Orbital AI Data Centers — Economic Competitiveness Timeline > Pages > What is the achievable thermal rejection rate (W_rejected/kg) for radiative cooling systems in LEO, given AI compute operating temperatures?

Radiative Cooling Specific Power

Answer

The achievable thermal rejection rate for complete radiative cooling systems in LEO depends critically on three variables: radiator operating temperature, panel areal density (kg/m^2), and system overhead (cold plates, fluid loops, pumps, plumbing). The values below represent the complete thermal management system -- not radiator panels alone -- expressed as watts of heat rejected per kilogram of total thermal system mass.

Scenario W/kg Assumptions
Optimistic 250 85C radiator, 2 kg/m^2 advanced panels, 1.2x system overhead, terminator orbit, ~600 W/m^2 net
Central 130 80C radiator, 3 kg/m^2 next-gen deployable panels, 1.3x system overhead, ~500 W/m^2 net
Conservative 50 70C radiator, 5 kg/m^2 conventional panels, 1.5x system overhead, ~350 W/m^2 net with environmental loading

For context, the ISS EATCS achieves approximately 5-11 W/kg at the system level, but this is a poor benchmark because it operates at much lower temperatures (20C) with 1970s-era radiator technology. A purpose-built compute-cooling system at elevated temperatures would perform 10-50x better.

Critical distinction: Many cited figures (e.g., Dwarkesh Patel's "320 W/kg") refer to radiator panels only and exclude system overhead. The complete thermal chain (cold plates, fluid loops, pumps, manifolds, working fluid, and radiator panels) adds 20-50% to panel-only mass, which is the single most important adjustment when converting physics-limit calculations to engineering reality.

Evidence

Physics Fundamentals

E1. [evidence] Stefan-Boltzmann law: net radiative power per unit area scales as epsilon * sigma * (T_rad^4 - T_sink^4). At epsilon=0.9 radiating to 3K deep space, one-sided emission yields: 60C = 629 W/m^2, 70C = 708 W/m^2, 80C = 794 W/m^2, 85C = 840 W/m^2, 100C = 989 W/m^2, 127C = 1,308 W/m^2. Two-sided emission doubles these values. (First-principles calculation, verified against spacecomputer-cooling, peraspera-realities)

E2. [evidence] The SemiEngineering March 2026 analysis provides a practical rule of thumb: rejecting 1 kW of heat requires approximately 2.5 m^2 of radiator area. This implies ~400 W/m^2 effective -- consistent with a single-sided radiator at ~70-80C with some environmental derating. (spacecomputer-cooling)

E3. [evidence] A 1 m^2 surface at 80C radiates approximately 850 W per side; at 127C approximately 1,450 W/m^2. (spacecomputer-cooling)

E4. [evidence] The sun delivers ~1,361 W/m^2 (AM0) to sun-facing surfaces. On a sun-facing side, the radiator may become a "heat absorber" rather than a heat rejector, per Chinese aerospace publication Xianzao Ketang, February 2026. This makes radiator orientation and orbit selection critical. (spacecomputer-cooling)

ISS EATCS Benchmark

E5. [evidence] ISS External Active Thermal Control System rejects up to 70 kW using 422 m^2 of ammonia-loop radiators, achieving roughly 166 W/m^2 in practice. This is well below theoretical maximums due to solar exposure, Earth IR loading, and system inefficiencies. (spacecomputer-cooling, peraspera-realities)

E6. [evidence] ISS EATCS uses 6 radiator Orbital Replacement Units (ORUs), each spanning 23m x 11m with mass of ~1,100 kg per ORU. Total panel mass: ~6,600 kg for 1,500 m^2. Adding refrigerant, pumps, support beams, valve assemblies, rotary joints, and cold-side heat exchangers roughly doubles the mass to ~13,000 kg total system. This yields approximately 5.4 W/kg at the full system level. (hn-xai-spacex-thermodynamics)

E7. [evidence] Mach33 analysis: ISS radiator ORU has system-level areal density of approximately 13.9 kg/m^2 (1,082 kg per ORU with ~77.6 m^2 deployed planform). The ISS operates at ~20C, which is far below the temperatures relevant to GPU compute. (mach33-cooling-model)

E8. [evidence] ISS radiator performance is a poor benchmark for orbital data centers because: (a) it operates at 20C vs. 70-85C for GPUs, (b) uses 1970s/1990s-era panel technology at ~14 kg/m^2 vs. modern concepts at 2-5 kg/m^2, and (c) was not optimized for mass efficiency. Operating at 70C vs. 20C dramatically improves radiative efficiency due to T^4 scaling. (hn-xai-spacex-thermodynamics, hn-xai-spacex-radiators)

