SpaceX Selects Nvidia for Starmind Orbital AI Compute

Yahoo Finance reports that Musk set a target of more than 2 GW of compute capacity by the end of 2026 and close to 10 GW by the end of 2027. He also said SpaceX intends to deploy Nvidia's Vera Rubin NVL72 rack-scale system, referred to as Kyber, in both terrestrial facilities and space through the Starmind program.

Musk said Starmind satellite launches would begin next year, according to Yahoo Finance. The reported targets are forward-looking statements from the earnings call, rather than installed-capacity figures.

Morningstar reported that SpaceX generated $7.8 billion in second-quarter revenue, up 92% year over year, while investing $3.5 billion in R&D and recording a $143 million operating loss. Its analysis characterized the company's AI ventures as dependent on deploying compute capacity and renting it to customers. Separately, Yahoo Finance reported AI revenue of $2.6 billion, attributing the growth to cloud-service agreements and increased Grok and X subscriptions; the available reports do not establish whether that figure uses the same revenue definition as Morningstar's total company revenue.

Nvidia launched its space-computing platform in March and named Aetherflux, Axiom Space, Kepler Communications, Planet Labs, Sophia Space, and Starcloud as early partners, according to BeInCrypto. SpaceX was not named in that initial group.

Wccftech reports that Nvidia's space-oriented Space-1 Vera Rubin module combines four Rubin GPUs and two Vera CPUs. Nvidia claims the module provides up to 25 times the AI-compute capability of an H100 GPU for orbital workloads. That is a vendor performance claim, and the sources reviewed do not provide independent benchmarks for the Starmind configuration or its expected end-to-end throughput.

For ML infrastructure teams, the salient architectural point is not simply that accelerators are being placed in orbit. In comparable edge-compute deployments, processing data near its source can reduce the amount of raw data transmitted to a central facility, but radiation tolerance, thermal design, power management, fault recovery, networking, and physical serviceability become first-order constraints. The reporting does not yet specify Starmind's model-serving stack, storage architecture, networking protocol, or workload mix.

BeInCrypto reports that SpaceX sought Federal Communications Commission authorization in January for as many as one million orbital data-center satellites across altitudes of 500 to 2,000 kilometers. The application describes a petabit laser mesh, according to that report. The FCC had accepted the filing in February but had not ruled at the time of publication, BeInCrypto reported.

The proposed scale remains a major open question. Public reporting describes the project as an orbital data-center constellation, but no source reviewed provides deployment milestones beyond Musk's statement about launches beginning next year. Industry experience with distributed compute systems indicates that aggregate accelerator capacity alone does not determine usable capacity: interconnect reliability, scheduling, data locality, and failure rates shape the performance available to actual workloads.

The Nvidia agreement provides a named hardware path for the first Starmind payload. The operational and regulatory conditions required for a large orbital AI-compute network remain unresolved in the reporting available so far.