Satellite GPU Capacity Scaling: How Many GPUs Per Satellite?

Our main analysis parameterizes cost per kW_IT without specifying GPU count, but the industry consensus is converging on ~100 kW satellites (~72 GPUs). The inference networking requirements page identifies that the most demanding workloads — frontier MoE models requiring wide expert parallelism across 64+ GPUs in a single NVLink domain — cannot be served across satellites with current inter-satellite link technology. Could a single satellite house an entire NVL72 rack (72 GPUs, ~120-130 kW) or larger? What are the tradeoffs between satellite size, thermal management, structural complexity, reliability, and the alternative of distributing GPUs across multiple smaller satellites connected by optical links?

Answer

A single satellite housing 72 GPUs (~130 kW_IT) is physically feasible and represents the baseline design point for multiple industry proposals, including SpaceX's AI Sat Mini (~100 kW, ~1 ton) and Starcloud-3 (~100 kW, ~2 tons). Independent analyses bracket the total satellite mass at 1-5.4 metric tons for 100 kW_IT -- well within Starship's 100+ ton LEO capacity. Volume is not a constraint: the full terrestrial NVL72 rack occupies only ~1.4 m³, and even at the compute-hardware-mass page's central estimate of 6.0 kg/kW_IT (~780 kg for a ~130 kW satellite), the compute hardware is far more compact than the support systems it requires.

However, thermal management becomes a major design driver above ~100 kW on a single satellite. Rejecting 137 kW of heat at 80°C requires ~275 m² of radiator area (vs. the ISS's 460 m² for only 70 kW at lower temperature). Thermal transport distances beyond ~10 m require mechanically pumped fluid loops with no direct flight heritage at this scale. Running chips hotter (SpaceX's D3 approach) or using heat pump temperature boosting can reduce radiator area by 30-60%, but these remain undemonstrated in orbit.

The practical single-satellite power ceiling is ~300-500 kW with near-term technology, limited by deployable structure mechanics, ground testing facilities, and attitude control challenges. A 1 MW satellite (12-50 tons depending on technology assumptions) fits in a single Starship launch but requires solar arrays exceeding 3,000 m² — beyond any single-deployment system.

The industry is splitting into two architectural camps: monolithic high-power satellites (SpaceX, Starcloud) betting on Starship economics and custom hot-running chips, and distributed formation flying (Google Suncatcher's 81-satellite clusters with 1.6 Tbps inter-satellite links) betting on the thermal and reliability advantages of smaller satellites. Both are viable, and the right choice depends on workload mix — monolithic satellites better serve tightly-coupled inference, while distributed clusters excel at embarrassingly parallel workloads and have ~3x lower standard deviation in catastrophic loss outcomes.

In-orbit assembly is not needed for near-term orbital compute. Starship can launch monolithic satellites well beyond current needs (up to ~2-5 MW). Assembly becomes relevant only for multi-MW individual platforms in the 2035+ timeframe.

Analysis

Physical Feasibility of Rack-Scale (72+ GPU) Satellites

The compute hardware itself is not the constraint. The Dwarkesh/Patel estimate of a stripped GB200 NVL72 at ~100 kg [nvl72-rack-physical-specs.1, spacex-ai-sat-mini-spacenews.3] is an unverified lower bound that likely understates bare-board mass — the compute-hardware-mass page finds it inconsistent with HGX B200 baseboard scaling (which implies ~288 kg for equivalent GPU baseboards alone) and adopts a central estimate of 6.0 kg/kW_IT, or ~780 kg for a ~130 kW satellite. Even at this higher figure, compute hardware contributes only 4.0-9.0 kg/kW_IT, or ~15-25% of total satellite mass — the support systems, not the compute, dominate the mass budget.

The dominant mass and area drivers are the support systems: solar arrays (~30-50% of mass) and thermal radiators (~15-20% of mass), with structural overhead adding 15-40% multiplicatively. This means the question "how many GPUs per satellite" reduces to "how much power and cooling can a single satellite platform support?"

Starship's payload capacity removes launch vehicle constraints entirely. At 100+ metric tons to LEO in a 9 m × 18 m fairing (~1,100 m³), even the most conservative 100 kW satellite estimate (5.4 tons) uses only 5% of capacity. Starship could deploy ~20 such satellites per launch, or a single satellite up to ~2-5 MW before exceeding structural volume limits [spacex-ai-sat-mini-spacenews.3, starship-payload-specs.1].

