Musk’s Space Data Centers Are a Bet That Moore’s Law Stays Slow
It's not a cooling bet, and barely a launch bet. It's a bet on how slowly Moore's Law moves.
Elon Musk wants to put AI data centers in space, and the internet has decided this is stupid. The favored reason is thermodynamic: space is a vacuum, there’s no air, how do you cool a gigawatt of GPUs with no air to carry the heat away? The dunk writes itself. The man forgot about heat.
I’m a physicist. I don’t have a horse in this race, and I find the dunk-or-defend reflex tedious. There is exactly one honest way to settle whether orbital data centers make sense, and it’s the way Musk’s own engineers would do it: write down the physics, write down the costs, and turn the crank. The answer falls out — and it turns out to be neither side’s answer.
So let’s turn the crank. No thesis. Wherever the arithmetic lands is where we land. I’ll cite every number; the full table is at the foot of this post. And every time I hit a judgment call, I’ll round it in Musk’s favor — generous emissivity, best-in-class solar, the friendliest orbit — so that whatever the ground wins, it wins on the orbital case’s best day.
First, the heat — because that’s what everyone’s arguing about
A hot object sheds heat three ways: conduction, convection, radiation. Conduction needs contact. Convection needs a fluid — air or water — to carry warmth off. A vacuum offers neither, so in orbit a radiator has one channel: thermal radiation, governed by the Stefan-Boltzmann law,
where q is watts radiated per square meter, ε is surface emissivity (a good coating reaches about 0.9), σ is the Stefan-Boltzmann constant (5.67 × 10⁻⁸ W/m²K⁴), T is the radiator’s temperature, and T_env is what it faces. Temperatures in kelvin, since the fourth power only behaves measured from absolute zero.
The intuition that excites people is the cold of space — 2.7 K, the leftover chill of the Big Bang. Surely radiating into a 2.7 K void beats a warm room? Plug in the numbers. A radiator at 75 °C (348 K), emissivity 0.9:
Radiating into deep space roughly doubles the heat you shed by radiation, versus a room-temperature sink. Real, and worth having.
But that doubling is the entire radiative prize, and it’s smaller than it looks, because radiation isn’t the only way to cool something — it’s just the only one left in a vacuum. On the ground that same panel also has convection: air moving across it, carrying heat off directly. With a fan, convection alone removes several thousand watts per square meter — multiples of what radiation manages in space. So on cooling, the honest tally is: orbit gives you a 2× bump on the radiative channel and takes away the convective channel, which was worth far more. On cooling alone, the ground wins. (This is the part the critics get right; Sam Harsimony made the case cleanly: build one radiator, fly a copy, bolt the other to a warehouse floor — the warehouse copy sheds more heat and never paid for a launch.)
One honesty note on my own number: a real radiator in low Earth orbit doesn’t face a clean 2.7 K universe. It sees the Sun, sunlight reflected off the Earth, and the Earth’s own infrared glow. NASA’s thermal-control models treat spacecraft temperature as a balance of all those inputs plus what the radiator sheds. So 750 W/m² is an optimistic, deep-space-facing figure — again, rounding in orbit’s favor.
So the critics are right that orbit is a worse place to dump heat. But here’s the move that the whole debate misses: being a worse radiator is not the same as being uneconomic. The question was never “is space good at cooling.” It’s “does the whole machine pencil out.” For that we need the radiator’s mass, not its efficiency — and then everything else.
Now turn the crank on the whole system
Build the satellite, per megawatt of compute, and price it against a data center on the ground.
How much radiator? At 75 °C, two-sided, at a spacecraft-radiator areal density of 5 kg/m² (real flight hardware runs heavier), the Stefan-Boltzmann number above gives 3.3 tonnes per megawatt. Three honesty notes, because a thermal engineer will check. A 75 °C radiator implies running the chip junctions hotter still — around 100 °C, since you need a temperature gradient to push heat from junction to coolant to panel — which is aggressive but plausible for purpose-built silicon (unlike a terrestrial DGX, which wants 30 °C air). The 750 W/m² assumes the panel mostly faces deep space; a face turned toward the Earth, which glows in the infrared at around 255 K, sheds noticeably less — which means the satellite has to hold an edge-on-to-Earth, sun-avoiding attitude to hit that number, and that attitude constraint isn't free, because it also dictates where the solar array is allowed to point. And the 5 kg/m² panel is the lightweight end; many flight radiators are heavier. All three assumptions are the generous-to-Musk end, which means 3.3 tonnes is a floor on radiator mass — and even at the floor it's almost an afterthought. The critics were right that the radiator is large in area — a megawatt needs about 1,300 m² of radiating surface, or roughly 670 m² of physical two-sided panel: a quarter of a football field if you count both faces, an eighth if you count the deployed footprint — but in mass, the thing a rocket cares about, it barely registers. Costed honestly, the heat problem everyone's dunking on is a few tonnes.
