On 14 May 2025, a Long March 2D rocket lifted off from the Jiuquan Satellite Launch Centre in the Gobi Desert at 12:12 Beijing time. According to Xinhua and CCTV, it carried 12 small satellites designed not to take pictures or relay phone calls but to compute. Each could process data on board, link to its neighbours by laser and run an artificial intelligence model in orbit. Its builders called it the Three-Body Computing Constellation, after Liu Cixin’s science-fiction trilogy.
Ten months later, on 21 March 2026, Elon Musk stood on a stage in Austin, Texas, and announced Terafab, a chip factory meant to produce more computing power every year than the entire semiconductor industry makes today, most of it destined for data centres in space. Taken together, the two events show that computing in orbit has become a question of industrial policy on both sides of the Pacific. This article looks at what China has built, what it plans, including a Beijing plan for a gigawatt-scale space data centre and state backing in the new five-year plan, how its state media greeted Musk’s Terafab, and why chips, export controls and launch capacity will decide how far China is involved in what comes next. Space computing attracts very large numbers, and many of them are targets, filings or estimates rather than hardware in orbit, so the article tries to keep what has been reported separate from what is analysis.
What Terafab actually is
This site has already covered Terafab as a chip-manufacturing story, so a short summary will do. Musk presented it at the Seaholm Power Plant in Austin as a joint project of Tesla, SpaceX and xAI, the AI company that SpaceX had absorbed in an all-stock deal announced on 2 February 2026, which valued the combined group at $1.25 trillion. xAI was later rebranded SpaceXAI. Terafab is meant to bring chip logic, memory, advanced packaging and testing under one roof.
The stated goal is one terawatt of computing capacity a year. Tesla put it bluntly: “That’s more than all the chip manufacturers in the world combined can provide today, or even by 2030.” According to Manufacturing Dive’s account of the presentation, about a fifth of the output is meant for uses on Earth, such as Tesla’s self-driving systems and Optimus robots, and four-fifths for space. SpaceX’s own update describes the full-scale plant, planned for Grimes County, Texas, as producing “high-power chips designed for operating SpaceX’s space-based data centers”. Reuters reported in May 2026 an initial investment of $55 billion, rising to as much as $119 billion across all phases, and in August Texas Governor Greg Abbott announced a first phase worth $16.8 billion, with construction due to finish in 2028.
Terafab is the factory end of a bigger plan. On 30 January 2026, SpaceX asked the US Federal Communications Commission to approve a “SpaceX Orbital Data Center system” of up to one million satellites at altitudes of 500 to 2,000 km, linked by lasers. The filing called it a “first step towards becoming a Kardashev II-level civilization – one that can harness the Sun’s full power”. The FCC accepted it for filing on 4 February, though SpaceNews noted that it contained no schedule or cost estimate. When announcing the xAI merger, SpaceX argued that the cheapest place to do AI computation “within two to three years will be in space”.
China’s first computing constellation
China’s effort predates Terafab and already has hardware in orbit. The Three-Body Computing Constellation (三体计算星座) is a project of ADA Space, known in Chinese as Guoxing Aerospace (国星宇航), a commercial satellite company in Chengdu, working with Zhejiang Lab, a research institute in Hangzhou. Its builders said the 12 satellites launched in May 2025 could each perform up to 744 trillion operations per second (TOPS), with a combined 5 peta-operations per second (POPS), 30 terabytes of storage, laser links of up to 100 gigabits per second and an eight-billion-parameter AI model on board. Some secondary reports give different figures; a Global Times article in May 2026 referred to “50 petaflops”, which does not match the official 5 POPS, and figures for low-precision AI operations are not the same as floating-point measures.
The first aim is not to move Earth’s data centres into space but to deal with data that satellites already generate. Li Chao of Zhejiang Lab told a conference in January 2026 that about 90% of data produced in space is not effectively processed, according to CGTN. Most observation satellites store images until they pass over a ground station, which can take hours. A satellite that analyses its own images and sends down only the answer saves bandwidth and time. Xie Lina of the China Academy of Information and Communications Technology (CAICT), a think tank under the Ministry of Industry and Information Technology, told Global Times that this could cut response times from hours to seconds.