Mass Budget: The 1-5.4 Ton Range for 100 kW

Independent mass estimates for a 100 kW orbital compute satellite converge within a 5x range:

The wide range reflects fundamentally different technology assumptions. SpaceX's aggressive ~1 ton figure likely requires next-generation solar arrays (200+ W/kg vs. flight-proven 100-120 W/kg), a custom chip designed to operate at elevated temperatures (reducing radiator mass), and minimal batteries in a dawn-dusk SSO. The Mach33 figure conservatively scales existing Starlink V3 mass ratios and is the most defensible near-term estimate [spacex-ai-sat-mini-spacenews.3, peraspera-realities.6, mach33-cooling.1].

At the 100 kW scale, the Mach33 breakdown shows solar arrays dominating at ~48% of total mass (~2,600 kg), with radiators at ~18% (~1,000 kg), and the bus/structure at ~34% (~1,800 kg). Radiators are "a line item, not a veto" at this power level mach33-cooling.1. This aligns with Elon Musk's observation that "the solar array is most of the weight on the satellite" musk-2026.3.

Thermal Scaling: The Binding Constraint Above 100 kW

Thermal management is the most challenging engineering dimension for concentrated satellite designs, and the one most sensitive to satellite size.

Radiator area scales linearly with thermal load at a given temperature — there are no economies of scale. Theoretical Stefan-Boltzmann emission at 80°C is ~850 W/m², but net rejection in LEO is substantially lower (~500 W/m²) due to Earth albedo heating (~30% of solar constant reflected), Earth IR emission (~240 W/m²), and non-ideal emissivity — a ~40% derating consistent with ISS experience (theoretical 418 W/m² at 20°C vs. measured 166 W/m²) [spacecomputer-cooling.1, nasa-atcs-overview.1]. At 80°C net ~500 W/m², a 137 kW satellite needs ~275 m² of radiator. At 70°C (~350 W/m² net), this grows to ~391 m².

For reference, the ISS External Active Thermal Control System — the largest operational heat rejection system in space — rejects only 70 kW across ~420-460 m² at low temperatures, with a system mass of ~13,000 kg [nasa-atcs-overview.1, spacecomputer-cooling.1]. A 130 kW compute satellite needs to reject roughly 2x the ISS's heat load but can operate at much higher temperatures (70-85°C vs. ISS operating temperatures) with far lighter radiator panels (2-5 kg/m² vs. ~28 kg/m² for the full ISS EATCS system including pumps, or ~5-8 kg/m² for panels alone), fundamentally changing the mass arithmetic.

Temperature is the single strongest design lever. The T⁴ Stefan-Boltzmann scaling means:

• 70°C → ~350 W/m² → 391 m² for 137 kW
• 80°C → ~500 W/m² → 274 m²
• 100°C → ~700 W/m² → 196 m²
• 127°C → ~1,000 W/m² → 137 m²

SpaceX's D3 chip, designed to "run hotter" than terrestrial GPUs, directly exploits this relationship spacex-ai-sat-mini-spacenews.4. ESA-funded heat pump technology (Celeroton) can boost radiator temperature from 80°C to 150°C, reducing required area by ~60% at the cost of 5-10% compute power celeroton-space-thermal.1.

Thermal transport distance is an underappreciated constraint. Conventional spacecraft heat pipes (CCHPs) transport heat passively up to several meters — practical designs reach ~4-5 m act-cchps-space.1; loop heat pipes (LHPs) extend transport distances further, with spacecraft LHP transport lines exceeding 5 m act-cchps-space.1. Beyond that, mechanically pumped fluid loops (MPFLs) are required, adding mass, power draw (increasing PUE), and leak-based failure modes — the ISS has experienced multiple ammonia leaks in its pumped loops spacecomputer-cooling.2. A 130 kW satellite with 275+ m² of deployed radiators requires active thermal transport across 10-20+ meters. The ISS uses MPFLs successfully, but no autonomous satellite has operated MPFLs at this scale.

Distributed architectures have a fundamental thermal advantage. Smaller satellites have better surface-area-to-volume ratios for heat rejection (square-cube law). Splitting 130 kW across 4 satellites at ~34 kW each brings each satellite within flight-proven thermal technology: ~68 m² radiator area (comparable to a single ISS radiator ORU), 5-8 m transport distances (within loop heat pipe range), and conventional deployable structures. At 8 satellites (~17 kW each), the thermal problem simplifies to heat-pipe-class transport with body-mounted plus small deployable radiators — well within heritage.

