Datacenters in space aren't going to work

Role of sci‑fi and hype

• Many see “datacenters in space” as shallow sci‑fi cargo culting: latching onto space aesthetics while ignoring the cautionary, societal focus of real speculative fiction.
• Several comments frame the idea as investor/PR narrative rather than serious engineering: something to reassure AI/infra investors and distract from terrestrial siting, regulation, and NIMBY issues.

Thermal management and “vacuum cooling”

• Core consensus: cooling is vastly harder in space. No air or water means essentially no convection; only radiation to deep space is available.
• Vacuum is an excellent insulator (thermos analogy). To dump multi‑MW of heat, you need gigantic radiators—football‑field to square‑kilometer scale for modern DC loads.
• Moving heat from chips to those radiators requires complex multi‑stage liquid loops and pumps; any leak or failure is catastrophic and hard to service.
• A minority argue that with very hot radiators, better coatings, and huge structures, it’s “just engineering,” but even they concede it’s difficult and expensive.

Radiation and electronics reliability

• Space datacenters would face high rates of single‑event upsets even in LEO, aggravated in regions like the South Atlantic Anomaly.
• True rad‑hard CPUs/GPUs exist but are generations behind and extremely expensive; triple‑modular redundancy further slashes effective performance.
• Some note ML inference is numerically tolerant to bitflips, but for large, precise workloads the reliability penalty is severe.

Economics, scale, and maintenance

• Launch costs, station‑keeping, gigantic radiators, shielding, and ground stations make per‑MW cost orders of magnitude above terrestrial DCs, even assuming Starship‑level prices.
• GPU lifetimes (~5 years) clash with “launch once, leave it there” dreams; maintenance missions are prohibitively expensive, and fail‑in‑place designs waste enormous capital.
• Comparisons to Microsoft’s underwater project: cooling “worked,” but logistics and maintenance killed scalability; space would inherit those problems plus worse cooling and radiation.

Latency, bandwidth, and realistic use cases

• Space links are tiny compared to intra‑DC fiber; Starlink‑class bandwidth/latency is hopeless for large training clusters that depend on ultra‑fast interconnects.
• More plausible niche: processing space‑originating data in orbit (imaging, surveillance, autonomous spacecraft), where local compute reduces downlink needs.

Alternative locations (ocean, poles, Moon, asteroids)

• Underwater, Arctic/Antarctic, rural, and bunker DCs are repeatedly cited as far more practical ways to get cheap cooling, isolation, or security.
• Moon/asteroid concepts face similar radiation and worse thermal issues; lunar regolith is an insulator, not an effective heatsink.

Security, jurisdiction, and dual use

• Some speculate about evading nation‑states or enabling resilient crypto/“sovereign” infra in orbit; others point out space assets are traceable, treaty‑bound, and trivially targetable by ASAT weapons.
• More credible “dual use” story: on‑orbit compute for military sensing, tracking, and battle‑management—though that still doesn’t justify general AI datacenters in orbit.

Environmental and solar‑power arguments

• Space solar gets more consistent, stronger insolation, but critics stress you still must radiate the same energy away; the thermal problem dominates.
• Climate impact of frequent launches is flagged as unclear but potentially serious; relying on rockets to “green” AI compute is viewed skeptically.

Optimism vs. “fundamentally dumb”

• A small camp argues “hard ≠ impossible” and that billionaires funding R&D can advance space thermal tech and on‑orbit compute for other missions.
• The dominant view: this isn’t merely difficult, it’s structurally worse than ground DCs on every important axis—cooling, cost, bandwidth, maintenance, and legal risk—so the idea is, for now, fundamentally uneconomic and mostly marketing.