Progress since launch has been steady. In February 2026, Science and Technology Daily reported that six satellites had been networked in orbit, ten AI models deployed, and that the system could support in-orbit inference for models of up to 140 billion parameters. Zhejiang Lab’s director, the academician Wang Jian, summed up the rationale in a phrase widely repeated in Chinese coverage: “人工智能不能因为缺失算力而缺席太空”, roughly “AI cannot be absent from space for lack of computing power”. Two further satellites went up on 24 July 2026 carrying what was described as an embodied-intelligence model.
Guoxing says it ran Alibaba’s Qwen3 large language model in orbit in November 2025, with a full question-to-answer loop from the ground taking under two minutes. CGTN later gave the month as January, so the date is uncertain, but both accounts agree that a commercial Chinese language model has run on a satellite. I could not find any source publicly identifying the chips inside the satellites, and it would be a mistake to assume they are either Western or Huawei parts.
From 12 satellites to thousands
The targets sitting on top of this modest hardware are much larger. Zhejiang Lab aims for 1,000 satellites with 1,000 POPS of combined computing power, which it expects to reach by around 2032. Near-term numbers differ: Wang Jian told Global Times in April 2026 that more than 50 computing satellites would be deployed in 2026, while CGTN reported in February that 39 were under development, with 100 targeted by 2027. The publicly reported launches so far add up to 14.
Guoxing is also planning a commercial network called Star-Compute (星算). According to IT之家 and Jiemian in January 2026, it would have 2,800 satellites, 2,400 for inference and 400 for training, at 500 to 1,000 km. Two further groups were in production for 2026, with 1,000 satellites in commercial operation by 2030 and completion by 2035. The company holds International Telecommunication Union (ITU) approval for frequency and orbit resources covering 3,145 satellites and is pursuing a Hong Kong listing. A filing and a production plan are not a constellation, and much depends on how many launches the company can buy.
A third project belongs to Beijing’s municipal government. On 28 November 2025, Xinhua reported that the Beijing Municipal Science and Technology Commission and the Beijing Astro-future Institute of Space Technology (北京星辰未来空间技术研究院) planned a space data centre in a dawn-dusk orbit at 700 to 800 km, which follows the line between day and night so that solar panels stay almost permanently lit. The plan calls for power capacity above the gigawatt level and sub-centres hosting clusters of millions of AI accelerator cards, built in three phases and reaching large scale by 2035. A consortium of 24 organisations was formed to develop it.
The first hardware, a test satellite called Chenguang-1 (辰光一号), was launched from Jiuquan on 24 July 2026, according to Beijing News. It carries space concentrator solar cells and flexible graphene radiative cooling panels, aimed at the two main engineering problems of orbital computing: generating enough power and getting rid of the heat. By then the consortium had grown to nearly 30 members. Some online reports attach specific computing and power figures to the project that I could not trace to an official source, so they are not repeated here.
The state steps in
What began with a provincial laboratory and a start-up has been drawn into national planning. On 29 January 2026, the China Aerospace Science and Technology Corporation (CASC), the state-owned prime contractor for China’s space programme, set out five major projects for the 15th Five-Year Plan period (2026 to 2030). According to Xinhua, one was to build “吉瓦级太空数智基础设施”, gigawatt-class space digital-intelligence infrastructure, integrating computing, storage and transmission so that both space data and ground data could be “computed in space”. Reuters reported this under the headline “China plans space-based AI data centres, challenging Musk’s SpaceX ambitions”, adding that a CASC policy document from December 2025 envisaged solar-powered hubs forming an industrial-scale “Space Cloud” by 2030.
The Ministry of Industry and Information Technology (MIIT) has also stepped in. Global Times reported in April 2026 that the ministry would accelerate a space computing industry ecosystem, and quoted an MIIT official, Zhao Ce, on the need for systematic planning. China’s first professional committee for space computing was set up to cover on-board AI chips, laser links, thermal control and space photovoltaics, and CAICT and others issued an initiative on building a “computing satellite network”. The national 15th Five-Year Plan calls for an integrated space-air-ground network combining communication, navigation, sensing and computing.
Part of the motive is energy. Global Times, citing official data, reported that China’s computing centres used 196 billion kilowatt-hours of electricity in 2025, up 18.1% on the year. The same pressure is driving American interest: orbit offers near-constant sunlight and no local planning objections. The telecoms analyst Xiang Ligang told Global Times that “the future of the human information industry will inevitably be an integrated space-air-ground system”. Analysts quoted in Chinese media have floated a market worth $1 trillion by 2035 and operating costs of about 5% of ground-based centres. These are projections, not established facts.