Solar Array Scaling

Solar arrays dominate satellite planform area and mass but are not the binding constraint — the technology exists or is in near-term development for the 100-300 kW range.

A 130 kW satellite (at PUE 1.05) needs ~137 kW of generation. At practical efficiencies:

• 200 W/m² (30% efficient cells, packing losses): ~685 m² array area, ~685-1,370 kg (at 100-200 W/kg)
• 300 W/m² (optimistic): ~457 m² array area, ~457-913 kg

For the SpaceX AI Sat Mini at 100 kW, the ~180 m wingspan exceeds the ISS (108.5 m) but at a fraction of the mass (~1 ton vs. 420,000 kg) spacex-ai-sat-mini-daniel-marin.2. This relies on strain-energy deployable boom technology that is advancing rapidly.

The solar technology progression supports these power levels:

• ROSA (flight-proven, ISS iROSA): ~100 W/kg nasa-smallsat-power-soa.1. NASA's Gateway PPE will deploy 60 kW from 2 ROSA wings nasa-rosa-gateway.1.
• MegaFlex (ground-tested, TRL 5-6): up to 200 W/kg, 175 kW per wing — a single wing nearly covers a 130 kW satellite's needs megaflex-sbir.1.
• NASA SEP studies have designed and validated 300 kW array structures nasa-300kw-solar-array-structures.1.

The practical single-satellite ceiling appears to be ~300-500 kW, where a two-wing MegaFlex system could provide up to 400 kW. Beyond this, attitude control (solar radiation pressure and drag on large areas), ground testing facilities (largest thermal vacuum chamber: ~30 m), and deployment mechanics push toward modular or multi-satellite architectures. A 1 MW satellite would need ~3,300+ m² of arrays — exceeding the entire ISS solar array area — with no single-deployment system in existence or development.

The Architecture Decision: Monolithic vs Distributed vs Formation Flying

The evidence reveals three architectural paths with different maturity timelines and workload suitability:

Monolithic high-power satellites (SpaceX AI Sat Mini, Starcloud-3, K2 Space Giga-Class):

• Single satellite, 100+ kW, fully integrated and ground-tested
• Internal NVLink interconnect (TB/s) — best for tightly-coupled inference
• Simpler operations: no formation keeping, no inter-satellite networking
• But: thermal transport challenge at scale, large deployable structures, higher per-event loss risk
• SpaceX explicitly plans megawatt-class follow-on satellites spacex-ai-sat-mini-spacenews.3

Formation flying (Google Suncatcher, Kepler):

• Multiple smaller satellites at 100-200 m spacing within a ~1 km cluster
• Google demonstrated 1.6 Tbps bidirectional optical links in bench tests, targeting tens of Tbps via DWDM google-suncatcher-research.1
• At 100-200 m, propagation latency is ~0.3-0.7 μs (negligible) and terrestrial-style COTS transceivers can be used directly
• Per-satellite thermal challenge stays within flight-proven technology
• No structural connections — avoids all assembly and deployment risks
• But: inter-satellite bandwidth remains ~18x below NVLink per-link; all-to-all patterns for wide EP remain infeasible across satellites
• Google Suncatcher prototype: 2 satellites launching by early 2027 google-suncatcher-research.1

Modular tile architectures (Sophia Space TILE):

• 1 m × 1 m tiles with integrated solar, compute, and passive cooling per tile
• 92% power-to-compute efficiency because the entire tile surface acts as radiator, eliminating thermal transport distance spacecomputer-cooling.1
• Scalable from single tiles on host spacecraft to ~2,500-tile data centers
• Bypasses the monolithic-vs-distributed debate: compute becomes a component on any platform
• Ground test 2026, orbit demo 2027-2028

The workload determines the right architecture. For Tier 1 workloads (1-8 GPUs), both monolithic and distributed work equally well — the GPUs don't need inter-satellite links. For Tier 2 workloads (8-72 GPUs, including frontier MoE at EP=64), a monolithic 72-GPU satellite serves the workload entirely within its internal NVLink domain; a distributed architecture would require cross-satellite expert parallelism, which is infeasible at current ISL bandwidths. For Tier 3 workloads (NVL144+, 72+ GPUs), even monolithic satellites approach their power and thermal limits — formation flying or next-generation satellite platforms become necessary.

Reliability Implications of Satellite Size

The choice between many small and fewer large satellites is a portfolio diversification question. The expected value of total GPU losses from catastrophic satellite failures is identical regardless of satellite size — both architectures lose the same fraction of their fleet annually at the same per-satellite failure rate.