What else flies? Compute, first. As a mass proxy, an NVIDIA DGX H100 is 130 kg and draws 10.2 kW, which is about 13 tonnes per megawatt. A real orbital machine wouldn’t be a DGX rack bolted to a satellite — it would need space-qualified packaging, shielding, thermal interfaces, and robotic-serviceable mounting, all of which add mass I’m ignoring. Then the solar array, using best-in-class flexible panels at 150 W/kg (typical space arrays manage a fifth of that): 6.7 t/MW in the best orbit. Add batteries to run through Earth’s shadow, plus structure and power electronics at a 30% allowance, and the whole satellite comes to roughly 30 tonnes per megawatt in the friendliest orbit, 41 in a generic one.
Now lift it. At today’s Falcon 9 price of about $2,700/kg, thirty tonnes costs $80M per megawatt — far above the roughly $11M/MW it takes to build a data center on the ground. The critics who say “launch is too expensive” are right about today. But Musk isn’t betting on today; he’s betting on Starship. SpaceX doesn’t publish an operational price per kilogram, so the figures everyone quotes — $100 to $300/kg — are projections for a mature, fully reusable system, not posted fares. Grant the optimistic end anyway. At a hypothetical $100/kg, thirty tonnes launches for $3M/MW — below the ground build. So the first launch is not the killer. On the narrow question “can you afford to put it up there once,” Musk wins the day Starship hits its targets.
Heat: not free, but no longer the obvious killer — 3.3 tonnes. Launch: also not the obvious killer, in a mature-Starship world. So what is? This is where the crank turns up the thing neither camp is shouting about.
The clock nobody priced
GPUs don’t last. NVIDIA ships a new architecture every 18 to 24 months, each generation improving performance per watt — though by how much is workload-specific and vendor-reported, not a clean law; NVIDIA’s own roadmap claims range from modest to 10× depending on what you measure. For the toy model I’ll assume frontier performance per watt doubles every D years and use D = 3 as a central case. That is not a law of nature; it’s the single parameter the whole result turns on, which is precisely why I’m flagging it. (I’ll call this “Moore’s law” colloquially, but the real variable is accelerator useful-work-per-watt, not transistor density.) Hyperscalers book six-year depreciation schedules, but that’s an accounting choice under pressure; Michael Burry has argued loudly the real economic life is closer to two or three years. Meanwhile the building that houses everything — concrete, power feed — lasts 15 to 30 years. Three clocks badly out of sync: chips obsolesce in a few years, the structure lasts decades.
On the ground this mismatch is boring and solved. When the chips age out, a technician wheels a cart down an aisle and swaps the cards. The expensive long-lived part — the building — stays put and amortizes over thirty years. You replace only the cheap, fast-moving part.
In orbit you cannot wheel a cart down an aisle. And this is where the entire economic question lives — but it’s subtler than “can you send a repair robot,” because there are two failure modes, not one.
The crude version: relaunch the whole satellite each refresh. If upgrading the chips means flying a fresh satellite every few years, then over a 30-year facility life you pay the full launch bill five to ten times. In this capex-only comparison, that only beats the ground if launch drops to roughly $25 to $76 per kilogram depending on refresh interval — below even the optimistic Starship projections. Possible in principle. But you’re betting the entire concept on launch prices nobody has demonstrated.
The sophisticated version: service in orbit. Suppose a robot can swap the GPU modules while the bus, radiator, solar array, and batteries all stay in place. This is the case the optimists are implicitly counting on, and it’s a real saving — you stop relaunching 30 tonnes of satellite. But it is not free, and here’s the trap the optimists miss: the replacement chips still have to reach orbit. You don’t relaunch the radiator and the solar panels, but you relaunch the compute payload — about 13 tonnes per megawatt — every single refresh cycle. And the compute is the part driving the refresh in the first place. Serviceability saves you the radiator-and-solar fraction. It cannot save you the compute fraction, because that’s the fraction that keeps going obsolete.
