This week, Reuters reported something quieter and more revealing than the usual orbital-data-center hype. Space companies have started talking to insurers about covering AI hardware in orbit. Blue Origin, plus a cluster of startups — Starcloud, Lonestar Data Holdings, Orbital, Cowboy Space — have all signaled they want to put compute above the atmosphere.1
The pitch is clean enough to fit on a napkin… AI is starving for power, terrestrial grids can’t keep up, and the sun is free up there, twenty-four hours a day, with no permitting fight and no water table to drain.
The reporting buried the actual story in the back half. The insurers, the people whose entire profession is putting a number on tail risk, are struggling to price it.
- Lack of historical data.
- No way to model the failure modes.
- No clean method to value an AI accelerator that’s been turned into shrapnel.2
That detail is not a footnote. It’s the whole thing.
When the industry built around pricing catastrophe says it cannot price your catastrophe, that is not a paperwork delay. That is the market telling you the tail is fatter than the pitch deck admits.
And it isn’t academic, either. Securing coverage is effectively the gate to the debt financing these projects need to scale. No model, no policy. No policy, no capital stack. The thing that can’t be priced is the thing that stalls the whole enterprise.
So let’s do the math, the pitch decks skip.
A penny arriving like a shaped charge
Start with velocity, because everything downstream is velocity squared. NASA’s Orbital Debris Program Office… in low Earth orbit, debris circles the planet at roughly 7 to 8 km/s. But the average impact speed between debris and another object is about 10 km/s, and it can reach 15 km/s… more than ten times the speed of a rifle bullet.3
The European Space Agency draws the damage thresholds. A 1 mm fragment can take out a subsystem. An object larger than 1 cm will most likely disable a spacecraft outright and punch through ISS shielding. A 10 cm cataloged object means catastrophic disintegration of the target.4 ESA’s working rule for “catastrophic” is brutally simple: an energy-to-mass ratio above 40 joules per gram.5
A U.S. penny is 19 millimeters across. That is comfortably over the 1 cm “will disable a spacecraft” line. And a penny is 2.5 grams. Run the kinetic energy:
- At 10 km/s: about 125,000 joules: roughly 30 grams of TNT equivalent.
- At 15 km/s: about 281,000 joules: roughly 67 grams of TNT.
A penny is not a coin up there. It is a shaped charge with a return address you’ll never read.
And there is no air to slow it down. On Earth, the atmosphere is a free, planet-wide debris shield that vaporizes most of this before it matters. In orbit, that shield is gone. The penny arrives with all of its energy intact.
The trap is in the business model itself
Here is the part the “abundant solar” crowd never says out loud: the variable you optimize is the variable that kills you.
To collect a gigawatt of power, you need area. Sunlight in space delivers about 1,361 watts per square meter; high-end space cells convert maybe 30 percent of it. That math forces roughly three square kilometers of solar array for a single gigawatt-class facility… before you’ve added the radiators. And you must add the radiators, because in a vacuum, you cannot convect heat away. You can only radiate it slowly, which means another square kilometer or more of panel just to dump the waste heat your chips produce.
On Earth you bolt a data center to bedrock. In orbit you unfurl it into a kilometers-wide sail and ask the debris field to please aim elsewhere.
Every square meter you add to collect more sun, or shed more heat, is another square meter of cross-section for the debris flux to find. The thing that makes orbital power abundant is the same thing that makes the facility an enormous, unmissable target.
The numbers, run honestly
Collision risk follows a clean model used by NASA and the National Research Council alike: expected impacts equal flux × area × time, and survival probability is the exponential decay of that.6 Plug in conservative, low-orbit debris flux and a five-year operating life.
For ≥1 cm debris, the “disables a spacecraft” threshold: a gigawatt facility with even a discounted half-square-kilometer cross-section expects on the order of fifty such impacts over five years. The probability of going un-hit rounds to zero. Not “unlikely.” Zero, to twenty-some decimal places.