The difference is in loss volatility: the standard deviation of annual GPU losses is ~3x higher with 72-GPU satellites vs. 8-GPU satellites for a 10,000-GPU fleet [nonuniform-tensor-parallelism.1, jacklin-small-satellite-failure-rates.1]. Practically:

• Worst single loss event: 8 GPUs (0.08% of fleet) vs. 72 GPUs (0.72%)
• A 72-GPU loss may force immediate session rebalancing and temporarily violate SLAs
• An 8-GPU loss is a routine scheduling event

Satellite size is not the primary reliability driver. Empirical data shows manufacturing maturity dominates: Starlink's failure rate improved from 13% (prototypes) to 0.2% (mature production). A 2010 study of 1,394 satellites found microsatellites and minisatellites equally reliable (~98%) after successful launch jacklin-small-satellite-failure-rates.1.

NVLink domain failure amplification is irrelevant for inference. A single GPU failure in a 72-GPU NVLink domain can halt tensor-parallel execution across the entire domain for training workloads nonuniform-tensor-parallelism.1. But inference typically uses TP1-TP8 per request, and each request is independent — a failed GPU is simply removed from the scheduling pool without affecting others.

Insurance and replacement economics favor smaller satellites. Large constellation operators (Starlink, OneWeb) self-insure, relying on strength-in-numbers. Replacing a small satellite costs ~$0.5-1M and can rideshare on any launch; replacing a 72-GPU satellite costs ~$5-10M and may need dedicated capacity leo-insurance-market.2.

In-Orbit Assembly: Not Needed Near-Term

In-orbit assembly is immature (TRL 4-7 in 2026) and unnecessary for current orbital compute scales payload-space-isam-2025.1.

Starship changes the equation. With 100+ tons to LEO and a 9 m × 18 m fairing, Starship can launch monolithic satellites well beyond near-term needs. A 1 MW satellite (12-50 tons) uses only 12-50% of Starship capacity. Full ground integration testing — impossible for assembled structures — remains a major reliability advantage.

Current assembly milestones:

• DARPA NOM4D Phase 3 (2026): small-scale orbital demos of 1.4 m trusses darpa-nom4d.1
• GITAI S2 (2024): autonomous robotic ISAM tasks outside ISS, TRL 7 gitai-iss-demo.1
• EROSS: full autonomous assembly targeted "after 2035" cordis-eross-iod.1
• NASA's flagship OSAM-1 was cancelled in 2024 after costs reached $2.05B nasa-osam-1.1

Formation flying sidesteps assembly entirely. Google Suncatcher's proposed 81-satellite cluster would provide multi-megawatt aggregate compute without any physical connections between satellites — no joints, no vibration management across structures, no assembly operations google-suncatcher-research.1.

Likely progression:

• 2026-2032: Monolithic satellites + formation flying (no assembly needed)
• 2032-2040: Modular docking for component swap/upgrade (building on servicing heritage)
• 2035+: Large-scale autonomous assembly for multi-MW individual platforms

Quantitative Summary: Satellite Configurations by GPU Count

Note: Suncatcher per-node estimates (~28 kW, ~575 kg) and cluster totals (~2,260 kW, ~47 tons) are derived from the 81-satellite count google-suncatcher-research.1 and mass/power estimates consistent with Google's disclosed orbital parameters. Google has not published per-satellite power or TPU count.

Implications for the Main Analysis

Our main model assumes a generic satellite without specifying GPU count, using mass-per-kW_IT and cost-per-kW_IT as the fundamental parameters. This side page confirms that this parameterization is appropriate — the economics are driven by mass-per-kW_IT regardless of whether that kW_IT comes from 8 GPUs or 72 GPUs on a single satellite.

The key insight for the main analysis: the 100 kW satellite is the industry consensus design point, not the 15-20 kW satellite our framing might imply. SpaceX, Starcloud, K2 Space, and the Handmer/Mach33 independent analyses all converge on 100-130 kW as the near-term target. This means:

1. Our mass budget estimates are consistent — the satellite-mass-budget page estimates 13-35 kg/kW_IT (1.3-3.5 tons for 100 kW), which falls within the 1-5.4 ton range from industry proposals. The SpaceX and Mach33 estimates bracket our model range, with SpaceX's ~10 kg/kW below our optimistic case and Mach33's ~54 kg/kW above our conservative case.
2. A monolithic 72-GPU satellite resolves the NVLink domain question for many current frontier inference workloads — Tier 1 and Tier 2 workloads (up to EP=64) fit within a single satellite's internal NVLink domain, conditional on current batching/context assumptions. Long-context workloads and future NVL144+ architectures may exceed this capacity.
3. The thermal challenge is real but addressable at 100 kW — it shifts from "line item" to "major design driver" but does not constitute a physical veto.
4. Scaling beyond 300-500 kW per satellite requires either formation flying (Google's approach) or technology not yet demonstrated. This doesn't affect near-term feasibility but matters for long-term scaling projections.

