Technology
Starlink pays the bills, Starmind is the bet
The short version
In about August 2023 I installed Starlink for a farmer, to get good internet into all the farm buildings. Where a cable run made sense I used 30-metre lengths of network cable, and where it didn't I used Ubiquiti NanoBeam wireless bridges, which cost about £83 each and can link points several kilometres apart. We even talked about sharing the connection with the neighbours over the same bridges. Nothing came of it, and I was never sure it would have been legal. So when I write about Starlink, I am writing as someone who has fitted one and seen what it does for a farm, not only as someone reading the headlines.
On 12 November 2025, after Elon Musk said SpaceX would build satellites that were also data centres, I started chewing over what that meant. My own explanation was that Starlink has to keep replacing its satellites anyway, so the replacements might as well be cleverer. That was eight months before SpaceX's name for the project, Starmind, surfaced in July 2026.
I should admit that I cannot leave this subject alone. As I put it in April 2026: "Sometimes I look at what Elon Musk is doing with his satellites and I bore my friends to death… and my wife, she's very stoic and listens to me… but I can tell she wishes I would shut up." This article is my attempt to set it all down in one place, with the numbers checked.
Here is the short answer. SpaceX already runs one very good business in space. Starlink brought in about $11.4 billion in 2025 and was the only part of the company that made an operating profit. SpaceX now wants to add a second business: Starmind, a planned fleet of satellites that would work as data centres for artificial intelligence (AI), powered by sunlight and cooled by radiating heat into space. Starlink was built on Falcon 9. That rocket, whose first stage lands and flies again, has launched nearly every Starlink satellite in orbit today, and cheap, frequent Falcon 9 flights are the reason Starlink exists at all. The next step for both businesses needs something bigger. The larger V3 Starlink satellites and anything on the scale of Starmind are designed for Starship, the giant fully reusable rocket that is about to attempt its first flight to orbit.
The big numbers online, such as hundreds of billions of dollars a year or gigawatts in orbit by 2028, are mostly analysts' scenarios, not SpaceX forecasts. What is real today is a profitable Starlink, a published satellite design for Starmind, a regulatory filing, and a handful of small experiments by other companies.
What Starmind is
In January 2026 SpaceX asked the US Federal Communications Commission (FCC), which licenses satellites, for permission to launch up to one million satellites acting as orbital data centres. The filing describes solar-powered satellites between 500 km and 2,000 km up, some in sun-synchronous orbit (SSO), a path that keeps a satellite over the line between day and night so its solar panels stay in sunlight. The satellites would talk to each other, and to Starlink, by laser.
In July 2026 Aviation Week reported that SpaceX would run this constellation under the name Starmind. It was not the only name in circulation. In February 2026 I noted that the project would draw critics and that "someone has termed starthink", and Brian Wang of the Next Big Future blog was writing about "Starthink AI V1 satellites" the following month. Starmind is the one SpaceX chose.
In August SpaceX published details of the first satellite, called AI1. According to heise, it has solar arrays spanning about 70 metres with 150 kilowatts (kW) of peak power, and it runs liquid-cooled Nvidia Vera Rubin servers drawing about 120 kW on average. The waste heat is pumped to a radiator of about 110 square metres that sheds it into space. heise calls the servers water-cooled, but I doubt it is plain water, because water freezes easily and ice bursts pipes. SpaceX has not said what the coolant is. On Earth these servers are normally cooled with water mixed with glycol antifreeze, and the International Space Station keeps water inside the station but uses ammonia in the loops that run out to its radiators, because ammonia does not freeze at the temperatures they reach. Notebookcheck reports that the first AI1 prototypes are planned for early 2027, with volume production targeted by the end of that year.
