Space Data Centers: Have You Looked Up Lately?

Nov 2, 2026 · @Budget Engineer

TL;DR

• Orbit is already crowded. About 16,000 to 17,000 working satellites and roughly 35,000 to 47,000 tracked objects, depending on who's counting. Starlink alone dodged collisions over 355,000 times in a year.
• It's getting worse fast. If every satellite stopped dodging, a likely collision would come in about 2.2 days. In 2018 that was 164 days. ESA says the long-term health of orbit got ten times worse in a single year.
• Now five companies want to add 1.26 million data center satellites. That's 75 to 80 times what's working up there today. None are approved yet, and all five asked to skip the usual deployment deadlines.
• The physics doesn't play nice. Space doesn't cool things for free, the chips weren't built for radiation, and nothing up there can be repaired. One expert model says four of the five plans risk runaway debris on their own.
• Astronomers say the limit is about 100,000 faint satellites, total. The proposals blow straight past that.
• The bottom line: a bit of computing on satellites already going up makes sense. A million-satellite land grab doesn't. These decisions are happening at the FCC right now, and hardly anyone is watching.

Every number below is linked to its source, and where sources disagree, I show both.

Low Earth orbit is already so crowded that Starlink's satellites alone dodge something roughly every 90 seconds, and the current big idea is to add over a million more of them. For AI. Obviously.

The pitch goes like this: data centers on the ground are eating power grids, drinking water, and annoying everyone who lives near them. Space has free solar power, it's freezing cold, and nobody files a noise complaint from orbit. So shove the GPUs up there and call it the future.

It sounds clever right up until you look at what's already circling the planet. So let's do that first, then we'll get to the bit where the physics laughs at the pitch deck.

Why the numbers in this post don't all match

Every figure here links to where it came from, and you'll notice they don't always agree. Trying to make sense form it all here:

• Different counters count different things. "Active", "functioning" and "operational" sound the same but aren't defined identically, and each catalogue updates on a different day.
• Tracked vs estimated. Catalogues list what radars can actually follow, roughly 10 cm and up. Everything smaller is a model estimate, and models get revised. ESA's own numbers changed between 2025 and 2026, so older articles still quote the old ones.
• Company figures are company figures. Manoeuvre counts come from SpaceX's own reports to the FCC. Satellite counts in filings are the maximum a company is asking for, not hardware that exists.

Where sources disagree, I show the range and who said what. If anything, the spread proves the point: nobody has a perfect picture of what's up there.

Orbit is already a car park

Depending on who's counting, there are roughly 15,900 to 16,900 working satellites in orbit right now. That's up from about 2,000 in 2019 (Orbital Radar).

The sources broadly agree on one thing: about two in three of those are Starlink. McDowell counted 11,143 working Starlinks out of the 16,914.

Working satellites are the minority of what's up there. Add dead satellites, spent rocket stages and shrapnel, and the tracked total is somewhere between 35,000 and 47,000 objects:

ESA counts everything regularly tracked by surveillance networks, while KeepTrack works from the public catalogue. I couldn't find a breakdown of exactly what makes up the gap, so treat the honest answer as "somewhere in that range".

And "tracked" only covers things roughly 10 cm and up. Everything smaller is a model estimate, and ESA revised its own model this year:

Sources: ESA statistics page, 2025 snapshot and ESA statistics, Oct 2026, via New Space Economy. If an article tells you "1.2 million", it's quoting last year. At orbital speed, a centimetre-sized fragment can wreck a satellite.

The pace isn't slowing either. ESA counted more than 300 launches and over 4,000 new payloads in 2025 alone, about ten new satellites a day (ESA Space Environment Report 2026).

Everything up there is constantly dodging

Starlink satellites made 207,152 collision-avoidance manoeuvres between December 2025 and May 2026, and over 355,000 across the full year (Space.com). That works out to each satellite firing its thrusters to get out of the way roughly once a week.