Sources

spacex-ai-sat-mini-spacenews

• URL: https://spacenews.com/spacex-offers-details-on-orbital-data-center-satellites/
• Title: SpaceX offers details on orbital data center satellites
• Description: SpaceNews reporting on SpaceX AI Sat Mini specifications and megawatt-class plans
• Summary: AI Sat Mini: 100 kW, more than 170 m length (illustration), ~100 m² radiator, custom D3 chip designed to run hot with radiation protection. Plans for megawatt-class follow-on.

spacex-ai-sat-mini-daniel-marin

• URL: https://danielmarin.naukas.com/2026/03/23/ai-sat-mini-los-centros-de-datos-orbitales-de-spacex-de-180-metros-de-longitud/
• Title: AI Sat Mini: los centros de datos orbitales de SpaceX de 180 metros de longitud
• Description: Daniel Marin's technical analysis of AI Sat Mini dimensions and deployment
• Summary: ~1 ton, ~180 m wingspan (exceeding ISS 108.5 m), SSO, ~100 per Starship V3 launch.

starship-payload-specs

• URL: https://www.eoportal.org/other-space-activities/starship-of-spacex
• Title: Starship of SpaceX - eoPortal
• Description: Technical specifications for Starship payload capacity
• Summary: 9 m diameter fairing, 18 m height, ~1,100 m³ volume, 100+ metric tons to LEO.

nvl72-rack-physical-specs

• URL: https://www.sunbirddcim.com/blog/your-data-center-ready-nvidia-gb200-nvl72
• Title: Is Your Data Center Ready for the NVIDIA GB200 NVL72?
• Description: Physical specifications of the GB200 NVL72 rack
• Summary: 0.6 m × 1.07 m × 2.24 m, 1,360 kg, 120 kW, 72 Blackwell GPUs.

nasa-atcs-overview

• URL: https://www.nasa.gov/wp-content/uploads/2021/02/473486main_iss_atcs_overview.pdf
• Title: NASA ISS Active Thermal Control System Overview
• Description: Technical reference for the ISS external cooling system
• Summary: 70 kW maximum rejection via 6 radiator ORUs (~460 m² total), ammonia loops, ~13,000 kg system mass.

celeroton-space-thermal

• URL: https://celeroton.com
• Title: Celeroton Space Thermal Management Systems
• Description: ESA-funded heat pump technology for space radiator temperature boosting
• Summary: Boosting from 80°C to 150°C reduces radiator area ~60%, COP 3-5, 5-10% compute power cost.

nasa-rosa-gateway

• URL: https://www.nasa.gov/missions/artemis/gateway/a-powerhouse-in-deep-space-gateways-power-and-propulsion-element/
• Title: NASA Gateway Power and Propulsion Element
• Description: Gateway PPE 60 kW ROSA deployment specifications
• Summary: 2 ROSA wings, 60 kW total, 100-120 W/kg, 40 kW/m³ stowed power density.

megaflex-sbir

• URL: https://www.sbir.gov/sbirsearch/detail/388526
• Title: MegaFlex Scale-Up to 175 kW/Wing
• Description: NASA SBIR for Northrop Grumman MegaFlex solar array
• Summary: Up to 200 W/kg, 175 kW per wing, fan-fold circular deployment, TRL 5-6.

nasa-300kw-solar-array-structures

• URL: https://ntrs.nasa.gov/citations/20140000360
• Title: Solar Array Structures for 300 kW-Class Spacecraft
• Description: NASA study validating 300 kW solar array structural feasibility
• Summary: Designed and ground-tested for Solar Electric Propulsion missions.

k2-gravitas-orbital-today

• URL: https://orbitaltoday.com/2026/03/23/k2-space-to-launch-satellite-that-could-pave-the-way-for-orbital-data-centers/
• Title: K2 Space to Launch Satellite for Orbital Data Centers
• Description: K2 Space Gravitas satellite launch and Giga-Class platform development
• Summary: Gravitas ~2 tons, 40 m wingspan, 20 kW, launching March 2026. Giga-Class: 110 kW, 15,000 kg payload.