My own view of where this came from has not changed since November 2025, when I wrote: "It's not about putting up data centres, Starlink have to replace all of their satellites on a running system as they devolve from orbit and burn up… so… you always want to be replacing them with something more sophisticated… and what's more useful than a space-based data center, along with its existing Starlink function." In January 2026 I put it more bluntly: "satellite data centres are the next evolution of an already existing system, they are nothing more than beefed up Starlink satellites… I don't mean that in a derogatory tone, I mean… the technology is already there begging for this upgrade." SpaceX has in fact filed Starmind as a separate constellation rather than adding computers to Starlink satellites, so the dual-purpose part of my guess has not happened. The family resemblance is plain, though: solar panels, lasers, mass production and the same rocket.
A note on what we now know
I am leaving those earlier words as I wrote them. More has become clear about Starmind since then, and it changes the picture I had in mind. The Starmind satellites are much bigger than Starlink satellites. They do not need to go into the Starlink shells. They need their own special shell, which orients them so they can collect solar energy 24 hours a day.
Why anyone would put a data centre in space
The argument starts with the problems of building on Earth. Speaking at the All-In Summit in September 2026, SpaceX president Gwynne Shotwell said that as soon as people learn a large computing centre is coming, local land can jump from $3,000 an acre to $180,000 an acre, and that permits, grid connections and generators can take years. In orbit, she argued, the real estate is free, cooling comes from a radiator facing deep space, and a satellite in the right orbit can face the sun almost all the time.
In December 2025 I summed up the logic this way: "Elon has identified energy as the new currency. He's planning to build a data centre in space because it can convert the solar energy into AI thought." The same month I gave a name to a pattern I keep seeing in Musk's businesses: "Elon Musk often has a deeply buried opportunity, not obvious to most… I call it a Musk 'discovered check'." In chess, a discovered check happens when you move one piece and reveal an attack by another piece that was sitting behind it. Starlink looked like an internet business. Behind it sat the thing that makes Starmind possible: a factory for mass-produced, solar-powered, laser-linked satellites and a rocket to carry them.
That is what orbit offers. The costs are that everything has to be launched, electronics must survive radiation, nobody can swap a failed server, and results must be sent back to Earth, which is why Starmind plugs into Starlink's laser network.
One rocket, two cargoes
The most important point for understanding any of the numbers is simple. Each Starship flight carries one kind of cargo. It is either a stack of Starlink internet satellites, whose value is measured in data capacity (terabits per second, or Tbps), or a load of Starmind computing satellites, whose value is measured in electrical power (megawatts, MW, or gigawatts, GW, where one gigawatt is 1,000 megawatts). The two also go to different orbits: Starlink's first V3 satellites are heading for a low orbit inclined at about 30 degrees, while computing satellites want sun-synchronous orbit for constant power.
That means the scenarios below are alternatives, not additions. A year of Starship flights can be imagined as all Starlink, or all Starmind, or some mix, but the same flight cannot count twice. SpaceX has not said how it will divide its flights between the two.
Early test satellites are the exception: they can fly low, where a failure is dragged down by the thin upper atmosphere and burns up quickly instead of lingering as debris.
Starlink today: the money
Starlink is the part of SpaceX that pays its way. SpaceX's stock market prospectus, filed in May 2026, shows its Connectivity segment, which is mainly Starlink, earning revenue of $11.4 billion and an operating profit of $4.4 billion in 2025, with revenue up about 50% on the year. That was roughly 61% of SpaceX's total revenue of $18.7 billion. By comparison, SpaceX's AI segment (the former xAI, which makes the Grok chatbot) took in $3.2 billion in 2025 while spending $12.7 billion on capital projects, mostly its Colossus data centres on the ground. In other words, AI at SpaceX today means computers on Earth. Starmind has not earned anything yet.
Hargreaves Lansdown's summary of the prospectus shows subscribers rising from 2.3 million in 2023 to 10.3 million in early 2026, while average revenue per user fell from $99 a month to $66. More customers are paying less each.
The next step is the V3 satellite. SpaceX says each V3 will provide 1 Tbps of download capacity and 160 gigabits per second of upload, about ten times the download of today's satellites, and that a full Starship load is planned to add about 60 Tbps, more than 20 times what a Falcon 9 launch adds today. The first real load will be smaller. Starship Flight 14, now set for no earlier than 28 September 2026, is due to carry 26 V3 satellites, adding 26 Tbps.