Run the corrected numbers. Best orbit, service in place, replacing 13 t/MW of compute every cycle, at a hypothetical $100/kg: a 3-year refresh costs $14.5M/MW (above the ground build), a 6-year refresh $8.1M/MW (below it), and at $300/kg the 3-year case balloons to $43.5M/MW.
The red line is the crude case — relaunch the whole satellite each refresh — and it sits far above the ground line everywhere worth discussing. The green lines relaunch only the chips: better, but the 3-year cycle still doesn’t beat the ground until launch falls near $80/kg. Serviceability helps — it stops you relaunching the radiator and solar — but it cannot stop you relaunching the compute, because the compute is the thing going obsolete.
Which is where it gets interesting, because I’ve been smuggling in an assumption that’s wrong, and correcting it is the most important move in this whole exercise.
You don’t throw old chips away. You demote them.
I assumed the orbital data center must refresh on the same clock as the ground — swap chips every three years to stay current. But that’s not how compute retires, on the ground or anywhere. A two-generations-old GPU isn’t garbage; it’s worth less. It still computes, just at lower performance per watt. So hyperscalers cascade hardware: newest silicon takes frontier training, last-gen drops to inference, older still to cheap batch work. Chips retire by demotion, not deletion.
So the orbital operator doesn’t have to match the ground’s cadence at all. It can run its silicon for years past the frontier, keeping aging chips busy on lighter, latency-tolerant work, and relaunch far less often. That breaks the punishing symmetry — and it’s a real win. Stretch the refresh from three years to nine and the recurring launch term collapses.
But — and conservation laws always have a but — old chips don’t compute as well, and here’s the precise asymmetry, because it’s the place a sharp reader will push hardest. Both the ground and orbit run aging chips; demotion happens everywhere. The difference is the cost of the power and space that an old chip occupies. On the ground, a demoted chip draws cheap grid power and sits in an already-built rack — the marginal cost of keeping it is basically its electricity. In orbit, that same old chip occupies solar array and radiator and structure that cost $100–300 a kilogram to lift and cannot be repurposed for anything else — so its true cost is the amortized launch of that fixed power budget, whether the chip running on it is frontier or four years stale. It’s an opportunity-cost argument, not a “space is worse at computing” one: orbit’s power is expensive and fixed, so filling it with obsolete silicon wastes capacity you paid a fortune to put there.
Here I also have to switch units, and I want to be explicit about it, because the switch quietly does real work. Up to now "per MW" has meant a megawatt of electrical IT load. From here it means a megawatt of *frontier-equivalent* compute, because an aging chip still draws its full orbital power but delivers less useful work relative to the current best silicon. Performance per watt improves over time (doubling every D=3D = 3 D=3 years, generously); your orbital power budget is fixed, sized once at launch. So a chip three doublings old — nine years, in this toy model — delivers maybe an eighth of the frontier compute per watt, on infrastructure you're still paying full freight to keep aloft. And here is the asymmetry I'm granting the ground for free, stated plainly so it isn't hidden: the orbital cost gets divided by this aging output fraction, but the ground's $11.3M does not, because cart-down-the-aisle refresh keeps terrestrial silicon at the frontier continuously. That's a real thumb on the scale — and it favors the ground, which is the direction I've been rounding throughout. The whole point is that orbit *can't* do the cheap continuous swap, so it's the side that pays the obsolescence tax.
(You might ask why the power budget is fixed — why not just launch more solar as the chips improve? You can. But that’s more launch mass every cycle, which is exactly the cost the whole optimization is trading against. Fixed power per deployed satellite is the right model; growth means more satellites, more launches, and the math below already captures that.)
