Orbital Compute: One Side Opened a District, the Other Announced a Partnership
On 31 August 2026, Shanghai's Songjiang district opened a dedicated space-computing hub at an inaugural forum that drew aerospace developers, semiconductor makers, AI specialists, cloud vendors, researchers and investors under one roof. The hub did not need to attract tenants. Songjiang already had them: SpaceSail, the Shanghai government-backed operator of the Qianfan low-Earth-orbit constellation; two satellite manufacturers, Gesi Aerospace and Xuntian Qianhe; and the Shanghai Aerospace Precision Machinery Research Institute. A municipal government designated a district that was, in every material sense, already a cluster.
Four days later and half a world away, the picture read very differently. SpaceXAI's Starmind AI1, the orbital AI compute satellite built with NVIDIA on an optimised Vera Rubin NVL72 system, is still working out what it actually is. Since NVIDIA's 24 August announcement, the power figures attached to the satellite have moved at least twice, and a widely repeated performance claim, that the system delivers twenty-five times the AI compute of an H100 GPU, belongs to a different NVIDIA product from five months earlier.
One side designated a district that already contained the operator, the manufacturers and the precision-machinery institute. The other announced a partnership whose own arithmetic has not settled. The Chinese programme's technical claims are not thereby shown to be more reliable, and neither bloc's technology is shown to be superior here. What differs is the kind of commitment each side has made public, and how much of it can actually be checked.
What happened this week
The Songjiang hub opening is confirmed against a South China Morning Post report filed the same day. Under a new action plan, the district is targeting more than five billion yuan in space-computing and related sales by 2030, layered onto a precinct that already hosts Shanghai's only designated 6G industry zone. The tenants named above, SpaceSail's Qianfan constellation (more than 200 satellites launched, a stated target of 15,000 by the end of 2030), Gesi Aerospace, Xuntian Qianhe and the precision-machinery institute, predate the hub's designation. The announcement formalised a cluster; it did not seed one.
Reported alongside the hub opening, from a single outlet this week's checking could not corroborate a second time, is a claim that this series' own standing case study, the Three-Body Computing Constellation, is now running eight-billion-parameter AI models aboard its satellites, with laser inter-satellite links holding 99.99 per cent uptime over eight consecutive days. The Three-Body Constellation is not a new entrant; it is the programme this series has tracked since its first twelve-satellite batch launched under Zhejiang Lab in May 2025, and this week's figures describe its current performance rather than a fresh deployment. But the specific numbers rest on one report this week's checking could not confirm a second way, and they should be read at that tier: a reported figure, not a settled one.
Two further international threads landed in the same seventy-two-hour window. Türkiye signed the Artemis Accords on 31 August 2026, becoming the 71st signatory, at a ceremony at NASA headquarters in Washington, signed on Türkiye's behalf by its Minister of Industry and Technology and witnessed by a US Deputy Assistant Secretary of State. The Accords are a normative-layer development, a set of shared principles for civil space activity, not an infrastructure announcement, and they belong in this account for what they are: evidence that the governance conversation around orbital activity keeps widening even as the compute conversation this week is really about two blocs. And on 3 September, at one minute past midnight New Zealand time, Rocket Lab lifted its 94th Electron from Launch Complex 1 at Māhia, carrying a single Japanese synthetic-aperture-radar satellite for the commercial operator Synspective. Three genuinely separate stories, arriving inside three days, none of them derived from the others.
Two ways to build, and only one of them can be checked twice
Space Mafia's Corporate and Sovereign Kardashev framework was built for a week like this one. Corporate Kardashev describes a single company stack extended by partnership: SpaceX's launch and orbital assets, joined to a purpose-built NVIDIA compute payload, announced as one initiative. Sovereign Kardashev describes a state directing the build of orbital infrastructure as policy. Songjiang complicates the second half of that split in a way the framework's original form did not anticipate. This is not one state agency building one programme. It is a municipal government designating a site that happened to already contain an operator, two manufacturers and a research institute, doing the work of vertical integration by proximity rather than by press release.