This is where the famous "$1 billion per launch" comes from. Analysts at ARK Invest, the investment firm led by Cathie Wood, estimate that Starlink earned about $19 million a year for each Tbps of capacity in 2025. Multiply that by about 61 Tbps for a full Starship load and you get roughly $1.16 billion a year, which ARK rounds to $1 billion. It is an estimate of the extra yearly revenue a full load could support if the new capacity sold at 2025 prices, not a launch price or booked money, and ARK itself expects revenue per Tbps to fall.
To see why that caveat matters, imagine a year of 200 Starship flights, all carrying full Starlink loads. That would add about 12,000 Tbps, and at ARK's 2025 rate it would imply around $200 billion a year of extra revenue. ARK's Brett Winton told CNBC in June that the whole Starlink network currently offers about 500 Tbps. Adding more than twenty times today's capacity in one year could not be sold at today's prices, so the $200 billion is a ceiling on the arithmetic rather than a projection. ARK's own published view, in its June 2026 newsletter, is more measured: 200 to 300 Starship launches a year could generate hundreds of billions of dollars of gross profit by 2030.
How big could orbital compute get?
Now imagine the same rocket carrying computing satellites instead. Brian Wang has published the most detailed public sketches. In his earlier sketch, one Starship lifts about 120 tonnes of satellites at about 70 kW per tonne, which is about 8.4 MW of computing power per flight, or roughly seventy AI1-class satellites. His later sketch assumes a bigger future Starship lifting 200 tonnes of denser satellites at 100 kW per tonne, which gives 20 MW per flight. That 200-tonne figure is his assumption about a future vehicle. SpaceX's published figure for a reusable Starship is 100 to 150 tonnes.
The table shows how many flights it would take to launch as much power as a British nuclear station or America's best-known dam.
| To match | Power | Flights at 8.4 MW each (earlier sketch) | Flights at 20 MW each (later sketch) |
|---|---|---|---|
| Sizewell B nuclear power station, UK | 1,198 MW | 143 | 60 |
| One gigawatt of computing | 1,000 MW | 119 | 50 |
| Hoover Dam, USA | 2,074 MW | 247 | 104 |
Sources: Sizewell B net output from EDF; Hoover Dam nameplate capacity from the US Bureau of Reclamation; power per flight from Brian Wang's sketches.
Wang's timeline has roughly 200 Starship flights in 2028. If every one of them carried computing satellites at 20 MW, that would add about 4 GW in orbit in that year. This is the other side of the Starlink scenario above, not an addition to it: the same 200 flights cannot also add $200 billion of Starlink capacity. Even on his optimistic assumptions, Earth still dominates. The same sketch has computing on the ground growing by 10 to 16 GW in 2028 alone. For 2027, Wang's cost model assumes a fleet-building year of roughly 100 to 140 Starship flights in total, before any split between cargoes.
At the far end, Elon Musk posted in August 2026 that SpaceX is aiming for more than 30 Starship launches a day in 2030, about 10,000 a year. That is a goal, not a costed plan.
What a gigawatt in orbit costs, and why estimates disagree
Published estimates for putting one gigawatt of computing in orbit range from well under $1 billion to about $170 billion. That looks like a wild disagreement, but mostly it is people pricing different things. Some count only the rocket ride. Others add the satellites, the solar panels and radiators, the chips and years of operation.
