SpaceX "banned" Nvidia's rival, all-in for the gamble of placing data centers in orbit

In early August, Elon Musk posted a rather brief line on Twitter: SpaceX and xAI will only use Nvidia chips, "because that's the best." It sounds like a polite compliment, but when Musk repeated this right in SpaceX's first financial report, it turned into a strategic statement with a completely different weight: a series of other AI processing chip companies such as AMD, Cerebras, D-Matrix, Rebellions, SambaNova or Tenstorrent were almost removed from the list of suppliers for Musk's two most important companies. And the most interesting part is not who SpaceX chooses, but where they plan to take those chips: straight into Earth orbit.

Why would a server architecture make Musk speak so publicly?

During the financial report, Musk further explained why he chose Nvidia: "We decided to build exclusively on the Nvidia platform, because we think the Vera Rubin architecture is the best architecture. We think it is the best AI computer, and we very much value our close cooperation with Nvidia on many levels." He added that the design of the NVL72 rack server system using the Vera Rubin VR200 processor is "much better" than the conventional standard rack design.

To Musk, NVIDIA's products are much better than other competitors The technical detail behind this praise is the cable-free design of the computing bays in the NVL72. It may sound small, but this is the most notable point: most problems when installing large-scale data centers come from connecting thousands of cables and pipes by hand, if just one technician misplaces one connector and the entire cluster can fail. Eliminating most of those manual connections increases assembly speed, reliability and maintainability, which are doubly important when the assembly is located not in a building in Texas but hundreds of miles above ground, where no one can fly up to fix a loose cable.

Starmind: when the data center is no longer grounded

According to Musk, SpaceX will launch a program called Starmind, with the core AI satellite being "basically an optimized NVL72 Vera Rubin computer." He emphasized that this is "not something far away in the future", SpaceX plans to start launching in 2027. The ambition here is much larger than the Space-1 Vera Rubin Module project that Nvidia itself is developing, which only aims to put a few, maybe a dozen, Rubin processors into orbit. SpaceX's plan is to send a complete NVL72 cluster, including 36 Vera CPUs and 72 Rubin GPUs, to space.

Starmind, Elon Musk's new project in which NVIDIA will have an important role in providing processing power The specific hardware for this ambition is a satellite called AI1, which SpaceX has revealed in detail. This is a machine of stunning proportions: when fully deployed, AI1 spans about 70 meters from one end to the other, which is wider than a Boeing 747, and is about 20 meters high. Its peak computing power is about 150 kilowatts, with an average of about 120 kilowatts. Because in a vacuum there is no air or water to dissipate heat in the usual way, AI1 uses a 110-square-meter liquid heatsink to expel excess heat into space in the form of infrared radiation, along with a large solar panel to maintain a nearly constant power supply, an advantage that terrestrial data centers do not have because they depend on the day-night cycle and the local power grid. The AI1's design also allows for interchangeable compute chips, meaning that in theory many different chip companies could supply processors for the same satellite chassis, although in this announcement Nvidia is almost certainly the sole supplier. SpaceX doesn't stop at just one or a few test satellites. The company has submitted an application to the US Federal Communications Commission (FCC) for a license to deploy up to one million orbital data center satellites, operating at altitudes from 500 to 2,000 km. This is a scale that I think, if compared, even exceeds SpaceX's current Starlink satellite network, which already has more than seven thousand active satellites. The first prototypes are scheduled for launch in early 2027, with commercial deployment and mass production targeted for later that year. Notably, SpaceX has signed at least one computing contract worth $920 million per month with Google, showing that this is not just a technology demonstration project but has real customers willing to pay.

The bigger picture: Terafab, and a closed ecosystem from chip to orbit

To understand why SpaceX dares to bet so big on a single architecture, you need to look at another piece of Musk's ecosystem: Terafab, the $16.8 billion chip factory that Tesla and SpaceX just announced they will build in Texas, with a scale described as larger than the Pentagon, Apple Park, Mall of America and Giga Texas combined. Terafab is designed to solve the exact problem that both Starmind and Musk's robot and self-driving car ambitions are facing: the current global supply of AI processing chips is simply not enough. When Tesla, SpaceX and xAI all run on the same Nvidia ecosystem, and at the same time Musk is building his own chip factory to partially control the supply, the picture that emerges is a closed chain: chips produced by a factory co-owned by Musk himself, running on architecture designed by Nvidia, serving projects also run by Musk, from Optimus robots, Cybercab self-driving cars, to orbital data centers.

Is this a reasonable step?

Actually, it is reasonable to be skeptical here because putting a data center into orbit is not an idea that no one has ever thought of, but an idea that many people have considered and then shelved it because the economic problem is not clear. Constant solar power and the cold air of a vacuum sound appealing, but launch costs, the reliability of the liquid cooling system when operating for years without on-site maintenance, and the risk of collisions with space debris are all variables for which there are no definite answers. Even technology analysts, when discussing competitors in the orbital data center race, frankly admitted that the economic problem of this model "has not yet been proven." But one thing may be different from previous times: SpaceX isn't just talking, they have paying customers, have applied for a million-satellite license, and have tied this project into a separate chip supply chain that is being built in parallel. Whether this gamble succeeds or fails, it represents a trend worth watching: as land, electricity, and cooling water on the ground increasingly become the real limits of AI ambitions, the idea of ​​leaving Earth is no longer a fantasy, but is becoming an investment category with a name, a launch date, and a specific contract.