Operating Temperature Effects

E9. [evidence] Dwarkesh Patel's model: GPUs run up to 90C, with ~30C temperature drop through heat pipes and fluid loops to radiator surface, yielding ~60C radiator temperature. At 2 kg/m^2 aluminum panels with Stefan-Boltzmann, this gives roughly 320 W/kg for the radiator panels alone. (dwarkesh-space-gpus)

E10. [evidence] Modern datacenter GPUs (H100, B200 class) throttle at junction temperatures of 85-100C. With ~10C thermal gradient across cold plates and interface materials, the radiator surface temperature is constrained to approximately 75C. (mccalip-space-datacenters via McCalip bifacial panel model)

E11. [evidence] Mach33 analysis: with ~30K temperature drop from chip junction to radiator surface, a 60-100C radiator implies ~90-130C junction temperatures. Consumer GPUs operate at ~80C; modern datacenter GPUs tolerate ~90-105C before throttling; industrial/military silicon is specified up to ~125C. Space-optimized AI ASICs could credibly operate at ~110-125C, corresponding to ~80C radiator temperature. (mach33-cooling-model)

E12. [opinion] HN commenter: "Your calculations are based on cooling to 20C, which is exponentially harder than cooling to 70C where GPUs are happy. Radiators would be roughly 1/3 the size of the panels for 70C." This is directionally correct but overstated -- the improvement from 20C to 70C is ~63% in W/m^2 (from ~418 to ~708), not a 3x reduction in area. (hn-xai-spacex-thermodynamics)

E13. [evidence] SatNews: "Running radiators at 60C instead of 20C can reduce the required area by half, but it pushes the silicon to its thermal limits, requiring a delicate balance between hardware longevity and system mass." (satnews-physics-wall)

Panel Areal Density (kg/m^2)

E14. [evidence] Spacecraft radiator weight typically varies from almost nothing (if existing structural panel is used) to around 12 kg/m^2 for a heavy deployable radiator with support structure. NASA has specified 2 kg/m^2 as a target for advanced thermal management systems. Only bare carbon fiber radiators operating at 800-1000K have reached this at high temperature. (toughsf-radiators, isnps-lightweight-radiators)

E15. [evidence] NASA advanced radiator R&D reports ~2.1 kg/m^2 specific mass for two-sided heat rejection (or ~4.2 kg/m^2 for single-sided). Agency targets are in the ~2-3 kg/m^2 class for high-temperature radiators. (mach33-cooling-model, citing NASA sources)

E16. [evidence] State-of-the-art heat rejection radiators with Ti-water heat pipe panels range from 5.8 kg/m^2 (S4-CBC system) to 7.16 kg/m^2 (Prometheus JIMO mission). NASA TFAWS 2024: embedded branching network heat pipes achieve finned surface efficiencies >70% at 500K input heat, with areal density within the 2-3 kg/m^2 range using additive manufacturing. (isnps-lightweight-radiators via web search)

E17. [evidence] A novel deployable radiator system has been designed at 1.9 kg/m^2 (or 3.9 kg/m^2 planform area). NASA OHP-embedded radiator demonstrator achieved effective density 40% lighter than solid aluminum. Development goal: reduce average areal mass density from ~10 kg/m^2 current to ~2 kg/m^2. (web search results)

E18. [evidence] Mach33 analysis brackets radiator areal density across three classes: 2 kg/m^2 (lightweight concepts), 5 kg/m^2 (conventional), 10 kg/m^2 (conservative/armored). At 80C and 100 kW, this yields radiator mass of ~420 kg, ~1,010 kg, and ~2,000 kg respectively -- a 5x range for the same thermal job. (mach33-cooling-model)

Mach33 System-Level Analysis

E19. [evidence] Mach33 Starlink V3 scaling analysis (20 kW to 100 kW): at 100 kW, radiator planform area is ~99 m^2 (only ~7% of total spacecraft planform). Radiator mass rises to ~1.0 ton (at ~5 kg/m^2 conventional), representing ~18% of the ~5.4 ton total spacecraft mass. Solar arrays dominate both area (>90%) and mass growth. (mach33-cooling-model)

E20. [evidence] Mach33: at 100 kW waste heat and 80C radiator temperature, required radiator planform area falls from ~130 m^2 at 60C to ~77 m^2 at 100C. This is "the story investors should internalize: as radiator temperature rises, required area drops quickly." (mach33-cooling-model)

E21. [evidence] Mach33: at 20 kW, radiators account for ~200 kg (~10% of ~2-ton spacecraft). At 100 kW, radiator mass rises to ~1.0 ton at 5 kg/m^2 conventional panels. "Radiators grow from a secondary to a meaningful subsystem, but they do not dominate the mass budget." (mach33-cooling-model)