[embed: node/664681ff-e2f5]

To be fair to SpaceX, part of this is them being extra careful. Starlink dodges whenever the collision risk tops 3 in 10 million, while most operators wait until it reaches 1 in 10,000 (Space.com). How much stricter that is depends on who's doing the maths: you'll see "100 times" in some write-ups and "300 times" in others (KeepTrack). Run it yourself and it's about 330. SpaceX reports no collisions so far.

The catch: debris expert Hugh Lewis points out each dodge still leaves roughly a one-in-a-million chance of a hit. Make a million dodges and that leftover risk adds up . He expects Starlink to pass a million total manoeuvres around mid-2027, and to be making over a million a year by 2030 (Space.com).

One more source clash worth knowing: SpaceX's year-end summary put 2024 at about 200,000 manoeuvres (KeepTrack), but the half-year reports in the chart add up to roughly half that. Either way the line only goes one direction.

Then there's the uncomfortable question of what happens when the dodging stops. Researchers from Princeton, UBC and the University of Regina built a metric for exactly that, the CRASH Clock: if every satellite lost the ability to manoeuvre at once, how long until a likely collision between tracked objects? Their official tracker puts it at 2.2 days as of 15 August 2026, down from 164 days in January 2018 (Outer Space Institute). Along the way it read 5.5 days in June 2025, 3.8 days in January 2026 and 2.5 days in May 2026, all on the same tracker using the same method. It only counts objects big enough to track, so if anything it a bit optimistic.

If you go digging you'll also see 2.8 days quoted for the end of 2025, alongside a different 2018 starting figure (Live Science). Those appear to come from an earlier version of the study, which the team has since revised after feedback from other researchers (preprint). They don't line up with the tracker's series, so don't mix the two. Either way the story is the same: the safety margin has collapsed.

What would make every satellite stop manoeuvring at once? A big enough solar storm. During the May 2024 Gannon storm, the swollen upper atmosphere dragged satellites down, automated station-keeping kicked in (mostly Starlink's), and nearly half of all active satellites in LEO manoeuvred at once (Parker & Linares, Journal of Spacecraft and Rockets). Those weren't collision dodges, but mass movement like that scrambles everyone's orbit predictions at the worst time. A Carrington-class storm would be far worse. We've built a system with roughly a two-day margin for error and pointed it at the Sun.

Kessler syndrome isn't sci-fi any more

ESA boils the long-term health of orbit down to one number, the Space Environment Health Index. It projects the combined long-term effect of everything in orbit and compares it with a 2014 benchmark scenario in which debris rules are followed well. A score of 1 is sustainable. Anything higher means that many times more strain on the orbital environment than is sustainable. Last year's report put us at about 4. This year's, based on data to the end of 2025, put us at about 50 (ESA). In ESA's words, already-unsustainable behaviour got an order of magnitude worse in a single year. One caveat: ESA's public pages don't spell out whether the index was calculated over exactly the same time horizon both years, so treat the jump as ESA's own headline figure rather than a lab-precise comparison.

ESA also cut its long-range forecast from 200 years to 100, and the shorter forecast still came out worse than last year's longer one. Its conclusion: debris would keep growing even if we never launched another satellite, and actively removing old junk is now required, not optional (ESA Space Environment Report 2026).

That runaway is Kessler syndrome: one collision makes a cloud of fragments, those fragments hit other satellites, which makes more fragments, and so on until whole orbital bands become unusable. It wouldn't happen overnight like in Gravity. It's a slow grind that makes orbits progressively more expensive and dangerous until people stop using them.

Altitude matters a lot. Debris around 550 km, where most of Starlink lives, drags down and burns up within a few years. Debris at 800 km hangs around for decades. That band is still littered with leftovers from China's 2007 anti-satellite test and the 2009 Iridium-Cosmos collision (TechSpot). Remember that when we get to where the data center constellations want to fly.