Put both effects together and the cost per unit of actual frontier-equivalent compute stops being monotonic. Refresh too fast and you relaunch chips constantly. Refresh too slow and you’re paying orbital prices to run silicon that can barely keep up. In between, there’s an optimum:
In a smooth continuous-cycle approximation, with performance per watt doubling every three years, the optimal refresh lands around nine years — and if you force integer launches over a 30-year life (deployments at years 0, 10, 20), the minimum nudges to roughly ten years with these inputs. The exact year isn’t the point, and it isn’t even sensitive to my assuming a smooth exponential decay: I reran it with the chip value falling linearly, and with a lumpy generation-by-generation step function, and an interior optimum near a decade shows up every time. That’s because the structure is general — launch cost falls as you refresh less often, while useful output also falls as the fleet ages, and any value curve that decays toward zero relative to the frontier produces that tradeoff. The precise optimum depends on the shape; its existence and its rough magnitude do not. And the point that matters is robust: the orbital optimum is much longer than the ground’s frontier-chip cadence — three times longer or more.
This is also the answer to the sharpest objection a cost analyst will raise: you optimized launch, but at $100/kg launch is the minority cost — satellite hardware has never been built for $100 a kilogram, so manufacturing dominates, and you optimized the wrong line item. True about the dollars, and beside the point about the structure. Changing the launch price scales the entire cost curve multiplicatively (it looks like a vertical shift only because the chart’s axis is logarithmic); it does not move the optimum. And the same is true of any per-kilogram cost of putting refresh mass on orbit — manufacturing, integration, insurance, all of it. Fold a hardware cost-per-kilogram into the launch term and the curve rises, but its minimum stays put. The refresh optimum and its dependence on the pace of chip improvement survive whatever the true all-in cost of a kilogram in orbit turns out to be. Only the absolute dollar figures move; the shape of the argument, and its crux, do not. The optimum is set by the tension between the durable orbital infrastructure you keep and the rate at which lifted silicon falls behind the frontier — not by what a kilogram costs.
At that optimum, orbital compute at $100/kg costs about $14M per megawatt of frontier-equivalent capacity — just above the $11.3M ground build on launch alone. But the ground also pays a grid-electricity bill, and orbit does not — it pays up front instead, in solar, batteries, radiators, and launch mass: thirty years of grid power adds $13–31M/MW to the terrestrial side, putting the ground build-plus-electricity comparator in the $24–42M range. So on launch cost alone, orbit at its optimal refresh comes in below the ground’s build-plus-electricity band — but that gap is what has to pay for everything I’ve left out of the orbital side: spacecraft manufacturing, servicing operations, insurance, stationkeeping, radiation-hardening, networking, and disposal. Below the band on launch alone is what a real business case would need as its starting margin, not its conclusion. At $300/kg, the optimized launch-only figure is about $42M/MW frontier-equivalent — landing right at the top of the ground build-plus-electricity band before a dollar of those orbital costs is counted. That’s a warning light, not a win. (Worth noting that the most-cited independent estimate lands in the same place: aerospace engineer Andrew McCalip’s calculator, written up in IEEE Spectrum, puts a gigawatt-scale orbital facility plus five years of operation north of $50 billion against about $16 billion on the ground — roughly 3×. Different method, same neighborhood as the arithmetic here, which is reassuring: when a back-of-the-envelope and a detailed model agree, both are probably close. Tellingly, McCalip gets even that far only by assuming the orbital fleet runs efficient, radiation-tolerant chips rather than frontier GPUs — which is the demotion argument arriving from the other direction.)
You might also note that you needn’t relaunch the whole compute payload each cycle — swap only the highest-value boards and keep the trays and interconnect — which helps the orbital case further, by an amount that depends on packaging choices nobody has made yet. I’ve left that on the table, conservatively, for Musk.
(And the sun fraction, the thing I was careful to pick favorably? Second-order. A generic 60%-sunlight orbit makes the satellite a third heavier and barely moves any of this. The refresh optimum dominates.)
Where the crank stops
Neither side had it right, and the truth is more interesting than either.
The critics are wrong: the heat can be rejected, the radiator is 3.3 tonnes, and “you can’t cool it in space” — though space is indeed a worse radiator — is not the argument that decides anything. They reached a reasonable instinct through the wrong variable.
Musk is right about the parts he’s loud about: heat is manageable, a single launch is competitive at projected Starship prices, and — once you account for the fact that you demote old chips rather than discard them — there’s an optimal way to operate an orbital fleet that becomes launch-and-electricity competitive with the ground in this toy model, at the optimistic end of his own launch-cost projections. He is not being stupid. Run the numbers his way, generously, and there’s a real case in there — one that still has to cover the spacecraft and servicing costs the model leaves out, but a real case.
