On May 14, 2025, a Long March-2D rocket lifted off from the Jiuquan Satellite Launch Center carrying 12 satellites. They were not designed for Earth observation or communications relay. Their purpose was computation.
The satellites — the first batch of the “Three-Body Computing Constellation,” developed by Zhejiang Lab and its partners — had a combined in-orbit computing power of 5 POPS (peta operations per second). The highest single-satellite computing power reached 744 TOPS. Within nine months, six of the twelve satellites had established inter-satellite laser links, and a space-based distributed operating system was managing their compute, storage, and network resources as a single pool.
By the time the system entered regular experimental service in August 2026, it had deployed multiple AI models directly in orbit. One of them — an astronomical observation model — could classify gamma-ray bursts onboard, compressing hundreds of megabytes of raw data down to tens of kilobytes and reducing processing time from hours to seconds.
The constellation is not a demonstration. It is a computing infrastructure. And it is already operational.
The Scale of China’s In-Orbit Computing
China has moved beyond single-satellite experiments. It is deploying computing capacity in orbit as a layered, networked system.
In September 2026, the “Supercomputing-1” satellite launched from Jiuquan carrying a 4-meter optical camera and an onboard AI computer. The satellite can process remote sensing imagery, identify targets, and transmit only high-value results back to Earth — reducing cross-domain data response time from hours to minutes. According to the developer, it is the first orbital node in a planned distributed space computing network.
A companion satellite on the same launch — “Pengcheng Nanke-1,” developed by Galaxy Space, Pengcheng Laboratory, and Southern University of Science and Technology — integrated a 5G NTN base station, a core network, onboard AI computing, and a 100 Gbps satellite-to-ground laser communication payload into a single 150-kilogram spacecraft.
The Tiansuan constellation, led by Beijing University of Posts and Telecommunications, has deployed 16 low-Earth-orbit satellites and seven ground stations. Its space computing cloud began offering regular in-orbit experimental services in August 2026, supporting more than 60 institutions in orbital tests and drawing researchers from more than 50 countries. The platform automates end-to-end operations across satellite platforms, space servers, ground stations, and data centers — from task submission to results delivery.
The capability is not limited to remote sensing. In 2025, Guoxing Aerospace and Shanghai Jiao Tong University demonstrated remote control of a ground robot via space-based computing. The satellite processed the control signals in orbit and transmitted them to the robot on the ground.
China’s national strategy reflects the ambition. The Beijing Space Computing Innovation Center was inaugurated in June 2026. The city of Chengdu is building a space computing hub targeting a 1,000-satellite commercial network by 2030, with full deployment by 2035, interconnected with more than 100 ground-based data centers. Multiple provinces have established space computing research institutes and industrial alliances.
The 2026 Space Computing Industry Conference in Beijing was not a research symposium. It was an industry conference — with companies, investors, and government officials discussing commercial deployment.
The American Approach: Capital and Constellations
The United States is not absent from the orbital computing race. It is approaching it differently — with larger capital commitments, larger constellations, and longer timelines.
Starcloud, a Seattle-based startup, put the first NVIDIA H100 GPU into orbit in November 2025 aboard its Starcloud-1 satellite. The company has since trained an AI model in space, run a version of Google‘s Gemini in orbit, and demonstrated high-performance inference on flight hardware. In August 2026, it announced a funding round with participation from NVIDIA and Cisco, at a valuation reported by the company.
Starcloud is working with NVIDIA on the Space-1 Vera Rubin Module, a space-rated GPU designed for the radiation and thermal conditions of orbit. Its long-term target is a constellation of 88,000 satellites providing 20 gigawatts of orbital computing capacity. Production lines for its next-generation spacecraft are under construction at a 100,000-square-foot facility in Woodinville, Washington.
SpaceX has filed an application with the Federal Communications Commission to deploy up to 1 million satellites in low Earth orbit to form an orbital data center network. The company disclosed the conceptual design of its first orbital computing satellite, AI1, in June 2026 — a spacecraft that, with solar arrays fully deployed, spans 70 meters. Each AI1 satellite is designed for 120 to 150 kilowatts of power, with stable operating consumption around 120 kilowatts. SpaceX’s target computing density is 100 kilowatts per ton, though AI1 is expected to reach only about 70 kilowatts per ton initially.