Advanced Concepts

E22. [evidence] Liquid Droplet Radiators (LDRs): generate controlled sheets of ~100 micron diameter droplets that radiate heat while traveling through vacuum, then recollect them. NASA research from the 1980s showed LDRs can be up to 7x lighter than conventional radiators. A November 2025 study in Applied Thermal Engineering demonstrated heat dissipation rates of up to 450 W/kg. LDRs remain developmental with ongoing research on droplet collection in microgravity, fluid contamination, and solar back-loading. (spacecomputer-cooling)

E23. [evidence] Sophia Space TILE architecture: flat, 1 m^2, few-centimeters-deep compute slabs where processors sit directly against a proprietary passive heat spreader. The high surface-area-to-volume ratio turns the entire structure into a radiator. Claims 92% of generated power goes directly to computation (only 8% thermal overhead). Plans to flight-test on Apex Space satellite bus by late 2027. (spacecomputer-cooling, satnews-physics-wall)

E24. [evidence] Google Project Suncatcher: constellations of ~81 TPU-equipped satellites at ~650 km altitude. Published research calls for "advanced thermal interface materials and heat transport mechanisms, preferably passive to maximize reliability." (spacecomputer-cooling)

E25. [evidence] SatNews: by 2027, industry expected to move toward active thermal control including space-rated heat pumps that can "boost" radiator temperatures to increase dissipation efficiency. (satnews-physics-wall)

System Overhead

E26. [evidence] The complete thermal chain for space compute follows the path: CHIP -> COLD PLATE -> FLUID LOOP -> RADIATOR PANELS -> DEEP SPACE. This is standard spacecraft thermal architecture. (balerion-space-bsv)

E27. [evidence] SatNews: "Unlike solar panels, which can be thin-film and flexible, high-performance radiators often require internal plumbing for liquid heat pipes or loop heat pipes (LHPs) to move thermal energy from the high-density chips to the external fins." (satnews-physics-wall)

E28. [evidence] Per Aspera: for a data center, a robust internal liquid cooling loop carrying heat from CPUs/GPUs out to radiators requires pumps, fluid, plumbing -- all of which add mass and potential points of failure. "Scaling up means essentially building the equivalent of a building's HVAC system into a satellite." (peraspera-realities)

E29. [evidence] Per Aspera: 100 kW system requires radiators of ~1,000+ kg. At the system level, this implies ~100 W/kg or less. (peraspera-realities)

E30. [evidence] HN discussion: ISS EATCS radiator panels alone are ~6,500 kg for 70 kW. Adding all the refrigerant, pumps, support beams, valve assemblies, rotary joints, and cold-side heat exchangers "probably together double the mass you need to put in orbit." (hn-xai-spacex-thermodynamics)

Scale Effects

E31. [evidence] SatNews: a centralized 1 GW orbital data center would require roughly 834,000 m^2 of radiators to dissipate waste heat (assuming 40% efficiency and radiators at 400K/127C). Radiators alone would weigh approximately 2,250 tonnes. (satnews-physics-wall)

E32. [evidence] SpaceComputer/Tara Jean: scaling ISS approach to 1 GW would require approximately 3,950 m^2 of radiator at optimistic operating temperatures, with mass of 19,750 to 39,500 kg at 5-10 kg/m^2. One first-principles analysis found thermal management system mass alone exceeds combined mass of computing equipment, power systems, and structural components at this scale. (spacecomputer-cooling)

E33. [evidence] SpaceComputer/Tara Jean cooling maturity table: at 1-100 kW (pumped fluid loops, panel radiators) the technology is flight-proven, constrained by fluid loop reliability and radiator orientation. At 100 kW-10 MW (distributed radiators), partially proven, constrained by radiator mass fraction. At 10 MW-1 GW+ (LDR, massive arrays), developmental, with radiator becoming dominant system mass. (spacecomputer-cooling)

Distributed vs. Centralized Architecture

E34. [opinion] Mach33: "The radiator question is often treated like a veto. Our radiator model, anchored on Starlink V3, suggests it should be treated as a line item." At 100 kW per satellite, radiators remain ~18% of total mass. (mach33-cooling-model)

E35. [opinion] HN commenter: "Datacenter capacity (and thus heat) grows by the cube law, but the ability to radiate heat grows by the square law, so it seems like it would be advantageous to have a bunch of smaller satellites." (hn-xai-spacex-radiators)

E36. [evidence] SpaceComputer: at individual satellite level (100 kW class), cooling is tractable. At constellation-aggregated levels (100 MW to GW), radiator mass compounds across thousands of spacecraft "but not to the point of becoming a constraint in contrast to area budget." (spacecomputer-cooling)

