peak and 175 kW average, 30 m deployed height, a 75 m wingspan, a 160 m² deployable liquid radiator, a 210 kW solar array, and 75 kW per ton.
The page at new.spacex.com, in a build stamped that same evening, gives 150 kW peak and 120 kW average, 20 m by 70 m, a 110 m² radiator (the fresher build carries a "RADIATIOR" typo), a 150 kW array, and 70 kW per ton. SCN retrieved both on the evening of August 24.
Determining which build aligns with current engineering plans remains impossible externally; a timestamp confirms build assembly, not authoritative status. Both sets are company claims with no independent verification, and this article names which envelope any derived number uses.
SpaceXAI, for orientation, is the AI division of the merged SpaceX: the company announced the xAI acquisition on February 2, at a combined valuation reported around $1.25 trillion, and listed on Nasdaq in June as SPCX.
On the ground, a Vera Rubin NVL72 combines 72 Rubin GPUs and 36 Vera CPUs into a single NVLink domain, a rack-scale supercomputer NVIDIA says is in full production, with partner products arriving in the second half of 2026. Supermicro's deployment blueprint provisions four 110 kW power shelves per rack and offers a 200 kW liquid-to-air option for a single rack in facilities without liquid infrastructure. A 200 kW-class machine, in other words, and one that does not strictly require facility water to deploy.
What "based on" that machine establishes is architecture. Flight configuration is a separate question, and it is open. Musk called the orbital design "a radical simplification of the normal NVL72 rack" on the same call, and said SpaceX expects to deploy the simplified design on the ground as well as in orbit. NVIDIA's language is an adaptation of a foundation. Between those two phrases sits everything undisclosed: how many GPUs a satellite carries, at what clocks, in what NVLink topology, at what power. Neither company has published any of it.
The terrestrial half of the announcement is more concrete. The Vera CPU is built for the orchestration side of agentic inference: 88 custom Olympus cores running 176 threads, up to 1.8 TB/s of coherent NVLink-C2C bandwidth to the GPUs, and, separately, up to 1.2 TB/s of LPDDR5X memory bandwidth against up to 1.5 TB of memory. "Vera gives us the CPU performance and memory bandwidth to run enormous amounts of orchestration, code, and data processing while keeping GPUs doing what they do best," said Mike Nicolls, SpaceXAI's president, in the release. NVIDIA claims up to 1.8x faster task completion than x86 CPUs; it names no baseline platform or methodology. The deployment context: SpaceX's CFO told the August 4 call the company ended June with 1.4 GW of nameplate compute, Musk expects "over 2 gigawatts" by year-end, and his 2027 figure came hedged: cumulative compute that "may, let's say, be closer to 10 gigawatts of compute than 5 gigawatts." That ramp is the terrestrial story SCN has been tracking since Rubin's pricing became an event in hyperscalers' 2027 budget planning.
Set the dated record side by side, and AI1 is a moving target. At a March 21 event in Austin, Musk showed an "AI Sat Mini" at 100 kW with roughly 100 m² of radiator. By mid-July, the Starmind site listed 150 kW peak and a 110 m² radiator. A capture checked August 9 recorded the larger 250 kW / 160 m² envelope; that is the sheet the www host, in its August 11 build, still served when SCN retrieved it on August 24. And the new host's August 24 build carries the smaller July-matching envelope again.
SCN is not inferring a rollback, a staging accident, a stale fork redeployed, or a mid-revision redesign; the public record does not support any of those readings over the others. What the record does establish is narrower and stranger: SpaceX's published AI1 envelope has changed repeatedly inside five months, and on the day NVIDIA formalized the satellite's compute architecture, the company's own pages disagreed about the satellite's power, size, and radiator area, with peak compute alone differing by two-thirds.
For a program whose FCC application contemplates up to 1,000,000 satellites, design fluidity this late in the public story is information in itself.
Orbit offers no useful external convective heat sink; whatever the pumped loops do inside the spacecraft, steady-state waste heat ultimately leaves by radiation. Both spec sheets pair their compute payload with a deployable liquid radiator, and the Starmind pages claim the design cuts cooling power overhead "by an order of magnitude" against terrestrial data centers. That is a company claim; here is what the published numbers let anyone check.
Run the two envelopes through the Stefan-Boltzmann law, as SCN's arithmetic, at an assumed emissivity of 0.9 and with average payload as the steady load. The two envelopes converge: 175 kW across 160 m² and 120 kW across 110 m² both work out to almost exactly 1.09 kW per square meter of quoted radiator area. If the quoted area is the total emitting surface, that flux requires an idealized radiator temperature of about 383 K (about 110 °C). If the quoted area is panel planform and both faces radiate, the effective area doubles and the requirement drops to roughly 322 K (about 49 °C). One undefined convention, a swing of roughly 60 °C in the radiator temperature the design must sustain. And the idealization is generous: it ignores solar loading, Earth infrared and albedo, view factors between panels, the temperature drop from coolant to radiating surface, fin efficiency, and end-of-life degradation. The underlying trade is not unique to SpaceX; SCN recently examined a patent that spends compute density to buy radiator margin because the arithmetic pushes orbital designs in the same direction.
None of this shows the radiator claim failing. It shows that SpaceX has published enough to reveal the scale of the thermal problem, and not enough for anyone outside the company to close the thermal budget. The missing variables- emitting-area convention and radiator operating temperature above all- decide whether the claim is extraordinary or merely aggressive. Musk's only on-record response to radiator questions, from March, is: "For some reason there's been a bizarre debate about radiators in space... SpaceX knows how to do heat rejection in space with 10,000 satellites in orbit" (SpaceNews). That is confidence, not a thermal design. SCN's survey of orbital-compute hardware, physics, and economics is the context for how far every flown system sits from either envelope.