google-suncatcher-research

• URL: https://research.google/blog/exploring-a-space-based-scalable-ai-infrastructure-system-design/
• Title: Exploring a space-based, scalable AI infrastructure system design
• Description: Google Research blog on the Suncatcher orbital AI compute architecture
• Summary: Illustrative 81-satellite constellation, 1.6 Tbps ISLs demonstrated (bench-scale), Trillium v6e TPUs, 2-satellite prototype by early 2027 with Planet.

nonuniform-tensor-parallelism

• URL: https://arxiv.org/html/2504.06095v1
• Title: Nonuniform-Tensor-Parallelism: Mitigating GPU failure impact
• Description: Analysis of NVLink domain vulnerability to GPU failures at different TP degrees
• Summary: TP64 at 0.1% failure rate: ~10% GPUs idle for training. Higher failure rates (~3x Llama baseline) drop training availability to ~80%. Proposes NTP to mitigate failure amplification in larger scale-up domains.

jacklin-small-satellite-failure-rates

• URL: https://ntrs.nasa.gov/citations/20190002705
• Title: Small-Satellite Mission Failure Rates (NASA)
• Description: NASA study of satellite reliability across mass categories
• Summary: After controlling for design maturity, micro/minisatellites equally reliable (~98%).

payload-space-isam-2025

• URL: https://payloadspace.com/the-state-of-isam-2025/
• Title: The State of ISAM 2025
• Description: Industry survey of in-space servicing, assembly, and manufacturing readiness
• Summary: Servicing TRL 7-9, assembly TRL 4-7, manufacturing TRL 5-7.

gitai-iss-demo

• URL: https://gitai.tech
• Title: GITAI S2 ISS External ISAM Demonstration
• Description: First autonomous robotic ISAM tasks outside the ISS (March 2024)
• Summary: Dual robotic arm achieved TRL 7 for autonomous assembly tasks.

ascend-project-specs

• URL: https://ascend-horizon.eu/activities/
• Title: ASCEND - Advanced Space Cloud for European Net zero emission
• Description: EU Horizon Europe feasibility study for orbital data center infrastructure
• Summary: 10 MW MVP, >1,200 tons, 4,000 m² solar + 2,000 m² radiator, ~120 kg/kW specific mass, requires in-orbit assembly.

starcloud-satellite-progression

• URL: https://www.geekwire.com/2025/starcloud-power-training-ai-space/
• Title: Starcloud plans its next moves after training first AI model in space
• Description: Starcloud satellite roadmap from single-GPU to 100 kW
• Summary: Starcloud-1 (1 GPU, operational), Starcloud-2 (multi-GPU, 2027), Starcloud-3 (100 kW, Starship).

sophia-space-tile

• URL: https://www.geekwire.com/2026/sophia-space-launches-orbital-data-center-plans/
• Title: Sophia Space TILE Architecture
• Description: Modular 1 m × 1 m compute tiles with integrated cooling
• Summary: 92% power-to-compute efficiency, passive cooling, ground test 2026, orbit 2027-2028.

kepler-comms-tranche1

• URL: https://www.kepler.space
• Title: Kepler Communications Tranche 1 Distributed Compute Cluster
• Description: First operational distributed on-orbit computing service
• Summary: 10 satellites, 4× Jetson Orin each, 100 Gbps ISLs, operational March 2026.

china-xingshidai

• URL: https://www.datacenterdynamics.com/en/news/chinese-ai-satellite-constellation-launches-12-satellites/
• Title: Xingshidai AI Satellite Constellation
• Description: China's first operational AI satellite constellation (May 2025)
• Summary: 12 satellites, 744 TOPS each, 8B model, 100 Gbps laser ISLs, target 2,800 satellites.

darpa-nom4d

• URL: https://www.darpa.mil/research/programs/novel-orbital-and-moon-manufacturing-materials-and-mass-efficient-design
• Title: DARPA NOM4D Program
• Description: Novel orbital manufacturing and assembly demonstrations
• Summary: Phase 3 orbital demos in 2026: 1.4 m truss construction and carbon fiber polymerization.

