Compute Above the Weather: The Week a 1,000-Satellite AI Data Centre Started Going Up
Six weeks ago I wrote about a land rush. More than a million orbital compute satellites had been filed with regulators on the American side, against roughly ten thousand actually operating. Paper, in other words, at a scale no one has ever deployed. This week the same race grew a second front, and this one is not paper. It is hardware, and it is already in orbit.
On Saturday 18 July 2026, a Shanghai company that most people outside the industry had never heard of put the opening pieces of a thousand-satellite plan into space. Shanghai Xingshu Tiansuan Space Technology, working with Fudan University, launched the verification phase of what it calls the Xingshu Plan: two computing satellites and twelve edge-computing satellites, built to process artificial intelligence and remote-sensing workloads in orbit and send back only the results, not the raw data. The launch landed in the same week that President Xi Jinping was on stage at the World Artificial Intelligence Conference in Shanghai, positioning China as a shaper of global rules for artificial intelligence.
The headline number, one thousand satellites, is real but distant. What went up this week is fourteen satellites in a verification phase. The company has described a path from there through a commercial-deployment phase of fifty computing and one hundred edge-computing satellites, to roughly a thousand at full scale. That is a plan, not a constellation. The significance of the week is not the count on day one. It is that Xingshu is now the newest of several distinct Chinese ventures pursuing the same idea, at national scale, timed deliberately to a state showcase on the governance of the technology itself.
The launch that actually flew
Strip away the framing and the engineering claim is specific and consistent across every source that reported it. Rather than beaming raw sensor feeds and model workloads down to Earth for processing, the Xingshu satellites are designed to run the computation in orbit and downlink only the derived answer. The stated benefits are less communication delay and less strain on the ground segment, because a processed result is a tiny fraction of the size of the data it came from. Move the compute to the sensor, and you stop paying to move the sensor's output.
That is the architecture the whole field is converging on, and it is worth being precise about what has been verified. Two computing satellites and twelve edge-computing satellites are confirmed in orbit, corroborated across separate Chinese and Western reporting. The company's own account of the deployment timeline, roughly two hundred satellites by 2030 and eight hundred by 2032, comes from its conference presentation and was not independently confirmed by a second source this week, so I am giving it to you as the company's stated intention rather than a fact. One wire report described a different phase structure again, with numbers that do not match and do not add up to a thousand; because a similarly named but separate company also exists, that report may be a conflation, and I have set it aside rather than blend it in.
This is not the Three-Body Constellation
Here is the point that matters most, and the one most likely to be got wrong. Xingshu Tiansuan is not the Three-Body Computing Constellation. Regular readers of this series, and readers of Space Mafia, know Three-Body as the standing Chinese case study: a twelve-satellite network launched in May 2025 by ADA Space, the trading name of Chengdu Guoxing Aerospace Technology, in partnership with Zhejiang Lab. Each of those satellites carries computing hardware rated at 744 trillion operations per second, for a combined total of five peta-operations per second, linked by laser connections and running an onboard model of eight billion parameters for astronomical observation.
That is a real, flying, fourteen-month-old system, and it is a different company, different satellites, and a different architecture from this week's launch. Xingshu is new. Three-Body is the incumbent. Treating the two as one, or reading Xingshu as an update to Three-Body, is the single accuracy error this story invites, and it changes the meaning entirely. The news is not that China's one orbital-compute demonstrator grew. It is that China now has a competitive field of them.
Four programmes, one governance week
Count the runners and the picture sharpens. ADA Space, the Three-Body operator, is expanding that constellation under a programme it calls Star-Compute, aiming at an eventual 2,800 satellites split between inference and training work, and it filed for an initial public offering on the Hong Kong Stock Exchange in January 2026. That is the same company as Three-Body, further along than any other Chinese entrant, and already at the capital-markets stage. Separately, SenseTime has partnered with the same manufacturer to launch four computing satellites of its own this year under its own brand, with its own stated long-term ambition. A fourth, older effort, a commercial spin-off tied to the Chinese Academy of Sciences, has been described in wire reporting as pursuing a near-Earth processor cluster; its specific figures rest on a single source and I am flagging it as background rather than fact.
The temptation is to collapse all of this into a single sentence: China is building orbital compute. Resist it. These are separate commercial companies with different customers and different technical bets, competing inside a country whose government has made a public point of wanting to lead on the governance of artificial intelligence. That is closer to a genuine competitive market than a single state programme, even though all of the runners benefit from a national environment that treats orbital compute as strategically important. The state did not launch these satellites. It provided the stage on which their launch was announced.
