A near term U.S.–China race to put AI compute in space is organized less by who is winning and more by which rung of a three tier architecture (edge satellite, dedicated compute satellite, full multi megawatt orbital data center) each country has actually reached.
China has already run a general-purpose AI model in space and used a satellite to command a robot on the ground. The U.S. has a high-end AI chip in orbit and a roadmap. The race to build data centers in space is real, but it looks less like a sprint and more like a ladder, and right now the two countries are on different rungs.
In November 2025, Chengdu-based startup ADA Space ran Alibaba's Qwen3 large language model on one of its operational satellites in low Earth orbit and got a response back in under two minutes. A few weeks later, working with Shanghai Jiao Tong University, the same company used compute on a satellite to command a robot moving across a ground lab. By Recode China AI's account of the demonstrations, these count as the first time a general-purpose AI model has run in orbit and the first time orbital compute has driven a machine on Earth.
On the U.S. side, the milestones look different. Nvidia-backed Starcloud flew an H100 GPU to low Earth orbit on a SpaceX rideshare in November 2025, ran a Google model on it, and has a Blackwell chip planned for around October 2026. Nvidia CEO Jensen Huang called the moment "space computing, the final frontier, has arrived." The day ADA Space filed for its Hong Kong IPO, SpaceX priced its own listing at a $1.75 trillion valuation on Nasdaq under the ticker SPCX, raising $75 billion and closing up roughly 30 percent on debut. The roadshow deck, by Recode China AI's reconstruction, sold orbital data centers alongside Mars.
The question is not which country is winning. It is which rung of the ladder each country is actually on, and what the engineering says about the climb.
Zhang Shancong, president of the Beijing Academy of Artificial Intelligence Sciences and Technology, or BAIST, lays out a three-tier framework that has become the de facto way Chinese planners think about this. The bottom rung is the edge computing satellite: a small platform under one kilowatt of power that handles inference or pre-processing close to a sensor, like a smart camera with a brain. The middle rung is the dedicated computing satellite, between one kilowatt and one megawatt, hosting model training or larger inference jobs. The top rung is the space data center proper, at multi-megawatt to gigawatt scale, with its own power generation, thermal management, and a swarm of spacecraft feeding it.
Right now, both countries are on the bottom rung with one foot on the middle. Neither has a true space data center.
China's most concrete deployed work sits in the Three-Body Computing Constellation, a partnership between Zhejiang Lab and ADA Space. The first 12 satellites launched in May 2025 on a Long March vehicle, with the opening batch delivering 5 POPS (peta operations per second, a measure of AI compute throughput) and 30 terabytes of storage. The roadmap runs to a thousand satellites and a pooled thousand POPS by 2032. ADA Space's own "Star Compute" plan calls for 2,800 satellites, split 2,400 for inference and 400 for training, spread across sun-synchronous, dawn-dusk, and low-inclination orbits.
The engineering constraints are real, and they explain why both sides are climbing in increments. Heat has to be shed into the vacuum of space, which means radiator panels, fluid loops, and in some designs pumped two-phase cooling where a working fluid boils on hot surfaces and condenses on cold ones. Power has to be generated continuously, which is why engineers keep returning to dawn-dusk sun-synchronous orbits, a near-polar low Earth orbit where the satellite never enters Earth's shadow and gets nearly constant sunlight. Comospace, a Chinese Academy of Sciences and Zhejiang Lab spin-off, is designing a ten-thousand-processor cluster around a 100-megawatt energy module with flexible solar and storage, a ten-terabit-per-second laser mesh, and a software-defined resilience layer running on commercial-off-the-shelf chips rather than radiation-hardened silicon. The bet is that software can compensate for hardware that is not built for space.
The Chinese Academy of Sciences' Aerospace Information Research Institute has separately demonstrated a 120-gigabit-per-second laser link between an orbiting satellite and a ground station over the Pamir Plateau, doubling the previous 60 gigabits per second by reconfiguring the on-board software in orbit, and downlinking more than 12 terabytes in a 108-second pass.
The U.S. side, by contrast, has concentrated more capability into fewer, more powerful units. The Starcloud H100 is a single, high-end data center chip in orbit, not a constellation of modest ones. SpaceX's pitch is that launch economics will eventually be there: the company has talked about targeting payload costs in the low hundreds of dollars per kilogram and eventually as low as $183 per kilogram with full reuse. The maiden flight of China's Long March 12B on June 1, a 72-meter partially reusable rocket in the Falcon 9 class, flew expendable on debut and carried payloads for the Qianfan megaconstellation, a planned Chinese broadband satellite network.
Two structural gaps make the Chinese climb slower. On hardware, domestic AI accelerators trail U.S. state-of-the-art chips by four to five years, so a Chinese system needs more chips to reach parity, which means more power, more weight, and a harder thermal problem. On launch, China is still mostly flying expendable boosters, and reusability at SpaceX's cadence is years away.
China's regulatory scaffolding is moving fast. The China National Space Administration's commercial space department is running a national feasibility study for a unified space-based intelligent computing constellation. CAICT, a think tank under the Ministry of Industry and Information Technology, has stood up a Space Computing Power Professional Committee to align standards for chips, lasers, thermal systems, and solar arrays, while MIIT drafts a full stack covering hardware, operating systems, networking, and cybersecurity. Regional clusters are forming in Beijing, Shanghai, Hangzhou, and Chengdu. CAICT's preliminary estimate puts China's space-based computing power industry at more than 250 billion yuan (around $36.6 billion) by 2030. CAICT frames that figure as a projection, not a confirmed forecast.
Wang Shangguang, dean of Beijing University of Posts and Telecommunications, has been blunt about the gap. SpaceX, he said, can "pull the entire industry chain together," while China is "small, weak, scattered, slow." His twelve-character diagnosis, 网网不同、星星并存、算算失衡, translates to fragmented networks, isolated satellites, and broken compute. On June 1, Beijing opened a space-computing innovation center in Haidian, led by BUPT, with six priority areas including radiation-tolerant chips; a parallel research institute is being set up in Beijing E-Town.
There are commercial claims worth flagging. ADA Space filed for a Hong Kong IPO at a pre-IPO valuation of 11.5 billion yuan, roughly $1.7 billion, on 2025 revenue of 703 million yuan and a net loss of 256 million yuan. Oriental Tiansuan and Guangbenwei announced what they called the "world's first space-based optical computing satellite" in May 2026; the claim is company- and ministry-aligned and has not been independently benchmarked. Orbital Chenguang closed a Pre-A1 round in April 2026 and reportedly secured a 57.7 billion yuan strategic credit line, around $8.4 billion, though the credit line figure should be verified against a primary disclosure before being treated as settled.
What to watch is whether China can compress the four-to-five-year accelerator gap through sheer system-level engineering, deploying more chips, smarter thermal design, and software-defined resilience on commercial hardware, and whether the U.S. can move from a single H100 in orbit to a working constellation before China's edge-and-mid-tier deployments compound into something harder to catch. The orbital AI race is real, but it is a ladder race, not a sprint, and right now the two runners are on different rungs.