But it lives in a specific corner, and the corner is defined by two numbers that are neither thermal nor propulsive. The first is launch price: the case looks interesting near $100/kg and becomes uncomfortable by $300/kg — where the launch-only figure already reaches the top of the ground build-plus-electricity band before orbital hardware and servicing costs are counted — and SpaceX has not yet flown anything at either price. The second is subtler and, I think, the actual crux: the pace of semiconductor improvement. The whole optimization turns on how fast orbital chips fall behind the frontier. If performance per watt keeps doubling slowly — every four or five years, as the leading edge arguably trends — old chips stay useful longer, the optimal refresh stretches out, and orbit looks good. If improvement reaccelerates to the historical two-year pace, the optimum tightens, the relaunch term comes roaring back, and the ground wins comfortably.
So an orbital data center is not a cooling bet. It’s barely a launch bet. It is, underneath everything, a bet that Moore’s law stays slow — that the chips you lift won’t be embarrassed too quickly by the ones you didn’t. That’s the number to watch. Not the temperature of space, and not even the price of a rocket. The rate at which silicon improves.
One honest boundary on all of this: it’s a launch-and-infrastructure model, not a full lifecycle one. I use two ground comparators — a capex-only baseline of $11.3M/MW, and that same figure plus a 30-year electricity band of $13–31M/MW — and the orbital side counts launch and durable hardware mass but not spacecraft manufacturing, servicing operations, insurance, or disposal. So the electricity band helps the orbital case; the omitted orbital hardware and operations hurt it; the two pull opposite ways. I’m also excluding the data-movement problem entirely: orbital compute only makes sense for workloads whose inputs can be staged ahead of time and whose outputs are compact or latency-tolerant, because you are not getting a fat low-latency fiber to low Earth orbit. And I’m setting aside radiation: leading-edge accelerators aren’t built on radiation-hardened processes, so a real design pays for shielding mass or accepts higher failure and faster degradation — either of which makes the orbital case harder, not easier, and tightens the very refresh clock the argument turns on.
Two of my “durable” components aren’t, and both cut the same way. Batteries cycling through Earth’s shadow see thousands of charge-discharge cycles a year; even space-grade cells are spent in five to ten years, not thirty, and solar arrays lose a percent or few of output annually to radiation. So in any eclipsing orbit the batteries and part of the array join the compute on the recurring-relaunch side of the ledger — they’re wear items, not failures. That only deepens the conclusion: more recurring mass per cycle makes the refresh cadence matter even more.
And the largest caveat is one my arithmetic simply can’t price. This model assumes the ground can buy electricity and a grid connection on demand at $0.05–0.10/kWh. The actual bull case for orbital compute is that it can’t — that transformer lead times, interconnect queues, and permitting mean the binding ground constraint isn’t the price of power but the availability of it, in which case scarcity rents push the ground comparator upward by an amount no one can yet quantify. That may well be the real argument for building in orbit. But notice it’s a bet on terrestrial grid bottlenecks, not on physics or launch cost, and it lives outside what a back-of-the-envelope can settle. I’m pricing power as available; if it isn’t, the orbital case looks better, for reasons this calculation doesn’t capture.
But the structure survives all of these corrections, because the refresh optimum — and the chip-progress bet behind it — governs the recurring side of the ledger no matter what. If someone tells you orbital data centers obviously do or don’t work, ask them two things: what launch price, and how fast they think accelerators will keep improving. Everything else is the temperature of space.
The calculation, every input cited
All figures are per megawatt — of electrical IT load through the mass and launch sections, and of frontier-equivalent compute in the demotion section (an aging chip draws the same orbital power but does less useful work). This is a launch-and-infrastructure model, not a full lifecycle one. I use two ground comparators: a capex-only baseline of $11.3M/MW, and that same figure plus a 30-year electricity band of $13–31M/MW. The orbital side counts launch plus durable hardware mass but excludes spacecraft manufacturing, servicing operations, insurance, failures, and disposal. Chip purchase cost is excluded, not cancelled: it would cancel only in a same-refresh comparison, but here ground and orbit refresh on different cadences, so treat the result as launch-and-infrastructure, not a full chip-capex model. The exclusion isn't neutral, and it leans the generous way: at the orbital optimum you deploy roughly 2.4 MW of fresh silicon every nine years (the oversizing needed to offset the aging fleet), about 0.26 MW of new chips per year, versus the ground's 1 MW every three years, about 0.33 — so the ground actually buys more new silicon per year, and leaving chip purchase out of both sides mildly favors the ground.