| Estimate | Cost per gigawatt | What it includes |
|---|---|---|
| Brian Wang, later sketch (200-tonne Starship, 20 MW per flight, about $5 million per flight) | about $0.25 billion | Launch only. Satellites and chips excluded. |
| Brian Wang, early flights (8.4 MW per flight, about $60 million per flight) | about $7.1 billion | Launch only. |
| IEEE Spectrum, based on Andrew McCalip's model | about $51 billion over five years, versus about $16 billion on Earth | Designing, building and launching about 4,300 satellites, plus five years of operation. |
| Futurum Group | about $72 billion, versus about $16 billion for an off-grid site on Earth | Building and launching the satellites, excluding the computer hardware. |
| Wood Mackenzie | about $170 billion, more than three times an equivalent site on Earth | Everything. Launch and satellites are about 60% of the total. |
Wang argues that the high figures, including a widely quoted Bank of America estimate of about $160 billion per gigawatt, assume today's rockets rather than a fully reusable Starship, and his headline claim is that SpaceX's plan comes in at under $1 billion per gigawatt for launch. His $5 million per flight is his own assumption about a mature, fully reused Starship; SpaceX has not published a price. Wood Mackenzie says orbital costs would need to fall by about 70% to match building on the ground, and calls orbital data centres "a bet on the cost curve". Both views can be true at once: launch may become very cheap while a gigawatt in orbit still costs more than one on the ground.
Who else is doing this
SpaceX is not alone, and it is not first. The useful distinction is between what has actually flown and what has only been announced or filed with regulators.
| Who | What has actually flown | What comes next |
|---|---|---|
| Starcloud (US start-up backed by Nvidia) | Starcloud-1, launched 2 November 2025, carried an Nvidia H100 data-centre graphics processor (GPU) and used it to train a small AI model on the works of Shakespeare. | Two 8 kW Starcloud-2 satellites on shared launches in 2027. It has asked the FCC for permission to operate up to 88,000 satellites. |
| Google, Project Suncatcher | Nothing yet. Ground tests found its Trillium tensor processing units (TPUs), Google's AI chips, survived radiation doses far above a five-year dose in orbit without hard failure. | A first prototype built with Planet Labs is due on a SpaceX rideshare flight in early October 2026, then two satellites in 2027 to test laser links between them. |
| China, Three-Body Computing Constellation (ADA Space and Zhejiang Lab) | Twelve satellites launched on 14 May 2025, with about 5 peta-operations per second (POPS) of combined computing and 30 terabytes of storage, linked by lasers. | Plans for up to 2,800 satellites. |
| SpaceX, Starmind | Nothing yet. Starlink V3 satellites, which Starmind would rely on to send results home, fly for the first time on Flight 14. | AI1 prototypes in early 2027 and an FCC application for up to one million satellites. |
China's cluster is small, but it is a working computing constellation. Filings for a million satellites are permission to try, not evidence of a build.
The awkward questions
The first is the one I kept asking myself. In December 2025 I wrote: "I still don't see the use case for these space-based data centers. I don't see any use above, say, 10, 20, or 100." A handful of orbiting data centres is easy to imagine. A million needs a vast, steady supply of paying work, and nobody has yet shown how that computing turns into revenue at that scale.
Radiation is the second. In April 2026 I raised it myself: "There is a weakness, and that's the radiation in space affects the chips… they need metal to surround the chip to protect it from the radiation, which adds considerable weight." My guess at the time was that older chips built on larger, cheaper manufacturing processes might cope better. Google's tests are encouraging, with no hard failures in its chips, but its memory showed irregularities at much lower doses, and a stray particle can flip a bit and corrupt a calculation.
Heat is the third. In a vacuum there is no air to carry heat away, so it can only be radiated, which needs large panels. AI1 needs a radiator of about 110 square metres to shed around 120 kW. Scaling that to gigawatts means a great deal of radiator.
Carbon is the fourth. Starship burns liquid methane and liquid oxygen. Counting only the burning of the fuel, a Stanford physics course estimate puts one launch at about 3,100 tonnes of carbon dioxide. Andrew Wilson of Glasgow Caledonian University, quoted by Space.com, puts a launch at about 76,000 tonnes of carbon dioxide equivalent (CO2e, a measure that folds other greenhouse gases into one number), using a much wider accounting. At a few hundred launches a year the totals are modest by global standards. At Musk's 10,000 launches a year, even the low combustion figure would be about 31 million tonnes, roughly a quarter of the 130.7 million tonnes the US Environmental Protection Agency (EPA) recorded for US commercial aviation in 2022. Soot and water vapour released high in the atmosphere, and metal ash from satellites burning up on re-entry, are further concerns that scientists do not yet fully understand.