For ≥10 cm debris, the “catastrophic disintegration” threshold: the same facility carries roughly a 22 percent chance of total loss over five years, about 5 percent a year. That’s per facility.
Now do what makes it a fleet. A space-based compute strategy that matters needs dozens of these. At fifty facilities:
- Expected catastrophic losses over five years: about eleven.
- Expected catastrophic losses per year: roughly two to three.
- Probability the entire fleet escapes a single catastrophic strike: about 0.0004 percent.
You are not insuring against a tail event. You are budgeting for a recurring one.
These aren’t back-of-the-envelope fantasies. The National Research Council’s technical assessment put the odds of a 10-square-meter spacecraft taking a ≥1 cm hit over a ten-year life at between one in a hundred and one in a thousand. Scale that cross-section up by four or five orders of magnitude, which is exactly what a kilometers-wide power-and-radiator structure does, and the comfortable “one in a thousand” curdles into “certainty.”
And then it gets worse
Everything above assumes each facility’s fate is independent. It isn’t.
Every catastrophic break-up is itself a debris generator — thousands of new fragments, each now a projectile for everything still in orbit. ESA calls the runaway version of this the Kessler syndrome, and notes that beyond 10 cm, impacts are assumed to trigger exactly that cascade. The 2007 Chinese anti-satellite test alone scattered more than 2,000 trackable fragments; the 2009 Iridium–Cosmos collision added roughly 1,800 more.7 These clouds don’t disperse on any human timescale.
Independence is the optimistic assumption. In reality, the first loss makes the second one likelier. The fleet doesn’t fail one at a time. It fails as a chain reaction.
ESA’s 2025 environment report contains the sentence that should end the conversation: at around 550 km altitude, there are now the same order of magnitude of threatening debris objects as there are active satellites.8 The neighborhood the optimists want to build in is already half junk.
The honest counterargument
To be fair to the people building this, and some of them are serious, there are real rebuttals.
Debris density isn’t uniform. The worst band sits around 800 to 1,400 km, where dead satellites and old fragment clouds pile up. Lower orbits, below ~600 km, partly self-clean as atmospheric drag drags debris down over years. Fly low, and you cut your exposure.
The cataloged≥10 cm objects can, in principle, be dodged, though a kilometers-wide structure maneuvers like a barge and presents a target hundreds of times an ordinary satellite’s size, so “dodge it” is a much weaker promise here than for a compact spacecraft. And not every ≥1 cm hit is fatal to a distributed structure; lose a panel section, route around it, keep computing.
All true. None of it rescues the economics, because the constraints that aren’t debris all get worse with scale, too. Heat rejection gets harder the bigger you build. Radiation degrades the chips regardless…
Starcloud flew an Nvidia H100 to orbit, and its own CEO conceded an H100 is “probably not the best chip for space.” Servicing, the thing that makes terrestrial data centers maintainable, is the one thing orbit takes off the table entirely.
What the insurers already know
The bull case optimizes a single variable, solar flux, and treats it as the whole equation. The other variables, the ones that don’t show up on the napkin, are heat you can’t shed, radiation you can’t escape, hardware you can’t service, and a debris field that grows denser every time something up there dies.
The clearest read on whether this works isn’t a physics paper. It’s a quote sheet. The people who price catastrophe for a living looked at orbital AI infrastructure and couldn’t write the number.
That blank space on the policy is the answer. Free solar. Fatal math.
Sources
A note on the math: kinetic-energy figures assume a 2.5 g modern U.S. penny; TNT equivalence uses 4,184 J/g. Collision estimates use a Poisson model (P[no hit] = e^−flux·area·time) with conservative low-LEO debris flux (~2×10⁻⁵/m²/yr for ≥1 cm, ~10⁻⁷/m²/yr for ≥10 cm) and an effective cross-section of ~0.5 km² for a gigawatt-class facility. These are order-of-magnitude inputs; the conclusion is robust across the plausible range, but swap in exact ESA MASTER / NASA ORDEM flux-by-altitude figures before treating any single number as precise.