That distinction matters because it changes what can actually be checked. SpaceSail, Gesi Aerospace, Xuntian Qianhe and the precision-machinery institute are independently verifiable entities that were resident in Songjiang before this week's designation. Starmind AI1's own supporting detail has moved while this series has been watching it. Chinese state industrial policy is not thereby a superior model to a corporate partnership for building orbital-compute capacity, and neither country's underlying technology is shown to be more advanced; the Three-Body Constellation's own headline figures rest on a source this week's checking could not corroborate a second time. What differs, this week, is the shape of the evidence on each side of a fast-moving industry.
The Starmind AI1 power figures illustrate the same point from the American side. Since SpaceX first specified the satellite in July 2026, the number attached to its power draw has moved at least twice: an initial 120 kilowatts average, 150 kilowatts peak; then, five days later, a revised 160 kilowatts average, 250 kilowatts peak, explicitly described by SpaceX as an increase; and a third figure, 175 kilowatts average against the same 250 kilowatt peak, that has circulated since early August without either earlier pair being formally retracted. None of the three should be treated as the current, settled specification. A satellite that has not yet flown, with a headline number that has moved twice in six weeks, is not a discrepancy at the margins. It is a live, unresolved question about what SpaceX and NVIDIA are actually building.
The clearest instance of the same discipline gap is the "twenty-five times the AI compute of an H100 GPU" figure now attached to Starmind AI1 in some secondary coverage. NVIDIA's chief executive did make that claim, at the company's GTC conference on 16 March 2026, for a genuinely different product: the Space-1 Vera Rubin Module, a size, weight and power-constrained payload built for six named partners, Aetherflux, Axiom Space, Kepler Communications, Planet, Sophia Space and Starcloud. Starmind AI1 and the NVL72 rack are not mentioned in that announcement. SpaceX appears once, in an unrelated aside. Five months and one entirely different product later, at least one piece of financial-press coverage attaches the same twenty-five-times figure to Starmind AI1 directly, with no qualification. It is a real number, correctly reported once, now travelling under someone else's name.
That is the pattern this series keeps returning to: a manufacturer's claim, made once, precisely, for a named product, drifting loose of its own context as coverage compounds it. It is also the discipline this article applies to itself. The Three-Body Constellation's eight-billion-parameter figure sits at the same single-source tier the twenty-five-times claim occupied before this week's checking traced it back to its actual source. Neither gets to be a headline until it can be checked a second way.
Read together, Songjiang and Starmind are a clean instance of the Two Space Clouds thesis this series has tracked since March: separate orbital-compute environments forming along geopolitical lines, each with its own supply chain, its own vertical-integration logic and, this week shows, its own evidentiary standard. Sovereign Kardashev by proximity, against Corporate Kardashev by announcement. Neither reading says one cloud is safer, more capable or more legitimate than the other. It says the two are not yet comparable on the same evidence.
None of this is settled by counting which side has the bigger number. It is settled, if it can be settled at all this early, by asking which claims trace to an entity that already existed and could be checked before the announcement, and which claims trace only to the announcement itself. Songjiang's tenants clear that bar. Starmind AI1's power specification, as things stand, does not.
The New Zealand position: launch site, not tasking authority
New Zealand's part in this week's picture is smaller, and more honest about it. Rocket Lab's 94th Electron, named "Owl Around The World," lifted from Māhia at one minute past midnight New Zealand time on 3 September 2026, carrying a single Synspective StriX synthetic-aperture-radar satellite to a 575 kilometre low Earth orbit. It was the fifteenth Electron launch of the year and the ninety-fourth of the programme's lifetime. Rocket Lab is Synspective's sole launch provider under a standing multi-launch agreement.