How China responded to Musk
As far as I can find, the Chinese government has not responded to Terafab specifically. I found no statement from the Ministry of Foreign Affairs, MIIT or any other government body addressing it by name. What exists is state-media commentary, and it is revealing.
On 26 March 2026, five days after the Austin presentation, Xinhua published a news analysis by its Los Angeles correspondent Tan Jingjing, which People’s Daily online republished the next day under the headline “马斯克的大型芯片项目Terafab落地不易”, or “Musk’s big chip project Terafab will not be easy to realise”. It noted that advanced chipmaking depends on extreme ultraviolet lithography machines from ASML, with lead times of more than a year. It quoted the Bernstein analyst Stacy Rasgon saying the project would be harder than sending rockets to Mars, and cited estimates from Morgan Stanley (about $45 billion for a fab producing 100,000 advanced wafers a month) and UBS (about $30 billion for the initial target). It also pointed to delays in SpaceX’s Starship and noted that some observers saw the announcement as supporting valuations. Its conclusion, in my translation, was that Terafab reflects Musk’s vision of fusing AI and spaceflight, but must still overcome technical thresholds, long construction times, capital needs and the question of commercial returns.
The tone was sceptical rather than hostile. The piece did not frame Terafab as a threat to China or mention China’s own space computing projects; it argued that chipmaking is hard, a lesson Chinese industry has learned under American export controls. It should be read as the view of an official news agency, not as government policy.
China’s broader response to SpaceX has been sharper. At an informal Arria-formula meeting of the UN Security Council on 29 December 2025, China’s representative criticised “the unchecked proliferation of commercial satellite constellations by a certain country”, saying Starlink, with more than 10,000 satellites, was “seriously occupying frequency-orbit resources”, had twice made dangerous close approaches to China’s space station and blurred military and civilian roles. In late December 2025, Chinese entities also filed with the ITU for about 203,000 satellites across 14 constellations, according to SpaceNews and the South China Morning Post, including two constellations, CTC-1 and CTC-2, of 96,714 satellites each. Such filings reserve a place in the queue for scarce spectrum and orbital slots; they are not launch approvals. None was described as a response to orbital data centres, and linking them to Terafab would go beyond the evidence.
The physics and economics problem
Both countries face the same physics. In a vacuum, heat can only be shed by radiation, so a computer drawing a megawatt needs large radiator panels to dump roughly a megawatt of heat; hence the graphene cooling panels on Chenguang-1. Cosmic rays flip bits and damage chips, so hardware must be shielded or designed to tolerate errors, at a cost in performance or mass. And every kilogram of chips, panels and radiators must be launched and eventually replaced. CGTN itself lists thermal management, radiation-hardened chips and launch cost as the key challenges, and reports estimates for when orbital computing could match ground-based costs: around 2030 according to the American start-up Starcloud, around 2032 according to Deutsche Bank. Both depend on launch prices that have not yet been achieved.
The American comparison helps. Starcloud’s first satellite, launched in November 2025, carried a single Nvidia H100 and was reported by CNBC to have trained a small language model in orbit. Google’s Project Suncatcher plans two prototype satellites with Planet by early 2027. China’s constellation is ahead in the number of computing satellites networked in orbit, which is why MIIT describes China as the first country to achieve in-orbit networking of a computing constellation. But nobody on either side has yet put anything close to data-centre scale computing into space.
Chips, controls and launchers
Three documented constraints shape how far China can go.
Chips
Since October 2022, the United States has restricted exports of advanced AI chips and chipmaking tools to China, tightening the rules in 2023 and 2024; this site has covered how China is building a chip industry around them. President Trump approved exports of Nvidia’s H200 in December 2025 in exchange for a 25% share of sales, but Beijing has rationed imports itself: Tom’s Hardware reported in August 2026 that ByteDance and Tencent had each received roughly 10,000 H200s under individual approvals from the National Development and Reform Commission. Huawei set out a roadmap in September 2025 for Ascend 950 chips in 2026, followed by the 960 and 970. It is barred from using TSMC and is believed to rely on SMIC, and industry estimates put its output at around 750,000 AI chips in 2026, shared across every Chinese AI company. Any Chinese space data centre will therefore be built largely on Chinese chips that are already scarce on the ground.
Controls on space technology
Space cooperation is restricted even more tightly. Since 2011, the Wolf Amendment in US appropriations law has barred NASA from bilateral cooperation with China without certification to Congress. American satellites and many components are controlled under the International Traffic in Arms Regulations (ITAR) or the Export Administration Regulations, and China is among the countries to which defence exports are generally denied. American companies do not launch on Chinese rockets or share satellite technology with Chinese firms.