The scoreboard: paper versus flying
Set the two sides of the race next to each other and the asymmetry is not the one the satellite counts suggest. On the American side, the announced filings are enormous: a roughly one-million-satellite regulatory filing following the merger of the world's dominant launch provider with a frontier artificial intelligence lab, and a separate eighty-eight-thousand-satellite filing from a company that has already flown a demonstration satellite carrying the most powerful accelerator yet sent to orbit. A third American programme plans to place its own processors on a solar-powered constellation, with a first in-orbit test targeted for 2027. By announced numbers, every Chinese plan is smaller.
But a regulatory filing is a paper commitment, not a deployed machine. What distinguishes this week's China news is that it is phased, funded, and flying. Fourteen Xingshu satellites are in orbit now; twelve Three-Body satellites have operated since May 2025; and ADA Space has reached a stock-market listing that none of the announced American mega-constellations has needed, because those are funded from existing balance sheets and private capital. The American lead is real on ambition and on the single most capable piece of hardware in orbit. The Chinese lead, such as it is, is on the unglamorous business of turning a plan into a company with satellites overhead and a prospectus filed.
Corporate and sovereign, at the same time
Space Mafia separates two patterns that usually appear apart. Corporate Kardashev is the accumulation of strategic orbital position by commercial actors faster than any government can write rules for them. Sovereign Kardashev is the state-capital version of the same move. The book's original cases tended to keep them distinct: private American firms on one side, sovereign wealth funds on the other. Xingshu collapses the distinction into a single actor. It is a commercial company, with a university as its research partner rather than a ministry, pursuing a build of national significance, launched to coincide with a state governance showcase. That is Corporate and Sovereign Kardashev in one entity, and it is a cleaner instance of the overlap than anything the book had to work with when it was written.
Whose law runs the computer?
The book's organising analogy is the pirate radio parallel: the 1960s broadcasters who ran from ships in international waters, beyond any nation's jurisdiction, until regulation caught up with them. In-orbit compute sharpens that analogy in a way the series has not explored before. If a satellite processes remote-sensing data as it passes over a third country's territory, and downlinks only the derived result, which nation's data-processing law governed the computation itself? Not the data at rest, and not the transmission, but the act of processing, performed in orbit, over territory that belongs to someone else. The raw feed may never touch the ground of any country whose law would have applied to it. Move the computation above the weather and you have moved it, at least arguably, above a good deal of the law as well.
This is the deeper reason the results-only design matters. It is sold as an engineering efficiency, and it is one. It is also a jurisdictional manoeuvre, whether or not anyone intends it as such, because the thing that would ordinarily anchor a legal claim, the data, is consumed and discarded in a place no court has yet reached.
What New Zealand holds this week
I will be honest about the New Zealand angle, because padding it would be worse than stating it plainly. This launch has no New Zealand company, no New Zealand ground station, and no New Zealand launch site. Unlike the last two Monday stories, both of which flew from Māhia, there is no physical thread running from this development back to the country. Inventing one would be dishonest. The New Zealand relevance here is regulatory and structural, not operational.
The one dated, genuine hook is regulatory. Under the Ground-Based Space Infrastructure regime created by the Outer Space and High-altitude Activities Amendment Act, operators of ground-based space infrastructure were automatically authorised during a transitional period. That period ended on 29 July 2026, five days before this article publishes. As of this week the authorisation requirement is no longer an approaching deadline; it is the enforced default. Every earlier article in this series that touched the Act described the deadline as upcoming. This is the first to describe the regime as fully in force. Operating without authorisation is now an offence carrying financial penalties, reaching into the hundreds of thousands of dollars for an entity, a figure this series has cited before and does not need to re-litigate here. What I will do is describe the mechanism and leave the assessment of it where it belongs, outside this article.
The structural point is about which side of the race New Zealand already sits on, and it is a description, not a recommendation. As a matter of existing commercial and security relationships, Rocket Lab's launch operations, membership of the Five Eyes intelligence partnership, and near-total dependence on United States-headquartered cloud and launch providers, the country sits inside the American and Western orbital sphere rather than the Chinese one. That is a fact about where the wiring runs, not a claim about where it should run.
None of this is new in kind, only in altitude. The Reserve Bank's own Financial Stability Report has already flagged the risk that a small number of overseas providers now sit underneath much of the artificial intelligence that New Zealand institutions rely on, a concentration that is hard to unwind once it is load-bearing. Orbital compute is that same dependency, formed earlier and higher. On the ground, a country can at least switch providers inside a single market. Between two orbital blocs that share no standards, there is no switch to throw. The lesson from the terrestrial layer is that these dependencies are easiest to shape before they harden, and orbital compute is still, just, at the soft stage.