Debris and crowded skies are the fifth. My instinct has always been to compare satellites with aircraft. In May 2025 I wrote that "there are 10,000 aeroplanes in the sky at any one time and I… hardly notice them, so I think the same applies to the satellites", though I was not sure of that figure and thought it might be nearer 20,000. But I have also conceded, back in November 2024, that "debris might be a problem particularly if several satellites crashed together." A million-satellite filing makes that concession matter far more.
Then there are the practical limits: broken servers cannot be repaired, and chips are scarce even on the ground.
What to watch next
Several near-term events will show whether this is moving from slides to hardware. Starship Flight 14, no earlier than 28 September 2026, is the first attempt to reach orbit and the first real deployment of V3 Starlink satellites; whether they work and start serving customers within weeks matters for both businesses. Google's first Suncatcher prototype is due to launch in early October. SpaceX's AI1 prototypes are due in early 2027, and Starcloud's two Starcloud-2 satellites in 2027, followed by Google's pair. The FCC's decision on the one-million-satellite application will set the rules. SpaceX's quarterly results as a listed company will show whether any AI revenue starts to come from orbit.
I also keep looking one step further. In February 2026 I set out the pattern as I saw it: "1) Starlink satellites, turning solar power into a money spinner for Musk. 2) Data centre satellites, turning more solar power into an up and coming money spinner!!! 3) Space based refineries, using solar to turn water and carbon dioxide into spaceship fuel. Has anyone seen this next step yet!" Nobody has announced such a thing, and it is speculation on my part, but Starship depends on refuelling in orbit, so fuel made in space is at least the right kind of question.
Until then, the honest summary is the title. Starlink pays the bills. Starmind is the bet.
Sources
- SpaceX Form S-1 prospectus, readable copy: https://fluiddocs.ai/spacex-s1
- ABI Analytics, SpaceX IPO prospectus summary (segment figures): https://www.abianalytics.com/SpaceX_IPO_Summary.html
- Hargreaves Lansdown, Inside SpaceX's IPO filing (subscribers, revenue per user): https://www.hl.co.uk/news/inside-spacexs-ipo-filing-revenue-starlink-ai-and-key-financials
- Aviation Week, SpaceX Brands Data Center Satellites As Starmind, 10 July 2026: https://aviationweek.com/space/commercial-space/spacex-brands-data-center-satellites-starmind
- heise online, SpaceXAI's Starmind AI1 Satellite Carries Nvidia Technology, 24 August 2026: https://www.heise.de/en/news/SpaceXAI-s-Starmind-AI1-Satellite-Carries-Nvidia-Technology-11424154.html
- Notebookcheck, SpaceX names its AI Orbital Data Center satellite constellation: https://www.notebookcheck.net/SpaceX-names-its-AI-Orbital-Data-Center-satellite-constellation.1328148.0.html
- AeroMorning, SpaceX FCC filing for orbital data centres, 2 February 2026: https://aeromorning.com/en/filing-request-to-launch-and-operate-orbital-data-centers/
- Advanced Television, SpaceX COO updates the market (Shotwell remarks), 17 September 2026: https://www.advanced-television.com/2026/09/17/spacex-coo-updates-the-market/
- Starlink Progress Report 2024 (V3 capacity): https://starlink.com/public-files/starlinkProgressReport_2024.pdf
- SpaceX, Starship Flight 14 mission page: https://www.spacex.com/launches/starship-flight-14
- Spaceflight Now, booster arrives at pad ahead of Flight 14, 21 September 2026: https://spaceflightnow.com/2026/09/21/spacexs-super-heavy-booster-arrives-at-pad-ahead-of-first-orbital-starship-launch/