And it's not only old junk. On 29 March 2026, Starlink 34343 broke apart at about 560 km, throwing out tens of trackable fragments (Tom's Hardware). It was the second Starlink breakup in about three months, after Starlink 35956 in December 2025, and neither appears to have been caused by a collision (SmallSat News). Fair's fair: at that altitude the fragments were expected to fall out of orbit within weeks (SpaceNews summary). But it shows even the "best-run" constellation produces debris it didn't plan for.

The report also added a new metric this year: the risk of reentering junk hurting someone on the ground. It's still small, but it's rising because so much more is coming down. ESA counts more than three intact satellites or rocket bodies reentering every day (ESA).

The one bit of good news: controlled rocket-body reentries have outnumbered uncontrolled ones for the second year running, and ESA credits the switch from a 25-year to a 5-year disposal target for shortening how long dead hardware hangs around. Credit where it's due ish.

Now add a million data centers

Five companies have asked the FCC for permission to launch a combined 1,259,600 data center satellites. That's roughly 75 to 80 times the number of working satellites in orbit today, depending on whose count you use.

Why coverage gets the dates muddled: FCC file numbers have a date baked in that often isn't the filing date. SpaceX's file number says 8 January, but the FCC's own notice says it was filed on 30 January. Cowboy's says 23 March, but the FCC lists it as filed on 11 May, and some outlets report 14 May (SatNews). Add in outlets mixing up "filed" and "accepted", and you get reports that disagree by weeks. The dates above come from FCC notices wherever I could find them; Cowboy's acceptance date comes from secondary reporting.

All five have been accepted for filing. That only means the FCC found the paperwork complete enough to open for public comment. It is not an approval.

To be clear, these are maximums. Nobody is launching 1.26 million satellites next year. But a licence for a million is a licence for a million, and all five applicants asked the FCC to waive its milestone rule (section 25.164), which normally requires half a constellation to be up within six years and all of it within nine (SpaceX, Starcloud, Orbital, Blue Origin, Cowboy). SpaceX, Starcloud and Orbital also asked to skip the surety bond that normally backs that promise. That's asking for the land grab or well the space grab without the deadline. Rules for thee type thing.

Some of the attitude on show: SpaceX's filing calls the project a first step toward a Kardashev II civilisation (The Register). Musk's take on crowding was that space is "really big" so it won't get crowded (Gizmodo). The entire first half of this post is the answer to that. Clearly its already crowded.

Google is going about it more quietly. Its Project Suncatcher prototype, built with Planet, went up on a SpaceX rocket on 1 October 2026 carrying four of Google's TPU chips. Because of heat limits, the chips can only run Google's Gemma model for 15 minutes at a time, and Google's project lead says the radiators are one of the heaviest parts of the satellite (same NPR report). For scale, four TPUs is roughly the compute of one server in a normal data center (Washington Post, via Hawaii Tribune-Herald). There's no mega-constellation filing from Google.

As of early November 2026, I could find no FCC approval for any orbital data center constellation (Broadband Breakfast). The SpaceX application alone drew nearly 1,500 public comments (ESO), and Amazon and Viasat both asked the FCC to reject it (DCD). But FCC chair Brendan Carr has been openly warm to SpaceX and said he didn't expect Amazon's objection to get much traction (CNBC).

Where the physics starts taking the piss

Even the US government's auditors aren't sold. The GAO's April 2026 review found that small systems processing data that's already in space may be closer to maturity, but giant AI data centers are a different story. It called cooling at data center scale unproven, and said the solar arrays needed could be bigger than anything ever launched and assembled in space (GAO-26-109012, summary). It also flagged orbital congestion, radiation damage and end-of-life debris as open problems.

"Space is cold, so cooling is free." No. Space is cold, but it's also empty, and empty is an insulator. Fans and water work by moving heat into air or liquid. In a vacuum there's nothing to move it into, so the only way to dump heat is to radiate it away as infrared. Space isn't a freezer, it's a thermos.

