SpaceX Goes All-In on Nvidia for AI, From Data Centers to Orbit
SpaceX's public commitment to use only Nvidia chips for all ground and orbital AI infrastructure — rejecting AMD and unveiling the Starmind satellite data-center program — represents the most consequential single demand signal Nvidia has received since hyperscalers began their AI buildout.
When Elon Musk declared this week that SpaceX would use 'only Nvidia chips — because they are the best,' the market responded with immediate precision: AMD shares fell roughly eight to nine percent on what was otherwise a record quarterly result, while Nvidia gained 3.4 percent to $219.22. The announcement is notable not merely as a procurement endorsement but as a statement about where the center of gravity in AI compute has settled — at least for the industry's most ambitious infrastructure builders.
The specifics, as reported across multiple outlets, are striking in scale. SpaceX has committed to Nvidia's H200, B200, and forthcoming Rubin GPU families for both its terrestrial data centers and the newly announced 'Starmind' orbital AI data centers — satellites equipped with Nvidia Rubin GPUs and Vera CPUs designed to deliver data-center-grade AI compute in low Earth orbit. Musk has stated publicly that SpaceX aims to deploy 10 gigawatts of AI compute capacity by 2027. Analysts cited in coverage suggest that target could require the procurement of more than two million Nvidia Rubin GPUs — a figure that, if accurate, would represent demand of an entirely different order than any prior single-customer commitment. An optimized Vera Rubin NVL72 configuration is reportedly slated for space deployment next year.
Nvidia's position as the default AI-infrastructure vendor is the product of a decade-long technical and ecosystem bet. The company's CUDA parallel computing platform, launched commercially in 2006, created developer lock-in that proved remarkably durable even as AMD and, more recently, custom-silicon programs at Google (TPU), Amazon (Trainium), and Microsoft (Maia) sought alternatives. Vera Rubin — entering volume production and scheduled to reach eight major cloud partners this fall, per Nvidia — represents the latest iteration of that roadmap: HBM4 memory reportedly triples bandwidth relative to predecessor configurations, and Nvidia cites a 10x reduction in per-token cost at inference. The HBM4 supply chain is itself a competitive battleground: SK Hynix, Samsung, and Micron are all reportedly vying for Nvidia's 16-layer HBM4 supply contracts, with SK Hynix targeting volume production in the third quarter.
The broader demand picture reinforces this momentum. Nvidia's B200 GPU systems are reported to be fully sold out within current shipment windows. Anthropic — simultaneously reported to be developing its own in-house AI accelerators — signed a ten-billion-dollar, six-year compute deal with Volta Infra, a months-old Nvidia-backed startup operating data centers in Norway, underlining how firmly Nvidia-powered infrastructure has become the default substrate even for frontier labs building toward alternatives. Nvidia is also reportedly the beneficial tenant behind a 704-megawatt, up-to-$50 billion lease on Hut 8's Beacon Point campus in Texas, per a Financial Times report that this outlet has not independently verified.
Against this demand backdrop, the risks are real and several. The most immediate geopolitical headwind: both Nvidia and Broadcom face potential exposure from a reported Chinese AI component export ban targeting custom silicon, networking, and interconnect products critical to data centers. CEO Jensen Huang met U.S. officials this week as regulatory scrutiny over China chip access intensified — a meeting that signals how central export-control policy has become to Nvidia's addressable-market calculus. Export controls have paradoxically also generated a gray market for H200 and B200 chips in Southeast Asia, complicating compliance and pricing. On the product side, reports that the Rubin Ultra variant will carry less HBM capacity than the standard Rubin have already prompted some procurement teams to revise their plans — a supply-specification mismatch that adds friction to an already complex upgrade cycle. Longer-term, the vertical integration trend is gathering pace: Anthropic's stated intention to design and manufacture its own AI accelerators follows analogous moves by Google, Amazon, and Microsoft, and represents a structural demand risk as frontier labs seek to reduce single-vendor dependence.
Three signals are worth tracking in the near term. First, whether Nvidia's export-control exposure to China — including the gray-market channel — draws regulatory action that materially impairs revenue, given that near-term supply is already heavily allocated to U.S. and allied customers. Second, how the HBM4 supply competition resolves: the outcome among SK Hynix, Samsung, and Micron will shape Nvidia's production-ramp velocity for Rubin-generation GPUs and, by extension, SpaceX's 2027 compute timeline. Third, whether Anthropic's and similar custom-silicon programs reach production scale fast enough to shift meaningful workloads off Nvidia silicon within two to three years. The SpaceX commitment provides Nvidia with a high-profile long-term anchor customer in a genuinely new compute vertical; how the company navigates regulatory risk, supply-chain concentration, and the gradual encroachment of in-house silicon will determine whether that anchor holds.