- ARK Invest newsletter #445, Starlink and Wright's Law: https://www.ark-invest.com/newsletters/issue-445
- ARK Invest newsletter #514, 15 June 2026: https://www.ark-invest.com/newsletters/issue-514
- Stocktwits, Cathie Wood on Starship and the $1 billion per launch estimate: https://stocktwits.com/news-articles/markets/equity/cathie-wood-starship-first-orbital-flight-ipo-deep-value-opportunity/cZtY7DfRB2g
- Invezz, Brett Winton on Starlink capacity (about 500 Tbps), 4 June 2026: https://invezz.com/ca/news/2026/06/04/ark-invest-starlink-alone-is-sufficient-to-command-dollar2t-valuation/
- Brian Wang, Next Big Future, Wall Street is pricing the wrong rocket, 19 September 2026: https://www.nextbigfuture.com/2026/09/bank-of-america-estimates-space-ai-at-160b-gw-wall-street-is-pricing-the-wrong-rocket.html
- Brian Wang, Next Big Future, SpaceX Starship Can Make Space AI Win on Cost, 29 July 2026: https://www.nextbigfuture.com/2026/07/spacex-starship-can-make-space-ai-win-on-cost.html
- Brian Wang, Next Big Future, SpaceX Space AI Ramp (uses the name "Starthink"), 12 March 2026: https://www.nextbigfuture.com/2026/03/spacex-space-ai-ramp.html
- Brian Wang, video on orbital AI cadence and costs: https://youtu.be/N0Du3v0AJJc
- Yahoo Finance, Musk on 10,000 Starship launches a year: https://finance.yahoo.com/markets/stocks/articles/elon-musk-says-spcx-could-044404010.html
- EDF, Sizewell B: https://www.edfenergy.com/energy/power-stations/sizewell-b
- US Bureau of Reclamation, Hoover Dam power FAQ: https://www.usbr.gov/lc/hooverdam/faqs/powerfaq.html
- IEEE Spectrum, How Stupid Would It Be to Put Data Centers in Space?, 26 February 2026: https://spectrum.ieee.org/orbital-data-centers
- Futurum Group, Orbital Computing Can Reach $1 Trillion Addressable Market by 2030: https://futurumgroup.com/research-reports/orbital-computing-can-reach-1-trillion-addressable-market-by-2030/
- Wood Mackenzie, Orbital data centres cost three times more than terrestrial alternatives, 18 June 2026: https://www.woodmac.com/press-releases/wood-mackenzie-orbital-data-centres-cost-three-times-more-than-terrestrial-alternatives-as-global-power-demand-heads-for-3700-twh
- CNBC, Starcloud trains first AI model in space, 10 December 2025: https://www.cnbc.com/2025/12/10/nvidia-backed-starcloud-trains-first-ai-model-in-space-orbital-data-centers.html
- TechCrunch, Starcloud raises $250 million, 21 August 2026: https://techcrunch.com/2026/08/21/starcloud-raises-200-million-for-orbital-data-centers-as-launch-options-dry-up/
- Google Research, Exploring a space-based, scalable AI infrastructure system design (Suncatcher): https://research.google/blog/exploring-a-space-based-scalable-ai-infrastructure-system-design/
- Reuters, Google plans first test of AI chips in space under Project Suncatcher, 24 September 2026: https://www.reuters.com/business/media-telecom/google-plans-first-test-ai-chips-space-under-project-suncatcher-2026-09-24/
- SpaceNews, China launches first of 2,800 satellites for AI space computing constellation, 14 May 2025: https://spacenews.com/china-launches-first-of-2800-satellites-for-ai-space-computing-constellation/
- Xinhua, China launches space computing satellite constellation, 14 May 2025: https://english.news.cn/20250514/420ee02233394205b0e9aaf4dbc344bb/c.html
- Stanford PH240, rocket CO2 emissions estimate (Dear, 2021): http://large.stanford.edu/courses/2021/ph240/dear1/
- Space.com, How environmentally friendly is SpaceX's Starship?, 21 March 2024: https://www.space.com/spacex-starship-rocket-launches-environmental-impact
- US EPA, Fast Facts on Transportation Greenhouse Gas Emissions: https://www.epa.gov/greenvehicles/fast-facts-transportation-greenhouse-gas-emissions