Amazon has built one thousand satellites. Three hundred and ninety-six of them are in orbit.
The company disclosed the thousandth build on 8 October 2026. Its plant in Kirkland, Washington took six months to produce the first twenty-seven Amazon Leo satellites. It now produces twenty-seven a week. That rate makes Amazon the second-largest satellite manufacturer in the world, behind SpaceX, on its own account of its own output [1][2].
The useful number this week is the gap. One thousand built against three hundred and ninety-six flying, with four more deorbited or failed, leaves roughly six hundred finished spacecraft on the ground. The first-generation constellation needs 3,232 to be complete [1].
Six hundred satellites are sitting in a warehouse. They are waiting for a ride.
What is actually holding them
Amazon contracted its Amazon Leo deployment across several launch-vehicle families. Two of them are currently out of service.
United Launch Alliance's Vulcan Centaur was grounded by an issue involving its solid rocket motors. Blue Origin's New Glenn has been out of service since 28 May 2026. A vehicle exploded during a pre-launch static-fire test at Launch Complex 36, Cape Canaveral Space Force Station. On 5 August 2026 Blue Origin identified the root cause as the main oxygen valve on one of the rocket's seven BE-4 engines [3][4]. The flight that explosion delayed was to carry forty-eight Amazon-owned satellites.
Trade coverage of the 8 October milestone states plainly that both vehicles are out of service and that hundreds of satellites are on standby in Florida waiting for a ride [2]. An Amazon executive is quoted in the same coverage to the effect that the company never wants a launch slot to open with no satellites ready to fill it. On the evidence of the build rate, that part is working.
Two groundings, one of them traced to a specific valve on a specific engine, is a documented engineering sequence. It is not a verdict on either company. Vehicles under development fail tests, and the public record here is a root-cause finding published by the company that made the engine. What the record does establish is the operational consequence: the satellites exist. The rides do not.
Three clocks, one company
Space Mafia carries a three-clock device: three clocks inside one operation, running at different speeds, with only one of them binding. Which three depends on the case. This week they are manufacturing, launch, and regulatory.
The manufacturing clock is the fast one. Twenty-seven a week, from a standing start of six months for the first batch, is a production ramp that has done what it was asked to do.
The launch clock has stopped for two of three vehicle families. That clock belongs to the engineering and qualification programmes of two other companies. The only vehicle currently moving the manifest belongs to a third.
The regulatory clock does not move to match the other two. It runs on a fixed calendar. Only it carries legal consequences.
The clock that binds
Amazon Leo was required to have half its planned 3,232-satellite constellation, 1,616 satellites, in orbit and operating by 30 July 2026. It missed by a wide margin. The Federal Communications Commission, the FCC, did not revoke the authorisation. It issued a conditional waiver instead: satellites launched after 30 July 2026 do not carry the spectrum priority granted in the Commission's 2020 and 2021 processing rounds [5].
That priority is restorable on three paths. An accelerated path, around October 2027, opens on documentary proof that the first-generation hardware is manufactured and the launch contracts to deploy it are secured. An automatic path opens the moment the constellation actually reaches 1,616 satellites in orbit. Failing both, restoration arrives by 30 March 2028 regardless. The terminal full-constellation deadline, 30 July 2029, is untouched [5].
Now read this week's finding against that structure.
The accelerated path has two limbs: hardware built, launches contracted. This week's disclosure settles the first limb comfortably. A thousand satellites exist, at a demonstrated rate of twenty-seven a week, against a requirement of 3,232. Whatever Amazon's problem is, it is not the factory.
The second limb is under pressure. The automatic path does not care about either limb. It counts satellites in orbit, not satellites built, launches contracted, money spent, or diligence demonstrated in the layer the operator actually runs.
That is the whole mechanism in one sentence, and it is worth sitting with before reaching for an analogy. An operator cannot argue its way back up the priority list by pointing at the warehouse, because the warehouse is not the thing being measured.
Outcome against process, and what it costs
Here is the comparison I keep returning to. It is mine, and it is not something the Commission has said about its own instrument.
Almost every frontier-AI control proposed this year conditions on process: document the evaluations, file the safety case, maintain the register, follow the procedure. The procedure is the thing that gets audited. The FCC's spectrum-priority mechanism conditions on an outcome instead. It does not care which layer of the supply chain produced the shortfall.