The bifurcation is also a standards story, and that is where open source earns its place in the argument. What makes two compute stacks unable to talk to each other is not physics; it is the absence of shared, openly published interfaces. The terrestrial internet scaled because its addressing and routing rules were open and anyone could implement them. Orbital compute has no equivalent yet. Open ground-segment projects such as SatNOGS, open flight-software frameworks like NASA's core Flight System and the Jet Propulsion Laboratory's F Prime, and the open data-handling standards of the Consultative Committee for Space Data Systems all show that an interoperable path exists. None of the announced constellations, Chinese or Western, has committed to it. Two closed stacks built in parallel is exactly the outcome open standards were invented to prevent, and it is being chosen by default, one launch at a time.
There is a sovereignty layer worth naming plainly. If orbital compute is splitting into two blocs that do not interoperate, then any nation that comes to depend on processing in orbit is choosing a side, and the choice carries a lock-in measured in decades rather than budget cycles. New Zealand's existing arrangements answer the question of which side before it is asked. Rocket Lab runs its launch operations here, the country sits inside Five Eyes intelligence sharing, and it takes part in the Combined Space Operations Initiative, the multinational military framework through which the United States, the United Kingdom, Australia, Canada, France, Germany, Italy, Japan, New Zealand and Norway share space-domain information. Those are facts about where the connections already run. The open question, not a prescription, is what sovereign capacity a country keeps when the compute it depends on is built, and governed, somewhere else.
The bottom line
Two years ago orbital compute was a projection. This week it is a competitive field with hardware flying on both sides of a widening divide. The American side leads on ambition and on the single most capable machine in orbit. The Chinese side has quietly done the less visible work of getting satellites up, companies listed, and a national programme named. Neither side has agreed a single shared standard with the other, which means the most important thing being built right now is not a constellation. It is a fork. The internet became one network because its rules were open. Orbital compute is on course to become two, because so far nobody has chosen to open theirs.
When the compute that matters most has moved above the weather, and split into two stacks that cannot talk to each other, the question every small country will eventually have to answer is not which bloc is ahead. It is which one you are already standing in, and what you can still do about it. Where do you think that leaves us?
The views expressed in this article are entirely my own, informed by more than 30 years of professional experience in architecture, security, and technology leadership in New Zealand. They do not represent the views of my employer, any government agency, or the New Zealand government. My commentary on legislation and policy is analytical, drawing on publicly available sources and my professional expertise in architecture, security, and AI governance. I follow the Public Service Commissioner's Code of Conduct for the Public Sector and social media guidance.
Andreas Hamberger is a New Zealand leader in Architecture & Security and Associate Member of the Institute of Directors. Space Mafia examines the sovereignty implications of orbital compute infrastructure.
I use AI tools, including Sudowrite, Claude, Perplexity AI, DeepSeek AI, ChatGPT, Grok, Copilot, Openart and Gemini, as deliberate production tools, not ghostwriters. This is consistent with my position: AI amplifies human judgement; it does not replace it. The frameworks, arguments, and editorial decisions in this series are original work. AI accelerated the process. The thinking is mine.
References
[1] China Daily. "China launches verification phase of Xingshu Plan space-based computing network." 18 July 2026.
[2] Global Times. "Space-based computing constellation launched; WAIC future-computing-power forum." 18 July 2026.
[3] eWeek. Aminu Abdullahi. "China's Xingshu Plan advances orbital AI computing." 21 July 2026.
[4] Reuters Science (wire). "Chinese company launches first satellites of space-based computing constellation." 18 July 2026.
[5] The News International (thenews.com.pk). Pareesa Afreen (AP-origin wire). "China begins building a space-based computing constellation." 18 July 2026.
[6] China.org.cn. "Verification phase of space-based computing network launched." 18 July 2026.
[7] Data Center Dynamics. "China's Three-Body Computing Constellation and the Star-Compute expansion (ADA Space / Guoxing Aerospace)." 2025 to 2026.
[8] SpaceNews. "China launches first of 2,800 satellites for AI space computing constellation." 2026.
[9] Zhejiang Lab. "Space Computing Constellation 021 (Three-Body Computing Constellation)." en.zhejianglab.com
[10] Introl. "The Orbital Data Center Race 2026 (SpaceX and xAI, Starcloud, Google Project Suncatcher context)." 2026; KrAsia; TechTimes.
[11] Ministry of Business, Innovation and Employment. "Ground-based space infrastructure: licensing and the transitional period ending 29 July 2026." 2026.
[12] New Zealand Defence Force. "Combined Space Operations (CSpO) Initiative Principals' Board Meets to Advance Combined Operations Efforts." 2026. nzdf.mil.nz/media-centre/news/combined-space-operations-cspo-initiative-principals-board-meets-to-advance-combined-operations-efforts/
[13] Reserve Bank of New Zealand. "Financial Stability Report." May 2026. (Concentrated dependence on a small number of overseas artificial intelligence providers, referenced as supporting context.)