On Sunday, November 2nd, NVIDIA sent its powerful H100 GPU to space for the first time to test how data centers could work in orbit.

The GPU, featuring an 80GB RAM, is a hundred times more powerful than any computer ever flown in space. It will test a range of AI processing applications including analyzing Earth observation images and running a large language model by Google. The test flight, on board of the Starcloud-1 satellite by Redmond, Virginia-based start-up Starcloud, is the first step in an ambitious plan to move the world’s power-hungry data-crunching infrastructure to space. Proponents think the idea makes sense: far above the planet, in the emptiness of space, data centers wouldn’t take up precious land, or take up as much energy and water for cooling. They wouldn’t release warming greenhouse gases into the atmosphere either. ****

NVIDIA chips have been to space before. The Jetson machine learning computing board has flown on several experimental and Earth-observing small satellites. But H100 test flight opens an entirely new chapter for AI data-processing in space, paving the way for commercial services to start as early as next year.

“The H-100 is about 100 times more powerful than any GPU computer that has been on orbit before,” Philip Johnston, CEO and co-founder of Starcloud, told IEEE Spectrum. “It will be the first time that a terrestrial-grade data center GPU will be flown and operated in orbit.”

The three-year mission will launched on SpaceX’s Bandwagon 4 Falcon 9 flight. The 60-kilogram Starcloud-1 satellite will orbit Earth in a very low orbit at an altitude of about 350 kilometers. While there, it will be receiving data from a fleet of synthetic aperture radar (SAR) Earth-observing satellites operated by U.S. company Capella, process it in real time and beam insights to Earth.

“Downlinking SAR data has historically been a huge problem because it’s extremely voluminous,” said Johnston. “But being able to process it in orbit will mean that we can only downlink the insight. The insight might be that there is a vessel at a certain location moving at a certain speed and in a certain direction. That will be just a one-kilobyte packet versus the hundreds of gigabytes you would need to downlink the data.”

The case for data centers in space

But orbital processing of data from Earth-orbiting satellites is only one part of Starcloud’s vision. The company thinks that with advances in rocket technology, especially with the cost reductions expected from SpaceX’ Starship, future large-scale computing infrastructure could be placed in orbit rather than taking up precious space on Earth.

“As the energy demands of AI development increase, orbital data centers represent a transformative environmental breakthrough—cutting greenhouse gas emissions by orders of magnitude and eliminating the need for advanced cooling,” Josh Parker, Head of Sustainability at NVIDIA told IEEE Spectrum in an email. “By harnessing low-cost, nonstop solar energy and avoiding land use and fossil-fuels, Starcloud’s technology allows data centers to expand rapidly and sustainably, helping protect Earth’s climate and critical natural resources as digital infrastructure grows.”

According to the United Nations Environmental Program (UNEP), the world’s data-crunching infrastructure is set to consume as much electricity by 2030 as the entire state of Japan. Data centers also require enormous amounts of water for cooling—a single one-megawatt data center consumes as much water as about one thousand people living in the developed world, UNEP data suggests. The demand for computing continues to grow in step with advances in AI and so does data centers’ hunger for consumables. The impacts are most intensely felt by communities in the vicinity of these sites, which are increasingly concerned about rising cost and disruptions to power and water supplies. Moving data centers to space could solve these problems, proponents of the technology think.

“My expectation is that within 10 years, almost all new data centers will be built in space purely because of the constraint that we’re facing on energy terrestrially,” said Johnston.

Making data centers on Earth run on fully green energy requires major investments into solar power generation and battery storage systems, he said. In space, no battery storage is needed as sunlight is available 24/7. On top of that, each solar panel produces eight times as much electricity as its equivalent on Earth, which further decreases the cost, he added.

“The only additional cost we have in space is the launch,” Johnston said. “We see a breakeven with launch cost of around $500 per kilo. With Starship, we’re expecting launch costs much lower.”

Per kilo launch prices for SpaceX’s Starship, once the mega-rocket is fully operational, are estimated to range from $150 to as little as $10. The launcher has so far completed six successful flights and is expected to send the first batch of satellites to orbit next year.

Starcloud is already planning its next mission, hoping to launch a ten-times more powerful data center than Starcloud-1 to space next year. The Starcloud-2 mission will be fitted with NVIDIA’s Blackwell GPU and several H100s. That mission, providing 7 kilowatts of computing power, is expected to provide commercial services to customers including Earth observation satellite operators and the U.S. Department of Defense, said Johnston.

An even larger, 100-kilowatt satellite, should reach orbit in 2027. By the early 2030s, Starcloud thinks they can have a 40-megawatt data center in space crunching data at a cost equivalent to data centers on Earth. Starcloud is one of several companies with plans to outsource computing to space. Axiom Space unveiled similar plans earlier this year. Florida-based Lonestar Holdings sent a small data center to the Moon with the Intuitive Machines-2 mission earlier this year and envisions building up major data centers on Earth’s companion in the coming years.