Why we should train AI in space - White Paper

Below top-k · 37

• contextReusable heavy-lift launchers are about to make launch costs negligiblec 0.90

  Partially and fully reusable heavy-lift vehicles are expected to reach roughly $30/kg to LEO long term, possibly as low as $10/kg, meaning launch is no longer the dominant cost driver for an orbital data center.

• mechanism24/7 unattenuated solar drives radically lower energy costc 0.85

  In orbit, solar arrays receive high-intensity sunlight without day/night cycles, weather, or atmospheric losses, yielding orders-of-magnitude lower marginal energy cost than terrestrial generation.

• evidenceA 5 GW training cluster would exceed the largest US power plantc 0.80

  Training models on the scale of Llama 5 or GPT-6 is projected to need ~5 GW clusters from 2027 onward — exceeding the output of the largest US power plant and rendering such clusters infeasible on today's grid.

• claimWaste heat must be radiated to deep space at gigawatt scalec 0.80

  With no conduction or convection available, all dissipated power must be radiated, requiring deployable radiators far larger than anything previously flown, aimed at the 2.7 K background.

• implicationA 5 GW data center needs fewer than 100 launchesc 0.80

  With ~40 MW of compute per launch and a comparable number of solar/radiator launches, a 5 GW orbital cluster is reachable in under 100 flights — conceivably deployable in 2-3 months at planned launch cadences.

• mechanismContinuous illumination roughly doubles power and removes batteriesc 0.80

  A non-eclipsing orbit nearly doubles average solar power versus day/night orbits, eliminates thermal cycling fatigue on panels, and removes the need for significant battery storage to keep compute running.

• mechanismPassive radiative cooling exploits the cold of spacec 0.75

  Orbital data centers can reject heat via passive radiative cooling directly to space, achieving low coolant temperatures and a more efficient cooling architecture than terrestrial chillers operating against hot ambient air.

• evidenceA 5 GW orbital data center needs roughly a 4 km by 4 km solar arrayc 0.75

  At 22% silicon cell efficiency and 90% fill, powering 5 GW requires about 16 km² of array, smaller and cheaper than the equivalent terrestrial farm thanks to higher capacity factor in space.

• evidenceOne 100-ton launch can deploy roughly 40 MW of computec 0.75

  A heavy-lift payload bay can fit about 300 racks at 50% capacity; at 120 kW/rack (GB200 NVL72 density) that is around 40 MW per launch, and rising rack power densities make this a conservative estimate.

• evidenceTerrestrial solar capacity factor is capped well below spacec 0.70

  US solar farms achieve a median 24% capacity factor and northern European projects under 10%, with >50% physically impossible on Earth due to the day/night cycle alone — while a space array can far exceed this.

• mechanismModular 3D assembly enables tightly coupled orbital clustersc 0.70

  Compute, power, cooling, and networking modules can be assembled in orbit in a 3D rather than 2D layout, keeping the cluster tightly coupled with low internal latency — a critical property for AI training.

• caveatOrbital debris and collision avoidance are the main regulatory hurdlec 0.70

  Because of their size, orbital data centers must demonstrate low collision probability via maneuverability, state-of-the-art tracking, ephemeris registration, and coordination with regulators.

• claimDesign is governed by five principles aimed at low cost and future-proofingc 0.70

  The concept design prioritizes modularity, maintainability, minimizing moving parts, resiliency through graceful degradation, and incremental scalability profitable from the first container.

• mechanismThin-film silicon cells enable mass- and volume-efficient launchc 0.70

  Cells under 25 μm thick deliver >1000 W/kg, can be folded or rolled for launch using proven Z-fold, roll-out, or picture-frame deployment, and anneal radiation damage without cover glass.

• implicationSpeed and agility, not just cost, are the real prizec 0.70

  Beyond saving on permitting costs, escaping terrestrial constraints lets operators scale faster and pivot plans as commercial requirements change, an agility terrestrial buildouts cannot match.

• claimA workable thermal design is possible without heat pumpsc 0.70

  Even with purely passive radiators at modest temperatures, the math closes for rejecting waste heat in this orbit, so active refrigeration is an optimization rather than a requirement.

• mechanismLarger containers shrink shielding mass per unit computec 0.70

  Shielding mass grows with surface area while compute grows with volume, so consolidating into very large containers reduces shielding mass per FLOP — making radiation a smaller relative concern than on typical satellites.

• contextAI training imposes tight latency and bisection bandwidth requirementsc 0.65

  Training large models forces all containers to sit within a few hundred meters of each other in a daisy-chained network with enough bisection bandwidth to support whole-cluster training jobs.

• mechanismModular swap-out keeps the cluster current despite component agingc 0.65

  Compute containers are designed to be docked and undocked individually, so faulty or outdated units can be returned in a launcher payload bay or designed to fully burn up on re-entry, with system-level redundancy handling graceful degradation.

• contextTiming aligns with reusable heavy-lift and in-orbit networkingc 0.60

  The opportunity is timed to the arrival of cost-effective reusable heavy-lift launchers and the proliferation of in-orbit networking, which together make orbital compute newly tractable.

• mechanismContainers dock via a single universal port for power, network, and coolingc 0.60

  Each compute container connects to the main structure through one mechanical port carrying thousands of fiber pairs, HV power, and high-volume coolant, mirroring terrestrial container data centers.