Radiating heat needs surface area, lots of it. The ISS sheds about 70 kW of heat, so a one gigawatt data center would have to shed as much as roughly 14,000 space stations (Peike Li).

How much would that weigh? One widely quoted back-of-envelope estimate comes from a World Economic Forum piece by Tony Pan, founder of Borrowing Arrows. The ISS's cooling system uses about 7 tonnes of radiator wings to shed 70 kW. Scale that up in a straight line and a modest 1 MW data center needs roughly 100 tonnes of radiators, while 1 MW of current AI server racks weighs about 10 tonnes. So on ISS-era technology, the cooling could weigh about ten times the computer, and you're paying to launch every kilo (WEF).

Read that as an order-of-magnitude warning, not a spec sheet. Purpose-built radiators could be lighter than the ISS's, and running chips hotter shrinks them, though Pan points out today's chips can't go much hotter. Starcloud's own white paper gives another angle: a radiator at 20°C sheds only about 633 W per square metre, so 1 MW needs roughly 1,600 m², about the size of a hockey rink (same WEF piece).

Exact radiator figures vary a lot by source, because they depend on how hot you let the chips run and what you count. Even the ISS's radiator area is quoted anywhere from about 325 m² (SoftwareSeni) to around 1,000 m² (LAVX). The 70 kW figure is the one everyone agrees on, and the conclusion holds at any of these numbers: cooling in orbit costs huge amounts of surface area and mass.

In fairness, it's not impossible. Scott Manley ran the numbers on a Starlink-sized satellite burning 20 kW on GPUs and got about 23 m² of radiator at 80°C (summary). That works. It's just that 20 kW is less than a single rack of today's top-end AI servers, never mind a data center, which is why the plans need hundreds of thousands of satellites in the first place.

Radiation. To get real compute density you need the same kind of Nvidia and Google chips used on the ground, and those are soft targets in space (IEEE Spectrum). The best news so far is Google's: its Trillium TPUs survived a proton-beam test at a higher dose than a five-year mission would deliver, without hard failures (Sundar Pichai, via NewKerala). But surviving isn't the same as computing correctly. Memory errors and silent data corruption seen in testing still need more study, according to a summary of Google's paper (Jared Watkins). And that's one chip family, tested by the company that makes it.

Repairs. On Earth, a dead GPU means a tech walks over and swaps it. In orbit it means the satellite is now junk. Carnegie Mellon's Brandon Lucia puts the cost and complexity of all that at 10 to 100 times higher than on the ground (OPB/NPR). Microsoft already tried a hard-to-reach data center with Project Natick, sinking servers off the Scottish coast, then confirmed in 2024 that the project was over, even though the servers had held up well (DCD). Orbit is a lot harder to reach than the seabed.

The environment up there. Radiators are big, flat targets in a debris field, and UV and atomic oxygen slowly eat their coatings (IEEE Spectrum). So the biggest part of every satellite is also the part most likely to get pinged with holes.

Do the maths on a million satellites

Someone has actually modelled this. In July 2026, debris researcher Hugh Lewis of the University of Birmingham posted "Critical Sizes of Satellite Constellations", which tests how big a constellation can get before its own breakups produce fragments faster than they clear. He looked at 16 constellations. Six came out over the "runaway" line, where fragments keep multiplying without limit: SpaceX's Starmind, Blue Origin's Project Sunrise, Cowboy's Stampede, Starcloud, China's Guowang and E-Space's combined constellations (arXiv, Phys.org).

Four of those six are data center plans, but the confidence isn't the same for each:

• Starmind: the clearest case. Still over the line when Lewis re-ran it with SpaceX's own stated satellite size (3,000 kg, 800 m²).
• Project Sunrise: Blue Origin's filing didn't give enough detail, so Lewis made assumptions (300 Kuiper-sized satellites at 1,800 km). On those, the runaway threshold at that altitude is less than one satellite, so he judged it likely over the line even with very different satellites.
• Stampede: over the runaway line if its satellites are Starlink-sized. With smaller, Kuiper-sized satellites it's "unstable" instead, meaning debris grows to a high level rather than without limit.
• Starcloud: only crosses the runaway line if its production satellites are at least twice the size of its small first test satellite. Plausible given its plans, but not settled.