Stefan-Boltzmann constant σ — 5.67 × 10⁻⁸ W/m²K⁴ — physical constant.
Radiator emissivity ε — 0.90 — high-emissivity coating (generous).
Radiator temperature — 75 °C / 348 K — chip-junction-limited assumption.
Radiator areal density — 5 kg/m², two-sided — aggressive lightweight-radiator assumption; many flight systems are heavier.
→ Radiator mass — 3.3 t/MW — computed; radiating area ~1,300 m²/MW, physical panel ~670 m².
Compute mass (proxy) — 12.8 t/MW — NVIDIA DGX H100: 130.45 kg, 10.2 kW — terrestrial hardware, mass proxy only.
Solar specific power — 150 W/kg (best-in-class) — NASA Small Spacecraft SOA (typical ~30); UltraFlex up to 150 W/kg EOL.
Battery specific energy — 150 Wh/kg — NASA: commercial Li-ion 150–270 Wh/kg; space-grade trades energy for life.
LEO orbital period — 1.5 h — standard.
Structure / electronics — +30% — engineering allowance (my assumption).
→ Total satellite mass — ~30 t/MW (best orbit) to ~41 t/MW (generic) — computed.
Ground build cost — $11.3M/MW (shell + core) — JLL 2026 Data Center Outlook; AI fit-out can add up to $25M/MW.
Ground electricity (sensitivity band) — $13–31M/MW over 30 yr — computed at $0.05–0.10/kWh, PUE 1.0–1.2.
Falcon 9 launch — ~$2,700/kg to LEO — Orbital Radar, May 2026.
Starship (hypothetical, mature) — ~$100–300/kg projected — SpaceX publishes >100 t to orbit, no $/kg; $/kg are projections.
GPU architecture cadence — new generation every 18–24 months — Introl.
GPU economic life (contested) — 2–3 yr (Burry) vs 5–6 yr (hyperscaler books).
Building life — 15–30 yr — three-clock mismatch.
In-orbit servicing status — demonstrated in GEO (Northrop MEV undocking); not yet industrial (NASA cancelled OSAM-1).
The model. Sun fraction f sets two coupled masses: solar sizing scales as 1/f (oversize to charge through the dark side), battery mass as (1−f) (store energy for the eclipse). With m_refresh the compute mass relaunched each cycle and R the refresh interval in years over a 30-year life:
where L is launch dollars per kilogram, treated as delivered to the target LEO or sun-synchronous orbit, with inclination and orbit-insertion penalties ignored — another assumption that favors the orbital case. The crude case relaunches the whole satellite each cycle; the serviced case relaunches only the compute. The ground comparators are the $11.3M/MW build, optionally plus the $13–31M/MW electricity band. The serviced case is the relevant one, and it still carries the full compute mass on every refresh — which is why the refresh cadence, not the radiator, governs the result. (The effective-cost optimization in the demotion section divides this launched mass by the cycle-average frontier-equivalent output
which is what produces the ~9-year continuous, ~10-year integer optimum.)
For more conservation-law thinking — where a cost you thought you’d eliminated reappears somewhere else in the ledger — see The Conservation Law of Tax Alpha. My book, The Science of Free Will, asks an equally uncomfortable question about an equally cherished story.
Thank you! This is a high standard writing. This is hard to find in the age of AI slop. Your logic and numbers are sound without marketing spin, kudos.
Btw, we share something. I have a BS Physics from China, and studied BioPhysics Ph.D in a big10 school in mid 1990s. I know a few Physicists (incl one Indian who were on SSC and Weinberg. I worked in software for 2 decades.
I think the biggest threat to space data center, beyond chips cycle, is the architecture change of AI models. As I read, the current LLMs are hitting scaling wall, and only architecture breakthrough can address it. The crowded frontier models reflect that. Without that, it’s hard to say we need that many essentially the same models.