Those are opposite design choices. They fail in opposite directions. A process-conditioned control can be satisfied in full by an operator who produces nothing at all, because the artefacts the control inspects are the documents, not the result. An outcome-conditioned control counts the result and cannot be satisfied by paperwork. It also counts against the operator a failure in a layer the operator does not run. Amazon did not design the oxygen valve that stopped New Glenn, and it does not qualify Vulcan's solid rocket motors. The mechanism does not distinguish. The second half is the price of the first.
The general form of that trade-off belongs to the Thursday series, where control design is the standing subject. This week is one worked instance of it, in a domain where the counting is unusually honest.
Flown against filed
The second half of this week's material is also a count, and it points the other way.
Two dedicated artificial-intelligence compute satellites are operating in orbit worldwide, carrying five data-centre-class accelerators between them.
Google's Project Suncatcher prototype launched on 1 October 2026 on a SpaceX rideshare mission from Vandenberg Space Force Base, with Planet supplying the satellite bus. It carries four Trillium-generation tensor processing units drawing on roughly one kilowatt of onboard solar power, with aggregate capability Google compares to a single data-centre server. The chips run in bursts of about fifteen minutes before the spacecraft has to cool. The mission is designed for about a year. Google's own post confirms the satellite is operating as expected and describes the work as a "long-term research moonshot" into whether space could one day host machine-learning infrastructure at scale [6][7].
Starcloud's satellite has been in orbit since November 2025, carrying one Nvidia H100 graphics processing unit in a package about the size of a small refrigerator. It ran inference on an open-weight Gemma model and trained a small language model on orbit [8].
Five accelerators, two spacecraft.
Against that, three separate applications before the FCC propose a combined 1,139,600 satellites for orbital compute. SpaceX's Orbital Data Center application proposes up to one million satellites, filed 30 January 2026 and accepted for filing in February. Blue Origin's Project Sunrise proposes 51,600, and Starcloud's constellation proposes 88,000 in sun-synchronous orbit [9][10]. None has been granted. A fourth, separate application from Orbital Compute Inc., for up to 100,000 satellites, was filed on 24 June 2026 and sits outside that total [11]. Fold it in and the four-application figure is 1,239,600, which is a different number describing a different set of filings.
This series has counted those filings twice before. What has changed since is the flown number, which went from one satellite to two.
The distance between five accelerators and 1,139,600 proposed satellites is a capability gap. The distance between a thousand satellites built and three hundred and ninety-six flying is a logistics gap. They are not the same shape. They do not resolve on the same timescale.
The physics constrains the first gap harder than launch cost does. Andrew Cavalier of ABI Research, speaking to IEEE Spectrum in June 2026, put radiators and solar arrays at 65 to 70 per cent of total satellite mass [12]. Space-grade photovoltaics run orders of magnitude more expensive than their terrestrial equivalents. A satellite must eventually radiate waste heat equivalent to the power it generates. The area that collects the power is close to the area needed to reject the heat. Cheaper launch does not change that ratio.
Scale brings a second constraint. The Secure World Foundation published an insight piece by Ian Christensen and Dr Alex Shi on 21 May 2026. It put the post-mission disposal success rate at ninety-nine per cent, and warned that the associated collision and casualty risks "compound quickly over tens of thousands of satellites and may become non-trivial" [13]. Apply that stated rate to a hypothetical million-satellite constellation and you get ten thousand disposal failures. That arithmetic is mine, not theirs; the Foundation published the rate, not the product. The habit of checking a widely-repeated derived figure against what its source actually said is the discipline the Saturday series exists to teach.
What New Zealand is selling into this
On 30 September 2026 Rocket Lab announced its largest commercial Electron contract to date: twenty launches for Synspective, a Japanese synthetic aperture radar operator and one of Electron's earliest customers. The launches are scheduled annually from 2028 through to 2031, all from Launch Complex 1 at Māhia. The deal takes Synspective to forty-seven contracted Electron missions, more than any other Rocket Lab customer, and pushes the company's overall launch backlog beyond one hundred missions [14]. Rocket Lab had already deployed thirteen StriX satellites for Synspective since the relationship began in 2020. Its founder, Sir Peter Beck, described the arrangement in terms of "tailored, dedicated launch" as the fastest and most reliable route to scaling a small-satellite constellation.
The contract's dollar value is not established. Two different derived estimates are circulating, both back-calculated from different assumed per-launch prices, and neither is a figure Rocket Lab's own release is confirmed to state. I am not going to repeat either one.
