Data Centers in Space: A Brilliant Idea or Delusional?

SpaceX is pitching orbital data centers as part of a $1.75 trillion IPO. The physics of cooling chips in vacuum has not agreed.

That is the core tension in a wave of investment and filing activity around Orbital Data Centers — ODCs — that has drawn in SpaceX, Blue Origin, the Nvidia-backed startup Starcloud, and China's state coordination apparatus. The pitch is logical enough: launch compute above the atmosphere, eliminate terrestrial power grid constraints, harvest unlimited solar energy, and sell AI training cycles back to earth. The execution problem is heat. Not metaphorical heat. The actual thermodynamic problem of keeping silicon alive in a environment where the only way to shed watts is radiation.

A single Nvidia H100 GPU, the chip running today's frontier AI models, consumes roughly 350 watts in terrestrial data centers. In orbit, dissipating that power requires a radiator surface of about 1.1 square meters — roughly the size of a door — per chip, according to ArkSpace. That is because vacuum does not conduct heat. There is no air. There is only radiation. A full-scale orbital data center generating the gigawatts that a terrestrial facility produces would need a radiator surface the size of several city blocks.

Low Earth Orbit compounds the problem. Satellites in LEO swing from plus 120 degrees Celsius in direct sunlight to minus 150 degrees Celsius in Earth's shadow every 90 minutes. That thermal cycle, repeating constantly, stresses materials in ways that terrestrial data center engineers never have to solve. "Thermal management" is not a software problem. It is the core unsolved challenge, and Voyager Space CEO Dylan Taylor told ArkSpace a two-year timeframe for commercial orbital data centers "would be aggressive."

SpaceX's own prospectus — the company that would know — admits ODCs "may not achieve commercial viability," according to its S-1 filing with the SEC. The filing, part of SpaceX's ongoing IPO process targeting a $1.75 trillion valuation, frames orbital compute as a growth pillar rather than a side experiment. The company is not waiting for that question to resolve. It spent $7.7 billion on AI infrastructure in the first quarter of 2026 alone — six times the $1.3 billion it spent on Starlink, which already generates revenue, according to Trending Topics EU. SpaceX has filed with the FCC for a constellation of up to one million satellites for orbital data centers, Wikipedia reports.

The thermal math at scale is punishing. A 10-megawatt orbital data center — a small facility by terrestrial standards — could require 200,000 to 1,000,000 kilograms of radiators alone, costing $20 to $100 million just for the thermal management infrastructure at current launch prices, ArkSpace calculates. Scale to the 5-gigawatt facility Starcloud has proposed — about the size of a small terrestrial hyperscaler — and the launch cost for thermal hardware alone reaches hundreds of millions of dollars before a single server rack arrives on orbit.

SpaceX projects cost parity with terrestrial data centers by 2028, according to ArkSpace. That projection depends entirely on Starship, its heavy-lift vehicle that has not yet reached operational cost targets. SpaceX moved 2,213 metric tons of payload to orbit in 2025. Its stated goal is one million metric tons per year — 450 times the current rate. It has no long-term GPU supply contracts; it procures chips on a purchase-order basis, per the S-1 filing, which means the compute backbone of an orbital data center business is exposed to spot market silicon pricing.

The competition is not idle. Starcloud placed the first Nvidia H100 GPU in space on November 2, 2025, aboard its 60-kilogram Starcloud-1 satellite, per Wikipedia, and has since deployed Google Gemma, an open large language model, running on that hardware. Starcloud-2 is planned for October 2026 featuring Nvidia's Blackwell platform. Blue Origin announced the TeraWave constellation of about 5,400 satellites for orbital data centers in January 2026. China announced a 200,000-satellite constellation focused on data sovereignty and in-orbit processing. Nvidia announced in March that it would manufacture GPU chips specifically for space-based computing.

Elon Musk, whose SpaceX operates Starlink and whose xAI runs one of the world's largest GPU clusters, said in January that within two to three years, the lowest-cost way to generate AI compute will be in space, according to Introl. Sundar Pichai, the CEO of Google, told the WSJ the same technology is "a decade away" from being normal.

The radiator math has not changed because the IPO calendar is aggressive.