Everyone agrees that data centers are eventually going to space. Very few companies can prove they have already put working hardware there. On September 11, 2026, at the space computing forum of the China Computing Power Conference, Beijing startup StarDetect (星测未来) made a claim that is unusually easy to verify: 37 advanced-process GPUs are running stably in orbit right now, backed by 39 operating payloads, and the company has just closed two funding rounds worth hundreds of millions of yuan to build China’s first cloud-collaborative space computing constellation, dubbed the “Fuyao Project.”
The announcement matters less as a funding headline than as a reality check. Orbital data centers have become the hottest narrative in hard technology in 2026, propelled by Elon Musk’s claim that AI could account for 99% of SpaceX’s value within four or five years. In that gold rush, PowerPoint constellations vastly outnumber launched ones. StarDetect’s numbers offer a rare measuring stick for what actually exists above the atmosphere and how close China is to turning orbital compute into a business.
What StarDetect Actually Announced
The company revealed completion of two consecutive funding rounds, A+ and A+, totaling hundreds of millions of Chinese yuan. Investors included Bei Venture Capital, China Creation Ventures, SunShine Capital, and NewDing Capital, with existing backers Wuxi Venture Capital and Houtian Capital increasing their stakes. The speed of the raise is the telling detail: StarDetect collected more capital in six months than in three times the total of the five rounds it completed over its first five years.
Behind the money sits an operating record assembled quietly since 2020:
- 37 advanced-process GPUs deployed and running stably in orbit, with 39 payloads in active service.
- First company in China to apply 12nm GPUs on orbit (February 2022), and the first anywhere in the world to fly a 7nm GPU in space (2024).
- A single computing device that has operated continuously in orbit for more than four years.
- An order book exceeding 100 million yuan, with revenue expected to double in 2026.
- More than 100 additional computing platforms reserved for launch, which would add up to 10 POPS of on-orbit capacity.
StarDetect was founded in Beijing in 2020 by three Tsinghua University doctoral graduates, CEO Cang Jirong, Cao Dezhi and Pan Xingyu, who all met working on Tsinghua’s “TianGe Project,” a student-led deep-space science mission that in October 2018 flew the world’s first gamma-ray burst detection payload on a micro-nano satellite. Their founding insight was simple: satellites are only as valuable as their ability to process data themselves. As 36kr reported in its exclusive profile, Cang has spent six years arguing that space computing power “should not be a concept that only stays on PPT, but a landed on-orbit productivity.”
Why This Is More Than Another Space-Compute Press Release
To understand why 37 GPUs is a meaningful number, you have to understand how noisy this market has become. Musk broke SpaceX’s valuation ceiling with the “AI plus space” logic and declared that the main battlefield for AI inference will eventually be in orbit. China absorbed and amplified that narrative fast. Startups have poured into the sector, and capital flows continuously: Zhongke Tiansuan is raising hundreds of millions of yuan toward a plan for 10,000 domestic GPUs in orbit by 2030, Orbital Chenguang has secured roughly $8.4 billion in credit lines from state-linked banks targeting gigawatt-scale orbital compute by 2035, and Zhejiang Lab’s Three-Body Computing Constellation already has a working cluster in low Earth orbit.
Independent analysis of China’s orbital compute sector, including a 2026 reality check by space industry consultancy SpacesGP, describes it as an early validation phase defined by a small number of genuine in-orbit facts surrounded by a large amount of financing narrative. That is precisely the gap StarDetect is trying to occupy. The company took the same route as Starlink: instead of expensive, slow-iterating, closed-ecosystem radiation-hardened chips, it uses commercially available ground GPUs and buys reliability with engineering, betting that iteration speed beats physical armor.
Is an orbital data center actually possible today?
Partially, and the word “data center” still oversells it. What exists today is real but modest: Starcloud’s satellite demonstrated in-orbit training of a small GPT-class model after its November 2025 launch, Zhejiang Lab’s constellation is running an 8-billion-parameter model across its satellite cluster, and StarDetect’s fleet performs real-time inference on satellite imagery and infrared scans. What does not exist yet is the gigawatt-scale orbital server farm implied by the marketing. The hard constraints remain launch cost, power generation, radiators for heat rejection, and the brutal economics of keeping hardware alive in a radiation environment for years. As we detailed in our earlier analysis of the three real bottlenecks facing data centers in space, none of those problems have been wished away by the hype cycle.
How Consumer GPUs Survive Space Radiation: A Three-Layer Defense
The core technical bet at StarDetect is worth explaining, because it inverts decades of aerospace orthodoxy. The traditional approach to radiation is to resist it physically, using specially fabricated chips that fight particle strikes at the silicon level. That works, but those chips are costly, slow to improve, and incompatible with the mainstream GPU software ecosystem. StarDetect’s alternative lets chips make mistakes and builds a system smart enough to catch and repair them before a fault becomes a mission failure. The defense stacks three layers:
- Structural shielding. Thickened structural components cut cumulative radiation damage to less than one thousandth of the unshielded level.