• mechanismTwo-phase cooling loops move heat from compute cores to radiatorsc 0.60

  Direct-to-chip liquid or two-phase immersion cooling inside containers feeds external cooling loops, with two-phase systems reducing mass flow and pumping losses; immersion can double as radiation shielding.

• claimModular incremental buildout is the realistic deployment pathc 0.60

  Rather than a rapid burst, the data center will be built up gradually by docking standardized compute containers and solar/radiator modules around a central hub, using only two primary structure types for manufacturing economies.

• implicationEnvironmental footprint shrinks: no grid emissions, no cooling waterc 0.55

  A European Commission study cited in the paper concludes orbital data centers significantly reduce greenhouse gas emissions from grid electricity and eliminate fresh water usage for cooling.

• examplexAI resorted to natural gas generators because the grid wasn't readyc 0.55

  xAI's Memphis cluster had to run on MW-scale natural gas generators temporarily because the local grid couldn't supply enough power, illustrating the severity of terrestrial bottlenecks.

• mechanismOptical communications dodge radio spectrum regulationc 0.55

  Laser links are currently unregulated and offer higher throughput and security than RF, making them well-suited to the high data rates orbital data centers require.

• mechanismData shuttles can ferry petabytes from ground to orbitc 0.55

  To complement laser and RF links, small docking modules launched from Earth can deliver petabyte- to exabyte-scale training data per trip, an approach already demonstrated by sending 7 GB to the ISS via an Amazon Snowcone.

• mechanismHeat pumps can dramatically boost radiator output via the T⁴ termc 0.55

  Because radiated power scales as the fourth power of temperature, using heat pumps to raise radiator temperature multiplies output per square meter — at the cost of additional electrical power to run the pumps.

• caveatSensitive electronics still need radiation shieldingc 0.55

  Even in low Earth orbit, storage and power-delivery components are vulnerable to latch-up, transients, and total ionizing dose effects, though logic devices — including those used in AI training — are notably more resilient.

• caveatTrue lifetime is set by cooling and power infrastructure, not chipsc 0.55

  The replaceable compute is not the limiting factor; the underlying cooling loops and power delivery subsystems — like those on the ISS — are expected to drive a roughly 15-year overall lifetime for the facility.

• evidenceProject Natick showed sealed environments can extend hardware lifec 0.50

  Microsoft's underwater data center demonstrated that sealed, thermally stable, non-corrosive environments can actually prolong electronics lifespan — a precedent for the benefits an orbital environment may offer.

• caveatSpace adds radiation shielding as a new cost linec 0.45

  Orbital operation introduces a radiation shielding cost — roughly 1 kg per kW of compute at $30/kg launch cost — that has no terrestrial analog, though it remains small relative to energy savings.

• evidenceISS data suggests solar arrays tolerate small debris impacts wellc 0.45

  Most of an orbital data center's surface is solar array, and ISS experience shows small debris collisions with arrays are largely passive over time.

• implicationVacuum gives a 35% speed-of-light advantage over fiberc 0.40

  Because light travels 35% faster in vacuum than in glass fiber, orbital interconnects offer a latency advantage that could be exploited in cluster networking.

• exampleTerrestrial cooling overprovisions for >45°C summer daysc 0.35

  Earth-based data centers must overprovision cooling capacity for peak summer days sometimes exceeding 45°C, an inefficiency that vanishes in the stable thermal environment of space.

• contextStarcloud positioning as first moverc 0.30

  Starcloud frames itself as the first company pursuing orbital data centers at gigawatt scale, with its own range of concept designs underpinning the white paper's claims.

• caveatAstronomy impact is limited by orbit choicec 0.30

  In the selected orbit the data centers are visible only at dawn and dusk when ambient light already precludes most astronomy, so impact on ground-based observation should be minimal.

Redundant with selected · 4

• claimOrbital data centers sidestep almost all Earth-side infrastructure bottlenecksc 0.95 · sim 0.88

  Putting compute in orbit avoids the permitting, transmission, and grid constraints that throttle terrestrial deployments, enabling faster and more agile scaling.

  overlapped with: Orbital data centers are a viable response to AI's power scaling crisis

• claimOrbital data centers sit at the intersection of four converging trendsc 0.95 · sim 0.86

  The case rests on cheap reusable launch, an imminent terrestrial electricity demand crunch, ballooning demand for large GPU clusters, and mega-constellation connectivity all arriving at once — making orbital compute feasible, economical, and necessary.

  overlapped with: Orbital data centers are a viable response to AI's power scaling crisis

• claimOrbit removes the physical and permitting ceiling on cluster scalec 0.85 · sim 0.87

  Space-based clusters can be linearly scaled nearly indefinitely, free of the siting, grid, water, and planning constraints that bottleneck terrestrial gigawatt-scale projects.

  overlapped with: Orbital data centers are a viable response to AI's power scaling crisis

• evidenceWater usage on land is non-trivial: 1.7M tons per clusterc 0.40 · sim 0.86

  A terrestrial 40 MW cluster over 10 years consumes roughly 1.7 million tons of water for cooling at 0.5 L/kWh — a cost item that simply disappears in orbit.

  overlapped with: 10-year cost comparison: $167M on land vs $8.2M in space