AI heads to space as orbital computing gathers momentum

The competition over artificial intelligence infrastructure is extending beyond terrestrial data centers and into orbit. As China moves ahead with efforts to process data aboard satellites, US companies are also testing how far computing can be pushed off the ground and into space.

China has reported new progress in this direction. The latest example came on Sept 20, when China launched PEGA-SUS1, an experimental satellite intended to integrate communications, sensing, computing and intelligent processing.

Instead of transmitting raw remote-sensing images of a possible fire to Earth, the satellite could analyze the data itself, determine whether a fire was present, estimate the affected area and send the results back.

Another progress is a space-based computing cloud built on the Tiansuan satellite constellation, which recently began offering regular in-orbit experimental services. Its capabilities include processing data and performing intelligent analysis in orbit, according to Xinhua News Agency.

US companies are also pushing ahead with demonstrations.

Joe Yaffe, chief operating officer of Cowboy Space, told China Daily that the company planned an April 2027 mission that would run a cluster of Nvidia GB200 chips in orbit and process weather data under a collaborative research agreement with the US National Oceanic and Atmospheric Administration.

The company, based in San Carlos, California, and Seattle, Washington, develops orbital infrastructure for AI data centers. It also plans to run a commercial workload for an unnamed AI lab during the April mission.

Yaffe said the company had already tested weather-data processing in a simulated space environment on the ground, reducing a task that took 12 hours to a mere 35 seconds.

"That's a real critical strategic advantage that needs to be developed," he said.

The technology, known as orbital or space-based computing, is moving closer to practical use as satellites become capable of processing more of the information they collect in space rather than sending vast quantities of raw data back to Earth.

Satellites traditionally collect images and other data and transmit them to ground stations for processing. Putting greater computing capacity in orbit would allow satellites to analyze information and send results back to Earth, potentially reducing the time needed to turn raw data into weather forecasts, disaster alerts and other actionable intelligence.

The race to put more computing power in orbit comes amid far more ambitious proposals to eventually build large-scale AI data centers in space. But speakers at the AI Infrastructure Summit held on Sept 15-17 in Silicon Valley cautioned that major challenges remain, including heat management, communications bandwidth and launch costs.

Sumeet Singh, chief data and AI officer at satellite communications company Viasat, said heat dissipation was one of the biggest technical obstacles. Unlike on Earth, computing hardware in space cannot rely on conventional cooling systems, requiring large radiators that add mass and therefore increase launch costs.

Moving data between orbit and Earth is another constraint, while current launch prices remain far above the level at which some studies suggest large orbital data centers would become commercially competitive.

Deepak Sachdeva, whom the summit organizers identified as chief information officer of the US Air Force, said orbital computing was not yet fully ready for demanding operational use, particularly when systems need to process and transmit large volumes of information in near real time.

"I think it'll be ready in about 12 months, 18 months or so," Sachdeva said in an interview after the session. "The rate is very, very fast now."

He said simply putting computing hardware in orbit was not enough. Systems must be able to handle rapidly increasing computing loads, maintain reliable communications and continue operating when physical repairs are impossible.

"You cannot just walk to a place and then fix it," he said. "If you press a button and you get an answer next day, then what's the point of it?"

But at the commercial level, Sachdeva said the technology is ready. "The best approach, in my view, is to do some sort of proof of concepts with the reconnaissance for the other equipment, and see how this actually operates."

Asked about China's progress, Sachdeva said a "huge amount of work" had been done and highlighted the pace of execution. He suggested deployment could be faster in China, while longer cycles in the US could slow the process.

Pushing back on Sachdeva's timeline, Yaffe said the technology would develop faster than many expect.

"I think the gold rush is here. I mean, the amount of capital flowing into the industry is huge," he said. "The AI infrastructure buildout by itself is a $7 trillion proposed capital expense." He suggested that even a small fraction of that capital flowing into space-based computing would represent a significant amount of investment.

As money and attention pour into the sector, the field is also increasingly viewed through the lens of US-China technological competition.

Asked whether he viewed orbital computing as another arena of US-China competition, Yaffe said: "It's 2026, and you are in the United States. Yes, there's a competition with China on everything."

"People talk about how we're way ahead, but frankly, I think China has done a lot in this area, from what I understand," said Yaffe.

"I cheer anybody on who's developing technology, whether that's here or over there," he added.