Analysis

Deriving the Three Scenarios

The specific power (W/kg) of a complete thermal management system depends on a chain of factors:

  1. Radiator temperature sets W/m^2 via Stefan-Boltzmann (E1)
  2. Environmental derating reduces ideal W/m^2 by 20-40% in LEO due to Earth IR (~237 W/m^2), albedo, and partial solar exposure (E4, E5)
  3. Panel areal density converts W/m^2 to W/kg for radiator panels alone (E14-E18)
  4. System overhead adds 20-50% mass for cold plates, fluid loops, pumps, manifolds, and working fluid (E26-E30)

Conservative (50 W/kg)

This represents near-term, conventional technology with conservative assumptions:

This is consistent with the Per Aspera estimate of ~100 W/kg for panels only at 100 kW (E29), adjusted for the complete system. It is a large improvement over ISS (5-11 W/kg) because it operates at a much higher temperature and uses modern panel technology.

Central (130 W/kg)

This represents next-generation technology achievable within a 5-year development horizon:

This is consistent with Mach33's analysis showing ~1,000 kg of radiator system for 100 kW at 5 kg/m^2 (E21), but assumes lighter panels. Their analysis at 80C and 2 kg/m^2 gives ~420 kg radiator for 100 kW (E18), which is ~238 W/kg panels-only.

Optimistic (250 W/kg)

This represents aggressive but physically plausible technology:

This is broadly consistent with Dwarkesh Patel's 320 W/kg (E9), which is panel-only at 60C with 2 kg/m^2. Our optimistic figure is lower despite higher temperature because we include system overhead. The LDR demonstration of 450 W/kg (E22) suggests the physics allows even higher figures, but LDRs are TRL 3-4 and not assumed for near-term deployment.

Key Uncertainties and Sensitivities

Temperature is the dominant lever. Due to T^4 scaling, the difference between a 60C and 100C radiator is nearly 60% more W/m^2 (629 vs. 989 per side at epsilon=0.9). The Mach33 analysis (E20) shows this translates to 130 m^2 vs. 77 m^2 at 100 kW -- a 40% area reduction. The challenge is GPU reliability at elevated junction temperatures.

Areal density (kg/m^2) is the second lever. For the same thermal job at 80C and 100 kW, Mach33 shows radiator mass spanning 420 kg to 2,000 kg depending on whether panels are 2 or 10 kg/m^2 (E18). This 5x range makes panel technology selection a critical economic variable.

System overhead is often omitted. Most W/kg figures cited in the discourse (Dwarkesh 320 W/kg, E9) are for radiator panels only. The ISS data (E6, E30) suggests system overhead can double panel mass. For a purpose-built compute-cooling system, overhead of 1.2-1.5x is realistic.

LEO environment derates ideal performance. Even in a terminator orbit, Earth IR loading (~237 W/m^2) and occasional albedo exposure reduce net cooling below the ideal deep-space calculation. Practical derating of 20-40% from ideal is standard (E5).

Distributed architecture reduces the per-satellite thermal challenge. At 100 kW per satellite (Starlink V3 scale), radiators are ~18% of mass -- a line item, not a blocker (E19, E34). The thermal problem becomes more severe at centralized MW+ scale (E32-E33).

Technology Readiness

Approach TRL W/kg (system) Notes
ISS EATCS heritage 9 5-11 Low temp, heavy panels, flight-proven
Conventional deployable panels + MPFL 7-8 40-80 Shenzhou, Starlink-class heritage
Advanced lightweight panels (3 kg/m^2) + MPFL 5-6 100-170 NASA R&D, AM heat pipes demonstrated
Ultra-lightweight panels (2 kg/m^2) 4-5 150-300 NASA target, limited demonstrations
Sophia Space TILE (integrated radiator) 4 150-250 Flight test planned 2027
Liquid Droplet Radiators 3-4 300-450+ Lab demonstration only, no flight heritage

Verdict

The conservative estimate (50 W/kg) is achievable with existing flight-heritage technology operated at higher temperatures. The central estimate (130 W/kg) requires next-generation panels at NASA target areal densities (3 kg/m^2) and optimized system integration, which is consistent with what Mach33 and SpaceComputer project for late-2020s hardware. The optimistic estimate (250 W/kg) requires either 2 kg/m^2 panels (at NASA R&D frontier) or integrated architecture approaches like Sophia Space's TILE -- achievable in physics, but not yet demonstrated in flight at scale.

The 450 W/kg LDR figure (E22) is excluded from the scenario range because it represents a technology at TRL 3-4 with no flight heritage. If LDRs mature to flight readiness by the early 2030s, the optimistic scenario could be revised upward to 350-450 W/kg.