nasa-smallsat-power-soa

• URL: https://www.nasa.gov/smallsat-institute/sst-soa/power-subsystems/
• Title: Small Spacecraft Technology State of the Art — Power Subsystems
• Description: NASA survey of space solar array technologies with specific power data
• Summary: Flown missions clustered ~30 W/kg. State-of-art rigid: up to 200 W/kg. ROSA: 100 W/kg. FOSA: 140 W/kg. Next-gen thin-film targets 500 W/kg (not flight-proven).

act-cchps-space

• URL: https://www.1-act.com/resources/blog/heat-pipes-in-space-cchps/
• Title: Heat Pipes In Space: How CCHPs Are Used In Spacecraft Thermal Control
• Description: Advanced Cooling Technologies overview of constant conductance heat pipes for spacecraft
• Summary: CCHPs transport thermal energy several meters in microgravity. Practical designs up to ~15 feet (~4.6 m). Aluminum extrusions with ammonia working fluid are the standard. Can be bent into 2D and 3D configurations.

Evidence

1. SpaceX's AI Sat Mini delivers 100 kW for AI processors, with more than 170 m length (per scale illustration), ~100 m² radiator area. Uses a custom D3 chip designed to run hotter than terrestrial chips with built-in radiation protection. Plans for megawatt-class follow-on satellites. — spacex-ai-sat-mini-spacenews

2. SpaceX's D3 chip is designed specifically for space, optimized to operate at elevated temperatures (reducing thermal radiator requirements) with integrated radiation protection. This directly addresses the thermal scaling constraint by allowing higher radiator operating temperatures. — spacex-ai-sat-mini-spacenews

3. The AI Sat Mini has approximately 180 m wingspan (larger than the ISS at 108.5 m), ~1 ton mass, and is designed for polar sun-synchronous orbit providing near-permanent sunlight. Starship V3 could carry ~100 units per launch. Future "AI Sat" (larger) planned for Starship V4 with 200+ ton capacity. — spacex-ai-sat-mini-daniel-marin

4. Starship payload fairing: 9 m outer diameter (8 m dynamic envelope), 18 m standard / 22 m extended height, ~1,100 m³ volume, 100+ metric tons to LEO. — starship-payload-specs

5. Mach33/Space Intelligence analysis scaling Starlink V3 from 20 kW to 100 kW: total mass ~5.4 tons. Solar arrays ~2,600 kg (~48%), radiators ~1,000 kg (~18%), bus ~1,800 kg (~34%). Radiators constitute ~7% of total planform area. Concludes radiators are "a line item, not a veto" — solar arrays dominate mass and area. — mach33-cooling

6. The ISS External Active Thermal Control System rejects a maximum of 70 kW (35 kW per loop) using two independent ammonia loops. The HRS consists of deployable radiator ORUs (8-panel systems), with 3 radiators per beam on each of the S1 and P1 truss segments. — nasa-atcs-overview

7. Stefan-Boltzmann heat rejection rates: 418 W/m² at 20°C, 850 W/m² at 80°C, 1,450 W/m² at 127°C. Rule of thumb: 1 kW requires ~2.5 m² radiator at 80°C. For higher power, mechanically pumped fluid loops are required (pumps can fail, loops can leak — the ISS has experienced ammonia leaks). Liquid droplet radiators demonstrated 450 W/kg (November 2025), up to 7x lighter than conventional. — spacecomputer-cooling

8. ESA-funded Celeroton oil-free turbo compressor heat pumps for vacuum/zero-gravity operation. Boosting radiator temperature from 80°C to 150°C reduces required radiator area by ~60%. COP of 3-5, costing 5-10% net compute power. Predicted industry adoption by 2027. — celeroton-space-thermal

9. NASA Gateway Power and Propulsion Element: 2 ROSA wings providing 60 kW total. Represents the highest-power single-deployment ROSA system planned. ROSA specific power is ~100 W/kg per NASA's Small Spacecraft Technology survey [nasa-smallsat-power-soa.1]. — nasa-rosa-gateway, nasa-smallsat-power-soa

10. MegaFlex solar array targets up to 200 W/kg, 175 kW per wing (350 kW two-wing system), fan-fold circular deployment, 10 m diameter ground-tested (TRL 5-6). Designed so each wing can be ground-tested in existing facilities. — megaflex-sbir

11. NASA actively designed and validated solar array structures at the 300 kW class for Solar Electric Propulsion missions, confirming structural feasibility at this power level. — nasa-300kw-solar-array-structures