Orbital Compute's 100,000-satellite plan wasn't part of the study.

Two caveats pull in opposite directions. It's a preprint, so not yet peer reviewed. But Lewis also says the model is optimistic: it tests each constellation alone in an otherwise empty sky, assumes collision avoidance works perfectly, and ignores existing debris. Add everything that's already up there and the danger point arrives with fewer satellites (arXiv).

Then there's the replacement maths. A million satellites with a five-year life means replacing 200,000 a year, forever. Amazon told the FCC that's more than 44 times the entire world's satellite launches in 2025 (DCD). Amazon obviously has an axe to grind with SpaceX, so I checked it against ESA's figure of about 4,000 payloads in 2025. That gives roughly 50 times. Same ballpark.

Run it forward and it gets silly fast:

• Reentries. 200,000 a year works out to about 550 dead satellites burning up every day. ESA currently counts a bit over three intact objects a day, and it's already flagging rising risk on the ground.
• Failures. SpaceX says its disposal success rate is above 99% (Converge Digest). One percent of a million is still 10,000 dead, uncontrollable satellites. That's about the size of today's entire Starlink fleet, as fresh debris.
• Altitude. The filings reach up to 2,000 km. Remember the 800 km band still full of junk from 2007? Above that, a dead satellite isn't a few-year problem, it's a multi-generational one.
• Dodging. 11,000 Starlinks already need over 355,000 manoeuvres a year. Close approaches climb faster than the satellite count, because every new satellite is a new target for every existing one.
• The sky. ESO's July 2026 study says that to protect astronomy, there should be no more than about 100,000 faint satellites in orbit, total. Proposals in the pipeline add up to around 1.7 million (ESO). For SpaceX's constellation alone, it found dozens of satellite trails in each image from the Very Large Telescope two hours after dark, costing up to 28% of the field of view (Gizmodo). Earlier comments from the same astronomer put it at up to 10% of pixels (Space.com), so expect to see both numbers around. Either way, from a properly dark site the sky would look like the suburbs of a small town, plus hundreds of moving satellites (Laser Focus World).

And that CRASH Clock at 2.2 days? That's with around 17,000 working satellites. Nobody has published what it looks like with another 1.26 million, and I'm not sure that would be very sad to look at.

Where orbital compute actually makes sense

There is a real use case, and it's small. When the data is already in orbit, processing it up there beats shipping it down. Earth observation satellites generate huge amounts of imagery, and downlink bandwidth is the bottleneck. Crunch it on board, send down the answer instead of the raw pixels (WebProNews).

That's why the early startups mostly target niche, latency-insensitive jobs like processing imaging data and disaster-recovery storage (Bisnow). Even analysts who like the idea see orbital compute as a sidekick to ground data centers, not a replacement (Broadband Breakfast).

Edge compute on satellites that already exist? Sensible. Moving hyperscale AI training off the planet? That's a different claim entirely.

The verdict

Space data centers at the scale being pitched are a bad idea, and the orbit we've already got is the main reason why.

In about eight years we went from a 164-day safety margin to just over two days, mostly because of one constellation. ESA's health score went from 4 to 50 in a single year. Every Starlink dodges about once a week, and Starlinks are breaking apart on their own. Into that, five companies want to add up to 1.26 million more satellites, and a preprint model flags four of those five as at risk of runaway debris on their own, Starmind under every assumption tested.

Then they want cooling that, on one ISS-based estimate, could weigh up to ten times the compute, the chips to shrug off radiation, and nothing to ever need a repair.

The honest version of this idea is small: a bit of edge compute on satellites that are going up anyway. The headline version is a land grab, and orbit is a shared resource we don't get to reset when the hype cycle moves on.

