Elon Musk wants to put Nvidia server racks in orbit

AI labs on Earth are currently fighting utilities and local governments for every stray megawatt of power. Elon Musk thinks he has a simpler fix: bypass the power grid entirely and put the server racks in space.

SpaceX dropped a detailed 30-minute design video showcasing AI1, its prototype orbital data center satellite. The solar-powered spacecraft is engineered to house frontier AI silicon in a roughly 600-kilometer orbit, soaking up constant sunlight away from terrestrial grid bottlenecks.

The timing is hardly a coincidence. The reveal came just three days before SpaceX prices what is expected to be the largest IPO in history, targeting a valuation near $1.75 trillion with trading set for June 12.

A full Nvidia rack in orbit

The scale of AI1 is hard to overstate. Once deployed, its solar arrays stretch 70 meters across, slightly wider than the 68.4-meter wingspan of a commercial Boeing 747-8.

That massive surface area exists to feed a massive computing payload. SpaceX disclosed an average draw of 120 kW and a peak of 150 kW, delivering a compute density of 70 kW per ton. That power profile effectively matches a ground-based Nvidia GB300 server rack, which pulls roughly 140 kW.

Musk noted that the spacecraft uses an interchangeable hardware bay. SpaceX is not locking itself to any single chipmaker, meaning it can drop in whatever AI silicon proves most competitive. That flexibility is also practical necessity: while SpaceX is currently constructing the Terafab chip facility in a joint venture with Tesla, its IPO filing warned that it still cannot guarantee adequate chip supply on its own.

Musk called the craft much simpler than a Starlink satellite. Instead of complex steerable phased-array antennas, AI1 trades high-density radio gear for solar arrays, high-speed optical laser links, and sprawling radiators.

The cooling elephant in the room

While orbital sunlight is plentiful, cooling server silicon in a vacuum is notoriously hard. Terrestrial data centers dump heat into air or circulating water, but in space, infrared radiation is the only way out.

AI1 integrates up to 110 square meters of deployable liquid radiators, backed by redundant pumping loops and micrometeoroid shields. Yet industry observers remain skeptical about the physics. According to SemiAnalysis, the International Space Station's external thermal control system relies on 422 square meters of radiators to dump just 70 kW of heat, half of what a GB300 rack demands, built at a cost of roughly $500 million.

Musk has dismissed the thermal critiques, arguing that operating over 10,000 Starlink satellites has taught SpaceX everything it needs to know about space-based thermal management.

The company is already preparing a manufacturing base to match that ambition. SpaceX is planning an 11-million-square-foot factory in Bastrop, Texas, aiming for volume production before 2028. Back in January, the company also filed with the FCC for permission to launch up to one million compute satellites.

Why it matters

Space-based compute spent years being dismissed as expensive tech theater. Earlier this year, OpenAI chief Sam Altman called orbital data centers flat-out ridiculous. Yet Google and Anthropic have already signed on as launch customers, with Google committing to a massive $920 million monthly compute contract.

For frontier model builders, orbital compute solves the single biggest bottleneck of the decade: physical grid delays. Getting a new gigawatt substation approved and wired on Earth takes years of permitting and political pushback. In orbit, the power is already on. For everyday builders, orbital compute will stay behind enterprise pricing tiers for the foreseeable future, but it opens an intriguing alternative route for bulk training workloads.

Strategically, this move gives SpaceX a compelling second act right on the eve of its IPO. We think packaging launch capability with orbital energy is a clever way to capture high-margin AI spend rather than remaining a pure rocket taxi service.

Still, the economics carry serious red flags. Satellites cannot be serviced when components fry, meaning depreciation is brutal. Unless Starship drives launch costs to rock bottom, dumping heat and replacing hardware every few years could keep orbital compute vastly more expensive than Earth-based facilities. AI1 proves SpaceX has a serious blueprint, but making the balance sheet work in orbit is the real test.