Nvidia just sent an H100 GPU to space to test orbital data centers

AI data centers on Earth are running out of power, draining local water supplies, and sparking neighborhood backlash. Nvidia's solution is straightforward: launch them into orbit.

Nvidia sent an H100 GPU into space aboard a SpaceX Falcon 9 rocket as part of the Bandwagon 4 mission. It marks the first time a terrestrial-grade data-center processor has ever been deployed outside our atmosphere.

The 60-kilogram satellite, called Starcloud-1, was built by Virginia-based startup Starcloud. Orbiting at a very low altitude of around 350 kilometers, the 80GB H100 packs roughly 100 times the computing power of any machine previously launched into orbit.

A full data-center rig at 350 kilometers

Starcloud-1 isn't just idling up there. Its first task is processing live synthetic aperture radar (SAR) data from Capella satellites and running a Google large language model directly on the hardware.

SAR data is notoriously massive, often creating massive bandwidth bottlenecks when downlinking to Earth. By crunching numbers in orbit, the H100 can extract actionable insights on the fly. Instead of beaming down hundreds of gigabytes of raw sensor data, it can broadcast a 1KB packet confirming a specific vessel moving at a set speed.

Beyond edge sensing, Starcloud and Nvidia see a massive energy arbitrage. The International Energy Agency expects terrestrial data centers to consume as much electricity as all of Japan by 2030, while each megawatt requires water cooling equivalent to 1,000 residents in developed nations.

In low Earth orbit, land is free and sunlight shines around the clock. Solar arrays generate eight times more electricity than they do on Earth, eliminating the need for expensive battery backup. Starcloud claims orbital compute breaks even with terrestrial costs at a launch price of $500 per kilogram, an expense SpaceX's Starship could slash to between $10 and $150.

Solar dreams versus orbital physics

Starcloud is moving fast. The team plans to launch the 7-kilowatt Starcloud-2 next year using Nvidia's Blackwell GPUs, aiming for commercial customers like the U.S. Department of Defense. Plans for a 100-kilowatt rig follow in 2027, scaling toward a 40-megawatt cluster in the 2030s alongside efforts from Axiom Space and lunar-focused Lonestar.

Yet running a beast like the H100 in a vacuum violates standard engineering comfort zones. Nvidia CEO Jensen Huang pointed out regarding the Vera Rubin Space 1 module that space lacks conduction and convection. Cooling must happen purely through thermal radiation.

Dissipating hundreds of watts without ambient air requires massive, heavy radiators that add significant launch mass. Furthermore, the H100 is commercial silicon, not radiation-hardened hardware. Whether it can survive three years of bombardment from cosmic rays remains an open question.

Why it matters

Terrestrial power grids are hitting a hard wall, forcing hyperscalers to scout nuclear reactors. Moving compute off-world represents an audacious pivot, but we think everyday developers won't be spinning up space-based instances anytime soon.

For defense contractors, intelligence agencies, and satellite operators, the value is immediate. Shrinking a multi-gigabyte data downlink down to a single kilobyte transforms maritime tracking and emergency response from an hours-long pipeline into near real-time telemetry.

Starcloud's playbook smartly bets on SpaceX driving launch costs into the ground while tapping unthrottled solar power. However, vacuum thermodynamics and cosmic radiation are unforgiving. Earth-grade chips might degrade rapidly, and radiator mass could easily wipe out projected rocket savings. Until thermal dissipation and radiation survivability are proven over multi-year cycles, orbital AI remains an exotic, high-stakes gamble.