Power balance. On the www envelope, peak compute (250 kW) exceeds the solar array's rating (210 kW). That proves only that peak could not be sustained from contemporaneous array output if the two ratings are comparable; batteries, duty cycling, load curtailment, or simply different rating conventions could each account for it, and SpaceX has published nothing on storage. On the new envelope, the question dissolves: peak compute and array rating are both 150 kW.
Radiation. Neither the new announcement nor NVIDIA's product materials address qualifying Rubin silicon for orbit. The nearest public datapoint comes from Google's Suncatcher work: Trillium TPUs under a 67 MeV proton beam showed no hard failures up to the maximum tested dose of 15 krad(Si), while the high-bandwidth memory subsystem proved the most sensitive component, showing irregularities from a cumulative 2 krad(Si). Google notes that it is still roughly three times the shielded five-year mission dose it expects. The relevance to Starmind is direct, because Rubin's compute stack is built on HBM4 memory, and the filed 500-2,000 km shells span meaningfully different radiation environments. SCN found no published SEE or total-dose characterization for Rubin-class parts.
Packaging. A four-GPU reference module displayed alongside the release (StorageReview) plausibly hints at how a rack becomes a payload, but neither company has presented it as Starmind's design. NVIDIA's Space-1 Vera Rubin module, announced in March with Aetherflux named as its deployer, shows the company is packaging Rubin for mass- and power-constrained missions; NVIDIA has not said Starmind uses it, and the module itself is listed as available "at a later date."
The Starmind pages state, on both hosts: "AI chip vendor agnostic. Our system architecture supports compute modules from any provider." Musk, on August 4: "we've decided to build exclusively on NVIDIA because we think the Vera Rubin architecture is the best architecture." Both statements are on the record, and SCN is not resolving them; the open questions are whether the exclusivity is contractual or merely current, and what workloads it covers; the company has not said. The stated long game runs through SpaceX's own silicon: the site names Terafab, the SpaceX and Tesla fab project Musk pitched in March at a terawatt of processors per year alongside a space-optimized D3 chip (SpaceNews).
NVIDIA, meanwhile, is placed on more than one side of the orbital bet. Three days before the release, it put $25 million into Starcloud's $250 million extension round at a $2.3 billion valuation; Starcloud already operates an H100-class satellite, holds an 88,000-satellite FCC filing of its own, and hopes to fly Rubin-class hardware in late 2028. Against the field SCN mapped in the gap between filed constellations and flown hardware, no Rubin-class silicon has flown yet.
The reason orbit is on the table at all is the standing power question: whether energy infrastructure for the next generation of supercomputing can be built fast enough, a constraint SCN has tracked as NVIDIA's order book collides with grid capacity. The independent numbers stay unfriendly. SemiAnalysis models a levelized cost of compute of $10.91 per GPU-hour in space against $2.49 terrestrial once reliability is priced in ($8.64 versus $2.37 on raw TCO), with parity around 2040 in its base case and near-parity in the early 2030s only if terrestrial capacity stays constrained. The model hinges on Starship reaching about $250 per kilogram; Falcon 9 flies at $1,400-1,800 per kilogram today. Their sharpest objection: accelerator supply is bottlenecked by chip production, "a problem that Space Datacenters cannot solve." The launch arithmetic, SCN's own and envelope-dependent: at the www sheet's 75 kW per ton, a gigawatt in orbit is roughly 13,300 tons; at the new. sheet's 70 kW per ton, closer to 14,300. Assuming approximately 100 tons per Starship flight, that is 130 to 145 fully loaded flights per gigawatt, before stowed volume, which neither sheet discloses, gets a vote.
The regulatory posture is early-stage. SpaceX applied on January 30 for up to 1,000,000 satellites across 500-2,000 km shells; the FCC accepted the application for filing on February 4 and sought comment. Acceptance for filing is not approval. Astronomy has produced the most quantified objection: ESO's July study of the more than 1.7 million satellites now proposed across all constellation filings concluded they would have "devastating consequences for astronomy", modeling field-of-view losses of up to 28% in Very Large Telescope images taken two hours into the night, and concluding that no more than 100,000 faint satellites should orbit Earth. That figure is a study threshold; no regulator has adopted or proposed it.
The stated timeline is short. Musk says launches start "next year"; the www Starmind page says Gigasat factory output supports "thousands of AI satellites starting as soon as late 2027." Between now and any of that, four things would move this story from spec sheets to evidence: SpaceX's two hosts converging on one AI1 envelope; a disclosed GPU count, clock target, or radiator operating temperature that lets outsiders close the thermal budget; movement on the FCC docket as comments come in; and the first radiation and thermal data from flown Rubin-class hardware, whoever flies it first.
The SCN Staff is a small AI editorial squad working under human direction. Each agent owns one job.
Scout does the research. It runs down primary sources and checks what's already been published, on SCN and everywhere else, before a story gets written. If a claim can't be traced back to a real document, Scout flags it.
Forge writes. It takes what Scout found and turns it into a draft, argument and sentences and all. Every SCN piece starts here, then gets sharpened.
Cipher handles search: the titles, descriptions, and keyphrase work that decides whether a good article ever gets found. Least glamorous job on the squad. Also one that matters more than it looks.
Pixel makes the visuals. Images, charts, the occasional diagram, all built to SCN's brand instead of pulled from a stock library. When something's easier to see than to read, it goes to Pixel.
Editorial judgment and the final call stay with the humans. So does the fact-checking.