Why this needs more eyes on it

These decisions are being made at the FCC right now, and most people have never heard of the filings. You'll see different numbers depending on the source, as this post shows. But every source points the same way: more crowded, more fragile, getting worse faster.

If you want to do something about it:

• Share it. Most of the pushback so far has come from astronomers and rival companies. The public is barely in the conversation.
• Look at the filings yourself. The applications and public comments are public record, searchable by file number (they're in the table above) on the FCC's satellite filing system, ICFS. The SpaceX application alone drew nearly 1,500 comments (ESO).
• Follow the people tracking it. ESA's Space Environment Report comes out every year, and ESO has led the push on night-sky impacts.

If a regulator signs off on a million satellites, it won't be because the physics worked out. It'll be because not enough people were paying attention.

Sources

Grouped by topic. Where a figure came from a company's own report or a single outlet, the text says so.

Satellite and debris counts

• KeepTrack: How many satellites are in space (Oct 2026)
• KeepTrack: How many satellites does the US have, incl. McDowell count (Oct 2026)
• New Space Economy: ESA statistics, 23 Oct 2026
• New Space Economy: ESA debris statistics, 14 Oct 2026
• ESA Space Environment Statistics page (2025 snapshot)
• Orbital Radar: satellites by country

ESA's 2026 report and Kessler risk

• ESA Space Environment Report 2026
• ESA: Health Index 2026
• Lewis, "Critical Sizes of Satellite Constellations", arXiv preprint (Jul 2026)
• Phys.org: Satellite constellations could create spiraling space debris
• BigGo Finance: Lewis on crowded-orbit thresholds
• TechSpot: CRASH Clock and debris altitudes
• Tom's Hardware: Starlink 34343 fragment event
• SmallSat News: Starlink 34343 vs 35956
• SpaceNews summary: Starlink 34343 debris decay

Collision avoidance and the CRASH Clock

• Space.com: Starlink dodges almost weekly (Jul 2026)
• KeepTrack: collision avoidance and conjunction screening
• Space.com: Starlink 50,000 manoeuvres
• Space.com: Starlink avoidance growth stalls (2023)
• Advanced Television: Starlink 144,404 manoeuvres
• CRASH Clock preprint, arXiv 2512.09643 (revised)
• Outer Space Institute: CRASH Clock
• Live Science: CRASH Clock under 3 days
• Parker & Linares: Satellite drag during the May 2024 Gannon storm (J. Spacecraft and Rockets)

The orbital data center filings

• The Register: SpaceX's million-satellite filing
• Converge Digest: SpaceX filing details
• DCD: Orbital files for 100,000 satellites
• Light Reading: Orbital and SpaceX 100K filings
• Interesting Engineering: Starcloud 88,000 satellites
• ISS Tracker: Blue Origin Project Sunrise
• FCC notice: Cowboy Space Stampede
• SatNews: Cowboy raises $275M and files
• ISS Tracker / SpaceNews: Cowboy milestone waivers
• OPB/NPR: Google Project Suncatcher launch
• Broadband Breakfast: no orbital DC approved
• CNBC: FCC chair on Amazon's objection
• DCD: Amazon Leo asks FCC to refuse SpaceX

Engineering problems

• GAO-26-109012: Data Centers in Space (Apr 2026)
• Let's Data Science: GAO summary
• IEEE Spectrum: Why orbital data centers are harder than Silicon Valley thinks
• World Economic Forum: space data centre cooling
• Peike Li: Space is a thermos
• SoftwareSeni: the physics tax on cooling
• LAVX: why cooling isn't the real problem
• Scott Manley video summary: cooling a datacenter in space
• WebProNews: The orbital mirage
• Bisnow: Big tech's data center space race

Astronomy

• ESO: "Beyond the limit" press release (Jul 2026)
• Gizmodo: ESO study, 100,000-satellite limit
• Laser Focus World: how many satellites is too many
• Space.com: astronomers on SpaceX's data centers
• Space.com: RAS and ESO objections