On August 24, 2026, SpaceX CEO Elon Musk revealed plans to send a jointly designed space version of the Vera Rubin NVL72 system—co-developed with NVIDIA—into orbit by the fourth quarter of 2027, with plans to scale up in 2028. On the same day, NVIDIA announced that SpaceX's first-generation "Starmind" AI satellite would be built on an optimized version of the Vera Rubin NVL72. This gives the long-standing concept of orbital data centers a product name and a timeline. However, no completed satellite or in-orbit computing results have been demonstrated. To gauge the weight of the 2027 target, we need to separate three distinct questions: what elements carry over from ground-based racks, how power and heat dissipation will be managed, and how far the regulatory process has actually progressed.
Musk names Q4 2027; NVIDIA commits to no date
Musk's post stated that SpaceX and NVIDIA had designed a "space-optimized" version of the Vera Rubin NVL72 and would launch it in the fourth quarter of 2027. He further stated that by 2028 the system would reach "significant scale," though he provided no figures for satellite count or computing capacity.
What NVIDIA's announcement confirms is a development plan: the first-generation Starmind will be built on an optimized version of the Vera Rubin NVL72. The company cited communication bandwidth, reliability, and integration into the spacecraft as conditions unique to orbital computing, alongside power and thermal management. However, the announcement contains neither a Q4 2027 date nor a 2028 deployment scale. Therefore, while the joint development with NVIDIA is officially confirmed, the launch timeline is a target set by Musk.
NVIDIA's announcement outlines two separate plans—one on the ground, one in orbit. On the ground, SpaceXAI will adopt Vera CPUs to accelerate code execution and data processing performed by AI agents. Starmind, by contrast, is not simply an adoption of standalone Vera CPUs but a plan built on an optimized version of the Vera Rubin NVL72. Even if Vera CPUs are successfully deployed on the ground, this alone does not confirm the satellite's completion or environmental testing.
SpaceX's vision is not simply about purchasing a ground-market product and loading it onto a rocket. NVIDIA's own wording—"built on"—does not promise that the same component count or performance will be maintained in orbit. Conflating these points risks turning a development direction into a finished specification.
The scope of the partnership also needs to be separated out. While NVIDIA is bringing Vera Rubin to the first-generation Starmind, SpaceX has stated that the AI1 spacecraft design can accommodate computing modules from multiple manufacturers. Based on current official statements alone, adoption in the first unit cannot be equated with exclusive adoption across all future satellites.
How much of the 72-GPU ground design will survive?
The ground-based Vera Rubin NVL72 connects 72 Rubin GPUs and 36 Vera CPUs via a copper NVLink spine. It houses 18 compute trays and 9 NVLink switch trays, with the entire rack containing roughly 1.3 million components. It weighs approximately 4,000 pounds, or about 1,814 kg. NVIDIA's design treats this entire rack as a single giant GPU.
These figures are not the confirmed specifications for the orbital Starmind. SpaceX and NVIDIA have not disclosed the number of GPUs and CPUs to be carried, operating clock speeds, or power consumption limits. Nor is it clear how much redundancy will be built in for an environment where ground crews cannot replace failed trays. The ground version's roughly 1,814 kg also cannot be used as the satellite's total mass, since the spacecraft's mass—including solar panels, radiators, and attitude control systems—adds separately.
What "space-optimized" actually means will likely come down to trade-offs against constraints rather than maximizing peak performance. Will GPU count be reduced to limit power draw? Will a configuration close to the ground version run at maximum load only for short periods? Or will computing nodes be distributed on the assumption that failures will occur? None of these questions have been answered publicly, and the product name alone cannot fill in the gaps.
Communication design, too, cannot simply extend from the ground version. The copper NVLink spine links GPUs within a single rack; the pathways between satellites and back to Earth are handled by SpaceX's laser network. SpaceX states that computing results will be relayed via Starlink, but it has not disclosed the speed at which input data can be sent to orbit or the effective bandwidth of inter-satellite links. Even if computing capacity increases, the useful work that can be accomplished remains limited if data cannot be moved efficiently.
Cooling an average of 175kW in space aboard a 30-meter AI1
According to SpaceX's published current plans, the AI1 satellite will stand 30 meters tall and span 75 meters in width when deployed. Its computing payload is rated at up to 250kW peak and an average of 175kW, with processed results relayed to Starlink via high-bandwidth lasers. The company states that operating in a sun-synchronous orbit allows continuous solar exposure, enabling it to expand computing resources without relying on ground power grids or land use.
The 250kW peak and 175kW average figures refer to power consumption by the computing payload—not solar panel output. SpaceX's current materials do not specify solar panel output, battery storage capacity, or the range of load fluctuation. As a result, it is impossible to determine how long peak load can be sustained.
However, vacuum itself is not a coolant. According to NASA thermal design documentation, there is no convection in a vacuum; heat generated inside a spacecraft must be conducted to radiators and ultimately released into space as infrared radiation. The rate of heat rejection depends on radiator surface area, emissivity, and operating temperature. SpaceX states that eliminating the chillers and cooling towers required by ground facilities can cut cooling's power burden by an order of magnitude, but it has not disclosed AI1's radiator surface area, cooling loop design, or how long peak load can be sustained.
A filing for up to 1 million satellites and 2028's "scale" are separate numbers
SpaceX is building a "Gigasat Factory" in Bastrop, Texas, aiming to produce and deploy thousands of AI satellites starting as early as late 2027. Musk's stated first-launch target overlaps with the factory's planned startup timeline. But factory completion, satellite mass production, launch, and in-orbit operation are each separate milestones.
SpaceX has stated that deploying heavy AI satellites will require Starship. Yet it has not disclosed which rocket will be used for the initial Q4 2027 mission, nor how many AI1 units it will carry. Even once the satellite factory becomes operational, launch capacity and regulatory approval will not necessarily align on the same timeline.
The regulatory figures involved are far larger still. On January 30, 2026, SpaceX filed an application with the U.S. Federal Communications Commission (FCC) for an "Orbital Data Center System" comprising up to 1 million non-geostationary satellites. The planned altitudes range from 500 to 2,000 km, using both 30-degree inclined orbits and sun-synchronous orbits, with satellites linked via optical connections. What the FCC did on February 4 was accept the filing and open it for public comment—not approve operation of up to 1 million satellites. The filing also includes requests for exemptions related to spectrum use and deployment obligations.
Optical links alone will not complete the regulatory process either. SpaceX is seeking authorization to use frequencies in the 18.3–19.3 GHz range for satellite-to-ground links and 28.6–29.1 GHz for ground-to-satellite links, and has agreed to accept conditions under which it would not be protected from interference. The target for the first launch and approval to operate the full communication network are proceeding on separate tracks.
Q4 2027, 2028's "significant scale," the factory's targeted thousands of units, and the FCC filing's ceiling of up to 1 million satellites are not the same plan expressed in different numbers. Rather, they represent a near-term launch target, a mass-production vision, and a regulatory ceiling being requested—laid out side by side. What would turn the 2027 launch from ambition into a concrete technical achievement is whether the first unit's configuration and launch vehicle are disclosed, whether it passes thermal vacuum testing and reliability verification, and whether it can return useful processing results from orbit. Only once that threshold is reached will the orbital data center move from concept to working experimental system.