- System fault tolerance. Error-correcting code memory checks, redundant backups and watchdog timers handle transient faults such as single-event upsets, where a particle flips a bit in memory.
- Active avoidance. A small onboard radiation sensor monitors particle threat levels in real time and triggers escalating responses: self-checks, targeted repairs, or full power-down protection when conditions deteriorate. This third layer, drawing on Tsinghua’s engineering physics heritage, is the piece StarDetect claims as unique.
StarDetect protects commercially available GPUs in orbit with three stacked defense layers rather than expensive radiation-hardened chips. (Credit: Intelligent Living)
That theory was tested in orbit, between 500 and 630 kilometers altitude, where the company’s team watched high-energy particles scramble storage devices until hard disks stopped being recognized, then recovered them by re-uploading software. Other single-event faults caused current anomalies and functional failures, fixed with power-cycle restarts and watchdog resets that restored tasks within minutes. “When no errors occur, we even doubt whether it is necessary to make so many redundant designs,” Cang told 36kr. “Until we experience problems, we find that many previous designs are valuable.”
Who Pays: From Hardware Payloads to a 100M RMB Order Book
Space computing pitches usually stop at technology. StarDetect’s more interesting claim is commercial traction, and the named customers are Chinese industrial heavyweights rather than research experiments.
- State Grid Corporation of China. StarDetect built the intelligent computing and algorithm platform for a power infrared satellite that launched in February 2026. Onboard processing identifies fire points and pins their locations from infrared imagery in under 100 milliseconds per scene, fast enough to matter for grid infrastructure monitoring and wildfire response.
- CCCC Water Transportation Consultants. Satellite remote sensing combined with on-orbit computing tracks congestion in busy waterways, extended into major engineering monitoring and global ship-tracking enhancement.
- Hainan Satellite Data and Application Research Center. A strategic partnership aimed at an international intelligent remote sensing constellation with a space-ground integrated sensing and computing platform.
The business model has also matured away from one-off hardware sales. StarDetect no longer simply ships payloads and walks away; after launch its delivery teams keep processing data and tuning algorithms alongside customers. That recurring-service structure underpins an order book above 100 million yuan and a planned expansion into what the company calls tokenized operation of space computing power, essentially selling orbital compute as a metered service rather than a widget.
China Versus the US in the Orbital Data Center Race
The comparison with American rivals is unavoidable and instructive. Starcloud, backed by Nvidia, raised $250 million in August 2026 at a $2.3 billion valuation and is preparing a Blackwell B200-class satellite for late 2026, as SpaceNews reported. The US companies lean on cutting-edge silicon and hyperscaler partnerships. China’s cohort, constrained by export controls on advanced chips, has responded with a different mix: state-aligned financing at enormous scale, domestic silicon roadmaps, and, in StarDetect’s case, an early-mover advantage in raw flight hours.
That produces two genuinely different strategies. The American approach maximizes performance per satellite, betting that spectacular chips make orbit worthwhile. The Chinese approach maximizes proven operations and unit economics, betting that a commercial loop closed on Earth today, however modest, compounds into something larger than any single launch. StarDetect is the clearest test case for the second strategy: no other company has publicly put 7nm silicon through years of continuous orbital service, and none has turned on-orbit processing into double-digit-million-yuan industrial contracts across power grids and shipping lanes.
What Comes Next: The Fuyao Constellation
The Fuyao Project is where StarDetect’s bet gets ambitious. The company’s six-year strategy compresses into four words: compute space data in space. Doing that at constellation scale requires three preconditions that only recently aligned: hardware reliable enough to survive unattended, enough orbital data to make processing worthwhile, and a maturing software ecosystem that lets large and small models run on satellites and coordinate with each other. Fuyao is designed to be China’s first constellation where computing payloads operate as a cloud-edge collaborative network rather than as isolated clever satellites.
The company frames this as the first step of a longer arc, from computing space-generated data in orbit to eventually processing ground data from space. If that sounds grandiose, the counterweight is the operating record: 37 GPUs that have not blinked, contracts that pay real money, and a team that survived its own cash-flow winters before the sector became fashionable. “We have almost been fighting against the wind all the time,” Cang said. “It is not until this year that we start to catch the favorable wind.”
For anyone tracking whether orbital data centers are a decade away or a year away, StarDetect’s answer is specific: the compute is already up there, quietly doing work somebody pays for. The rest of the industry still has to catch up to that sentence.