Launchers
China conducted a record 93 launches in 2025, according to official figures cited by Reuters, and reached its 65th launch of 2026 on 17 September. Reusability arrived late. CASC’s Long March 10B had its booster caught by cables on an offshore platform on 10 July 2026, and on 19 August LandSpace landed the first stage of its methane-fuelled Zhuque-3 on legs in Gansu, China’s first land-based recovery of an orbital-class booster, according to Spaceflight Now. (This site has looked at China’s commercial launch sector in more detail.) SpaceX had flown a reused booster for the 614th time the day before. China’s internet constellations show the gap: Guowang had 204 satellites in orbit after a launch on 17 September 2026, against a 2026 target of 310, according to China in Space, and Qianfan had around 250, against Starlink’s more than 10,000. Computing satellites have to compete with them for launch slots.
How involved will China be
This section is analysis. It rests on the facts above, but it is a judgement about the future.
In Terafab and SpaceX’s orbital data centres, China’s direct involvement is likely to be close to zero. US export controls, ITAR, the Wolf Amendment and the security politics around SpaceX, a major defence contractor, make Chinese participation in design, manufacturing, launch or operation hard to imagine. Beijing, for its part, has barred state-funded data centres from foreign accelerators, according to Tom’s Hardware, and is unlikely to want Chinese data processed on an American constellation.
The one plausible point of contact is the supply chain. China dominates global production of solar cells. Reuters, CNBC and the South China Morning Post reported in February 2026 that teams linked to SpaceX and Tesla had visited Chinese photovoltaic manufacturers, including JinkoSolar, TCL Zhonghuan and GCL, with interest in heterojunction and perovskite technology. JinkoSolar confirmed contact but denied any cooperation or orders. That suggests an American orbital compute effort may find Chinese industrial capacity hard to avoid entirely, but US tariffs, forced-labour import rules and pressure on defence suppliers all push the other way. Any Chinese contribution is likely to be indirect and heavily scrutinised.
In orbital computing as a field, China is likely to be one of the two main players, running a parallel system. It has a networked computing constellation in orbit, a municipal gigawatt-scale plan with a test satellite flying, CASC’s five-year commitment, MIIT’s backing and a national professional committee. The pattern is familiar from Musk’s wider relationship with China and from navigation and broadband: where the United States builds something strategic, China tends to build its own version, as it did with BeiDou and is doing with Guowang and Qianfan.
The limits on China’s scale are chips and launches, not ambition. If Chinese chipmakers cannot meet demand on the ground, sending large numbers of accelerators into orbit will be a hard choice, and Chinese space computing may stay focused on processing satellite data, where the payoff is clearest. If Starship reaches its promised reliability and cost, SpaceX will have lift capacity China’s fleet cannot match; if it keeps slipping, the gap narrows. China’s booster landings in 2026 make cheaper launches more plausible, but turning two successful recoveries into routine reuse will take time.
China may lead in early demonstrations while the United States leads in potential scale. Neither SpaceX’s million-satellite filing nor Beijing’s gigawatt centre has a firm, costed public schedule, and both depend on unsolved engineering problems, notably getting rid of heat at scale. The contest may be decided less by who announces the biggest number than by who first shows that computing in orbit can be cheaper than computing on the ground.
Conclusion
China’s orbital computing programme is real, small and growing. Fourteen Three-Body satellites have been launched, a Chinese language model has run in space, a Beijing test satellite is trialling the power and cooling technology a larger system would need, and the main state space contractor has written gigawatt-scale space computing into its five-year plan. The plans for thousands of satellites and gigawatt centres remain targets.
Terafab is a verified, if unfinished, project: a Tesla, SpaceX and SpaceXAI chip factory in Texas intended mainly to supply orbital data centres. China’s documented reaction came from its official news agency, in a measured analysis arguing that the factory will be very hard to build, not from any government statement. Beijing’s wider response to SpaceX has come through UN diplomacy, spectrum filings and its own industrial policy.
The likeliest outcome is two separate orbital computing systems, one American and one Chinese, with little direct Chinese role in the American one beyond, possibly, solar technology and materials. How large China’s version becomes will depend on how many chips its foundries can make and how quickly its reusable rockets become routine, questions that will be settled in factories and on launch pads rather than in announcements.