Google launched its first Suncatcher experimental satellite in October 2026, carrying a Tensor Processing Unit into orbit. The satellite will validate whether Google‘s AI chips can operate reliably in the space environment. Google’s own research estimates that achieving commercial viability for large-scale orbital data centers would require reducing launch costs to approximately $200 per kilogram — a target that would require roughly 1,800 Starship launches to accumulate the necessary payload mass.
Cowboy Space, a California-based startup, plans an April 2027 mission to run a cluster of NVIDIA GB200 chips in orbit and process weather data under a research agreement with NOAA. The company has already tested weather-data processing in a simulated space environment on the ground, reducing a task that normally takes 12 hours to 35 seconds.
The Cost Barrier
The American approach is constrained by a single variable: the cost of getting mass to orbit.
Deutsche Bank‘s orbital data center cost model, published in July 2026, estimated that deploying 1 gigawatt of space-based AI computing capacity currently costs approximately 6 times the cost of an equivalent ground-based facility. The gap is driven by two factors: launch cost and satellite hardware.
For a 2027 deployment using conventional satellites, launch costs run about $1,429 per kilogram and the satellite hardware that is not computing costs about $50,000 per kilowatt. For a 100-satellite deployment, those two line items alone come to $115 billion, against roughly $20 billion for the equivalent ground-based system.
Deutsche Bank’s model shows the gap narrowing as Starship becomes fully reusable. By 2029 — when SpaceX‘s AI1 satellites are expected to deploy — launch costs fall to about $398 per kilogram and satellite non-computing hardware to about $13,000 per kilowatt, bringing total non-computing deployment costs to roughly $27 billion, close to the $23 billion for a ground-based system and cutting the cost multiplier from six times to 1.2. By 2032, with launch costs down to about $170 per kilogram, the model puts total non-computing costs near $15 billion — below the $25 billion on the ground.
The entire trajectory depends on Starship. If launch costs fall as projected, orbital data centers become competitive. If they do not, the economics do not work.
Google’s research reaches a similar conclusion from a different angle. Its Suncatcher paper estimates that achieving $200 per kilogram launch costs — the threshold for commercial viability — would require approximately 370,000 tons of cumulative payload delivered to orbit, equivalent to about 1,800 Starship launches at 200 tons per launch. At an average of 180 Starship launches per year, that target could be reached in a decade.
The physics of cooling in space adds another constraint. Unlike ground-based data centers, orbital computing hardware cannot rely on air or water cooling. Heat must be dissipated through radiators, which adds mass — and mass adds launch cost. As one industry executive put it, heat dissipation is “one of the biggest technical obstacles” to scaling orbital computing.
Two Approaches, Two Timelines
The contrast between the Chinese and American approaches is not about technology. It is about sequencing.
China is building orbital computing capacity incrementally, using existing launch vehicles and commercially available components. The Tiansuan constellation uses 16 satellites and seven ground stations. The Three-Body Computing Constellation has 12 satellites with 5 POPS of combined capacity. The Supercomputing-1 satellite carries a 4-meter camera and an AI computer. None of these systems require Starship. None of them require a new launch paradigm. They are operational today.
The American approach is building for scale from the outset. Starcloud‘s target is 88,000 satellites and 20 GW. SpaceX’s application covers 1 million satellites. Google‘s Suncatcher model is designed around a future where Starship launches are measured in hundreds per year. These are not incremental deployments. They are infrastructure programs that require a launch ecosystem that does not yet exist.
The Chinese approach has a working system now, with modest capacity, that can be scaled as launch costs fall. The American approach has a plan for massive capacity later, contingent on a launch cost trajectory that has not yet been demonstrated.
Both approaches have merit. Incremental deployment generates data, refines operations, and proves the business model. Large-scale planning avoids the problem of stranded assets if the technology evolves faster than the infrastructure. But the timelines are different, and the risk profiles are different.
China’s orbital computing cloud is already serving users. It has processed gamma-ray bursts, controlled ground robots, and supported research from 60 institutions. It has done so with satellites that launched on existing rockets, using chips that are commercially available.