This is not a response to Songjiang or to Starmind, and it should not be read as one. The payload is a Japanese Earth-observation satellite for a Japanese commercial operator. New Zealand's role this week is as a launch-service participant in someone else's constellation, on someone else's licence, aligned with neither the cluster in Shanghai nor the partnership in the United States. That is the honest and slightly uncomfortable part of the picture: participating in the orbital supply chain as a launch site is real, and it is not the same capability as owning tasking authority over what gets observed, or owning the compute layer that turns a synthetic-aperture return into a usable image.
New Zealand's own ground-based space infrastructure regime, the authorisation framework created by the Outer Space and High-altitude Activities Amendment Act 2025, closed its transitional period on 29 July 2026 and continues to operate as standing background to every launch from New Zealand soil. Nothing in this week's findings bears directly on that regime's requirements, and nothing here evaluates whether it is adequate to what it covers. What it has not yet been tested against, because no source located this week says either way, is whether it addresses orbital compute or downlink activity originating from a state-directed foreign constellation such as Qianfan or the Three-Body Constellation. That is a gap, not a finding, and it should be treated as one: a question this week's evidence leaves open, not a conclusion this week's evidence supports.
For a New Zealand board or technical leadership team, the practical takeaway sits one layer below either headline. It is not which bloc is ahead. Three questions do most of the work when an orbital-compute or AI-infrastructure claim reaches a desk: where the number first appeared, whether it has changed since, and whether it has been attached to more than one product. Ask them of a vendor briefing, an industry report or a conference keynote alike. This week, the Starmind power figures failed the second question and the twenty-five-times claim failed the third. Both were caught by reading the original announcement, not by scepticism about orbital compute as a category.
What to watch next
The book's Heaven Vector and Skynet Vector framing applies cleanly to this week's pairing, and it is worth holding both readings rather than collapsing them into one. The Heaven Vector: a government-designated industrial cluster is at least transparent about what already exists in one place today, a form of legibility an individual corporate announcement does not automatically provide. The Skynet Vector: state-directed clustering of this kind concentrates orbital-compute capability inside institutions under no obligation to disclose beyond what a forum announcement states, and a corporate partnership's own unreconciled, twice-revised figures are their own form of opacity, a commercial one rather than a state one. Both readings sit on the same set of facts. The book's dual-pathway structure says to present them side by side rather than resolve them, and this week is a genuine instance of why: neither side of this comparison has yet supplied enough independently checkable detail to settle which vector it is actually on.
What to watch next: whether a second outlet corroborates the Three-Body Constellation's eight-billion-parameter and laser-link figures; whether SpaceX or NVIDIA issues a fourth, reconciling power figure for Starmind AI1, or lets the third figure stand uncorrected; whether Songjiang's more than five-billion-yuan 2030 sales target produces any verifiable interim milestone before then; and whether the financial-press coverage that already conflates the twenty-five-times claim with Starmind AI1 gets corrected or simply gets repeated by the next outlet that picks it up. None of those questions will be settled by the next headline. They will be settled, if at all, by whoever checks the primary source next.
Starmind AI1 is a closed stack by design: SpaceX's launch and orbital assets, NVIDIA's compute payload, announced and specified by the two companies alone. The orbital ground segment has a different tradition running alongside it. GNU Radio, the open-source software-defined radio framework widely used in ground-station signal processing, and F Prime, the Jet Propulsion Laboratory's own openly licensed flight software framework, flown on multiple small-satellite missions since its 2017 release, both show that vertical integration is a choice about who controls the stack, not a technical requirement of building one. Neither project resolves this week's power-figure dispute or the Songjiang cluster's evidentiary tier. What they demonstrate is that an openly inspectable ground segment and flight software layer is available to any operator, sovereign or corporate, willing to build on it instead of around it.