What the deal sells is the exact commodity Amazon's six hundred grounded satellites demonstrate the absence of. It is a launch on a date the customer picked, from a vehicle sized to the payload, on a cadence the launch operator controls from its own pad. It is the structural difference between a dedicated small launcher and a share of an oversubscribed heavy-lift queue. One of those is a schedule. The other is a position in someone else's schedule.
Rocket Lab's larger Neutron vehicle is the other half of the company's position. Company guidance as at the August 2026 reporting is a first-stage tank delivery to the Wallops, Virginia pad in the fourth quarter of 2026. The nine-engine Archimedes set needed for the maiden flight has entered production, with more than four hundred hot fires completed across both engine variants. Pad delivery is followed by static-fire and further qualification testing before anything launches. Several observers have assessed the window for a 2026 maiden flight as narrowing, and some expect the launch itself to slip [15]. That assessment is analyst and observer forecast on the contingency of qualification testing.
For New Zealand organisations, the regulatory frame around all of this is the Outer Space and High-altitude Activities Amendment Act 2025, the ground-based space infrastructure regime, which reaches ground infrastructure, launch, and payload. Its transitional automatic-authorisation period closed on 29 July 2026, so operators are inside the enforced regime rather than the lead-in to it [16].
The practitioner question is narrower than the regulatory one and lands sooner. Most organisations now hold at least one dependency whose timeline is set by a queue they do not control: a satellite imagery feed, a connectivity path, a chip allocation, a certification slot. The discipline is to name the binding clock rather than the busiest one. Amazon's busiest clock was manufacturing, and it ran beautifully. The clock that binds was somewhere else entirely, in another company's engineering programme, and no amount of additional effort on the fast clock moves the slow one.
What to watch
Four dates carry real information. Around October 2027, Amazon can attempt the accelerated restoration path on documentary proof of hardware and contracts. Restoration happens automatically whenever the constellation reaches 1,616 satellites in orbit, which is the figure to watch rather than any announcement. 30 March 2028 is the outer limit either way. 30 July 2029 remains the terminal full-constellation deadline [5].
Two further things are worth checking rather than assuming. Whether any of the four orbital-compute applications is granted, and on what deployment conditions, will say more about the regulator's posture than any of the filings do. And Starcloud has publicly planned a second satellite carrying newer hardware; whether it has flown is a question to verify against the record rather than a milestone to expect.
One more fact belongs here. The same Commission that applied a reversible, outcome-conditioned demotion to Amazon Leo authorised SpaceX on 7 October 2026. SpaceX can deploy fifteen thousand satellites in very low Earth orbit for direct-to-device service, on a two-stage deadline running to 2035 [17]. Those are two documented actions with two documented mechanisms in one season. Reading a motive into the pairing would be guesswork. The mechanisms themselves are on the public record and they are different instruments.
The open-source dimension
The open-source dimension sits exactly where the counting happens. A test denominated in satellites actually in orbit is only as good as the ability of somebody outside the operator to check the count. Yamcs is an open-source mission control system, used on real missions, whose telemetry and command path an engineer outside the programme can read rather than take on trust. The manufacturing half has an open-standards story of its own. The openly published CubeSat Design Specification, written at California Polytechnic State University, gave builders who had never coordinated with one another a single shared mechanical and electrical interface. That is how small-satellite production at rate became possible for anyone other than a prime contractor. Open tooling does not make a constellation trustworthy. It makes a claim about a constellation checkable, which is the more useful property.
Sovereign space capacity is usually discussed as a question of what a country owns. Cadence is the sharper measure. A satellite that exists but cannot fly is not capacity, and an imaging capability whose renewal sits behind a queue somebody else controls is capacity on loan. Synthetic aperture radar, which images through cloud and at night, is the sensor class that maritime domain awareness across New Zealand's exclusive economic zone depends on. Radar satellites have working lives measured in a handful of years. Replacement is a standing requirement rather than a one-off purchase. A country that hosts the launch end of that chain, as New Zealand does at Māhia, holds a position in the cadence. A country buying rides on a shared heavy-lift manifest does not.
Which dependency in your own organisation is paced by a queue you do not control, and what happened the last time that queue stopped?