The American orbital data center is not yet serving commercial customers. It is raising capital, building manufacturing capacity, and waiting for Starship.
What the Race Reveals
The orbital computing race is not a competition to build the first satellite with a GPU. Both countries have done that. It is a competition to build the first orbital computing infrastructure that can be used, scaled, and sustained.
China is deploying infrastructure. The Tiansuan constellation is offering regular services. The Three-Body constellation is running AI models in orbit. The Supercomputing-1 satellite is processing imagery in real time. These are operations rather than experiments.
The United States is building toward infrastructure. Starcloud’s H100 demonstration was a technical milestone. Its funding round reflects confidence in a future market. But the company’s 88,000-satellite constellation is a plan, not a deployment. SpaceX‘s AI1 is a design, not a satellite. Google’s Suncatcher is a validation mission, not a data center.
The difference is not capability. It is sequencing. China built the first orbital computing node with off-the-shelf technology and existing rockets. America is building the launch infrastructure required to deploy orbital computing at a scale that has not yet been attempted.
Both strategies have risks. China‘s incremental approach could be overtaken if Starship delivers on its cost promises and American companies can deploy capacity orders of magnitude larger. America’s scale-first approach could be stranded if launch costs do not fall as projected, or if the demand for orbital computing does not materialize at the scale its business models assume.
What is already clear is that the race is not theoretical. China has working orbital computing capacity today. The United States is building the infrastructure required to deploy it at scale.
Sources: Beijing Municipal Government (September 21, 2026); 21st Century Business Herald (September 21, 2026); TMTPost (September 23, 2026); China.org.cn via Xinhua (September 1, 2026); China Daily (October 6, 2026); Wall Streetcn via Deutsche Bank Research (July 14, 2026); 163.com via cnBeta (October 2, 2026); Data Centre Magazine (August 24, 2026); Interesting Engineering (August 23, 2026); C114 (July 22, 2026); China News Service (July 21, 2026).
Disclaimer
The information provided in this article is for general informational and educational purposes only. It does not constitute legal, financial, or professional advice. The author and publisher are not responsible for any actions taken based on the content of this article. Readers should consult qualified professionals for advice specific to their situation. All trademarks and references to third-party products, services, or organizations are the property of their respective owners. The performance data and benchmarks discussed are based on specific research studies and may not generalize to all use cases or environments. As of the publication date, the AI landscape continues to evolve rapidly, and readers should verify current information independently.
Limitations
This analysis is based on reporting and public data available as of the article date; figures may be revised as sources update.
Cost projections come from Deutsche Bank's published model as reported in the press; the full model is not public.
Launch cadence, cost and performance figures are company targets, not demonstrated results.
Details of Chinese constellations come from state media and developer announcements.
Comparative claims rely on public reporting, not audited operational data.
Sources
- Beijing Municipal Government (September 21, 2026)
- 21st Century Business Herald (September 21, 2026)
- TMTPost (September 23, 2026)
- China.org.cn via Xinhua (September 1, 2026)
- China Daily (October 6, 2026)
- Wall Streetcn via Deutsche Bank Research (July 14, 2026)
- 163.com via cnBeta (October 2, 2026)
- Data Centre Magazine (August 24, 2026)
- Interesting Engineering (August 23, 2026)
- C114 (July 22, 2026)
- China News Service (July 21, 2026).
The information provided in this article is for general informational and educational purposes only. It does not constitute legal, financial, or professional advice. The author and publisher are not responsible for any actions taken based on the content of this article. Readers should consult qualified professionals for advice specific to their situation. All trademarks and references to third-party products, services, or organizations are the property of their respective owners. The performance data and benchmarks discussed are based on specific research studies and may not generalize to all use cases or environments. As of the publication date, the AI landscape continues to evolve rapidly, and readers should verify current information independently.
Limitations: This analysis is based on reporting and public data available as of the article date; figures may be revised as sources update.; Cost projections come from Deutsche Bank's published model as reported in the press; the full model is not public.; Launch cadence, cost and performance figures are company targets, not demonstrated results.; Details of Chinese constellations come from state media and developer announcements.; Comparative claims rely on public reporting, not audited operational data.