The same question of who controls the stack applies to New Zealand's own position, one layer up from ground-segment software. Rocket Lab's Māhia launch this week carried a foreign-owned Earth-observation payload: real launch-site sovereignty, but not tasking sovereignty, and not the AI-enabled image analysis that turns a synthetic-aperture return into something a government or a business can act on. That upstream dependency, on someone else's tasking authority and someone else's compute, is the same earth-observation tasking sovereignty question this series raised when a commercial orbital compute cluster first went operational. Neither this week's Chinese industrial cluster nor the American corporate partnership changes New Zealand's answer to it. Both illustrate, from opposite directions, that owning the antenna and owning the analysis are separate capabilities, and a country can hold the first without the second.
Two claims. Two very different evidentiary states. A government that designated a district it did not need to build, and a partnership whose own numbers have not agreed with themselves for five weeks. None of that tells you which country will win the race to put usable AI compute in orbit. It tells you something more immediately useful: that the discipline of checking a claim against its primary source, not against how many times it has been repeated, is the same discipline whether the claim comes from a Shanghai forum or an NVIDIA press release.
Where has your organisation taken an orbital-compute or AI-infrastructure claim at face value, only to find the number had moved, or belonged to a different product, once you went looking for the original source? I would like to hear about it.
This is one of seven weekly series in The Hamberger Report. Subscribe on LinkedIn and the next one arrives in your feed.
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. I write as director of Te Pono Limited; the views are personal and do not represent the position of any client, 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, and it is politically neutral.
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.
This article was produced with AI assistance under my direction. Research, drafting and images pass through a pipeline I built and govern: automated gates for source verification, forbidden language and political neutrality, and my own review before anything is published. The tools include Claude, Gemini and Openart. The frameworks, arguments and editorial judgements are mine and are the same discipline I apply to the AI systems I audit for clients. AI accelerated the work; the thinking, and the responsibility for it, are mine.
[1] South China Morning Post. "China-US space race heats up as Shanghai unveils space-computing hub." 1 September 2026. https://www.scmp.com/tech/tech-war/article/3365839/china-us-space-race-heats-shanghai-unveils-space-computing-hub
[2] TechTimes. "China Has 8 Billion-Parameter AI Running In Orbit As Shanghai Opens Space Computing Hub." 1 September 2026. https://www.techtimes.com/articles/326180/20260901/china-has-8-billion-parameter-ai-running-orbit-shanghai-opens-space-computing-hub.htm
[3] NASA. "NASA Welcomes Türkiye as Newest Artemis Accords Signatory." 31 August 2026. https://www.nasa.gov/organizations/oiir/artemis-accords/nasa-welcomes-turkiye-as-newest-artemis-accords-signatory/
[4] Rocket Lab, via GlobeNewswire. "Mission Success: Rocket Lab Launches 94th Electron Mission." 2 September 2026. https://www.globenewswire.com/news-release/2026/09/02/3355246/0/en/mission-success-rocket-lab-launches-94th-electron-mission.html
[5] NVIDIA Newsroom. "SpaceXAI Adopts NVIDIA Vera CPU to Accelerate Agentic AI at Massive Scale." 24 August 2026. https://nvidianews.nvidia.com/news/spacexai-adopts-nvidia-vera-cpu-to-accelerate-agentic-ai-at-massive-scale
[6] NotebookCheck. "SpaceX lists AI1 satellite cooling specs for Starmind data center in space." 11 July 2026. https://www.notebookcheck.net/SpaceX-lists-AI1-satellite-cooling-specs-for-Starmind-data-center-in-space.1340218.0.html
[7] NextBigFuture. "SpaceX AI Satellites Will Have 250 Kilowatts of Power." 16 July 2026. https://www.nextbigfuture.com/2026/07/spacex-ai-satellites-will-have-250-kilowatts-of-power.html
[8] The Register. "Nvidia chips in spaaaaaace." 17 March 2026. https://www.theregister.com/2026/03/17/nvidia_chips_in_spaaaaaace/