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 morethan 30 years of professional experience in architecture, security, andtechnology leadership in New Zealand. I write as director of Te PonoLimited; the views are personal and do not represent the position of anyclient, any government agency, or the New Zealand government. My commentaryon legislation and policy is analytical, drawing on publicly availablesources and my professional expertise in architecture, security, and AIgovernance, and it is politically neutral.
About the Author: 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: automatedgates 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 aremine and are the same discipline I apply to the AI systems I audit forclients. AI accelerated the work; the thinking, and the responsibility forit, are mine.
[1] TheNextWeb. "Amazon Leo builds its 1,000th satellite." 8 October 2026. https://thenextweb.com/news/amazon-leo-1000th-satellite-europe
[2] Via Satellite. "Amazon Hits 1,000-Sat Milestone, Blue Origin Bolsters Sat Manufacturing." 8 October 2026. https://www.satellitetoday.com/manufacturing/2026/10/08/amazon-hits-1000-sat-milestone-blue-origin-bolsters-sat-manufacturing/
[3] Spaceflight Now. "Blue Origin's New Glenn rocket explodes during prelaunch testing at Cape Canaveral." 29 May 2026. https://spaceflightnow.com/2026/05/29/blue-origins-new-glenn-rocket-explodes-during-prelaunch-testing-at-cape-canaveral/
[4] US News. "Blue Origin Zeroes In on Rocket Engine Issue as Source of New Glenn Explosion." 5 August 2026. https://www.usnews.com/news/top-news/articles/2026-08-05/blue-origin-zeroes-in-on-rocket-engine-issue-as-source-of-new-glenn-explosion
[5] SDxCentral. "FCC waives Amazon Leo deadline after extending it as constellation gets ready for market." 2026. https://www.sdxcentral.com/news/fcc-waives-amazon-leo-deadline-after-extending-it-as-constellation-gets-ready-for-market/
[6] Google. "Project Suncatcher prototype." 2026. https://blog.google/innovation-and-ai/models-and-research/google-research/project-suncatcher-prototype/
[7] GCN. "Google Project Suncatcher puts four chips in orbit." 4 October 2026. https://gcn.com/google-project-suncatcher-puts-four-chips/22176
[8] CNBC. "Nvidia-backed Starcloud trains first AI model in space, eyes orbital data centers." 10 December 2025. https://www.cnbc.com/2025/12/10/nvidia-backed-starcloud-trains-first-ai-model-in-space-orbital-data-centers.html
[9] Federal Communications Commission. "DA-26-113A1" (SpaceX Orbital Data Center application accepted for filing). 2026. https://docs.fcc.gov/public/attachments/DA-26-113A1.txt
[10] SpaceNews. "Starcloud files plans for 88,000-satellite constellation." 2026. https://spacenews.com/starcloud-files-plans-for-88000-satellite-constellation/
[11] SpaceNews. "Orbital files plans for 100,000 orbital data centers." 2026. https://spacenews.com/orbital-files-plans-for-100000-orbital-data-centers/
[12] IEEE Spectrum. "Orbital data centers face a heat problem." 11 June 2026. https://spectrum.ieee.org/orbital-data-centers-heat
[13] Secure World Foundation. "To ODC or not to ODC: Insight for May 2026" (Ian Christensen and Dr Alex Shi). 21 May 2026. https://www.swfound.org/publications-and-reports/to-odc-or-not-to-odc-insight-2026
[14] NZ Herald. "Rocket Lab inks largest ever Electron launch contract." 30 September 2026. https://www.nzherald.co.nz/business/rocket-lab-inks-largest-ever-electron-launch-contract/premium/B6CDTUUNI5GCVL5HA6JVKZRBGI/
[15] Spaceflight Now. "Window for 2026 launch debut of Rocket Lab's Neutron rocket is narrowing as development continues." 10 August 2026. https://spaceflightnow.com/2026/08/10/window-for-2026-launch-debut-of-rocket-labs-neutron-rocket-is-narrowing-as-development-continues/
[16] Ministry of Business, Innovation and Employment (New Zealand). "Our regulatory regime." https://www.mbie.govt.nz/science-and-technology/space/our-regulatory-regime
[17] SatNews. "FCC Grants SpaceX Approval for 15,000-Satellite Very Low Earth Orbit Direct-to-Device Constellation." 7 October 2026. https://satnews.com/2026/10/07/fcc-grants-spacex-approval-for-15000-satellite-very-low-earth-orbit-direct-to-device-constellation/