12. Google Suncatcher: illustrative 81-satellite constellation at 650 km dawn-dusk SSO, 1 km cluster radius, 100-200 m inter-satellite spacing. ISL bandwidth: 1.6 Tbps demonstrated (bench-scale), targeting tens of Tbps via DWDM. Trillium v6e TPUs with 2 krad(Si) radiation tolerance. Two-satellite prototype launching by early 2027 in partnership with Planet. — google-suncatcher-research

13. K2 Space Gravitas satellite: ~2 tons, 40 m wingspan, 20 kW, 12 payload modules, high-power electric thruster. Scheduled to launch late March 2026 on Falcon 9. Founded by former SpaceX engineers. Giga-Class platform in development: 110 kW array power, 15,000 kg payload capacity. — k2-gravitas-orbital-today

14. Sophia Space TILE: tabletop-sized satellite modules combining solar power generation and radiative cooling, connected into racks for scalable LEO computing. First TILE module deliveries to customers targeted for 2028. — sophia-space-tile

15. Kepler Communications launched the first operational distributed on-orbit computing service: 10 satellites, each ~300 kg with 4× Nvidia Jetson Orin modules, 100 Gbps optical ISLs. Launched January 2026, operational March 2026. First on-orbit compute power sold to Axiom Space. — kepler-comms-tranche1

16. Larger scale-up domains increase the blast radius of GPU failures: with TP degree 64, just 0.1% of GPUs in a failed state can cause nearly 10% of allocated GPUs to not contribute to training throughput. At higher failure rates (~3x the Llama report baseline), availability drops to ~80%. A single GPU failure in a 72-GPU NVLink domain can halt tensor-parallel execution across the domain. — nonuniform-tensor-parallelism

17. A study of 1,394 satellites (Dubos & Castet 2010) found microsatellites and minisatellites equally reliable (~98%) within the first 20 years after successful launch. Satellite size is not the primary reliability driver; design maturity and testing investment are. — jacklin-small-satellite-failure-rates

18. 2025 industry survey of In-Space Servicing, Assembly, and Manufacturing: servicing (RPO, docking) at TRL 7-9 and commercially operational. Assembly (robotic construction) at TRL 4-7 with limited orbital demos. Full autonomous assembly not yet demonstrated at scale in orbit. — payload-space-isam-2025

19. GITAI S2 dual robotic arm completed autonomous ISAM tasks outside the ISS in March 2024: ORU maneuvering, flexible material manipulation, fastener attachment/detachment. Achieved TRL 7. — gitai-iss-demo

20. Xingshidai (ADA Space): first 12 AI satellites launched May 2025, each with 744 TOPS and 8B parameter model, 100 Gbps laser ISLs. Target: 2,800 satellites. Primary use: remote sensing with on-board inference. — china-xingshidai

21. The ASCEND project targets a 10 MW MVP at >1,200 tons total space infrastructure (specific mass ~120 kg/kW), requiring 4,000 m² solar panels and 2,000 m² radiators, assembled in orbit using robotic technology. 1 GW target before 2050. — ascend-project-specs

22. GB200 NVL72 rack physical specifications: 0.6 m × 1.07 m × 2.24 m, ~1,360 kg total, ~120 kW power (1.2 kW/GPU average), 72 Blackwell GPUs. — nvl72-rack-physical-specs

23. Casey Handmer estimates a Starlink-derived satellite could produce ~130 kW and host ~200 H100-equivalent GPUs. Variant concept distributes GPUs directly onto solar panel surfaces (~6 kW each), eliminating thermal transport distance while keeping everything on one platform. — handmer-2025-tweet

24. Starcloud progression: Starcloud-1 (60 kg, single H100, operational Nov 2025), Starcloud-2 (multi-GPU H100+B200, ~100x power of Starcloud-1, launching Oct 2026), Starcloud-3 (2-ton, 100 kW, Starship-optimized). — starcloud-satellite-progression

25. Per Aspera estimates a 100 kW orbital data center at 3-5 metric tons total, itemizing solar panels (~700 kg for ~100 kW generation), batteries (a few hundred kg), radiators (~1,000 kg), plus structure, cooling loops, and computers. An earlier back-of-the-napkin calculation in the same source sizes 140 kW generation (for 100 kW average with LEO eclipses) at ~930 kg solar panels and ~500 kg batteries. — peraspera-realities

26. Elon Musk on satellite design: "the solar array is most of the weight on the satellite" and chips should be designed to "run hot" since raising operating temperature by 20% in Kelvin cuts radiator mass roughly in half. — musk-2026

27. Heat pipes and loop heat pipes transfer heat passively via phase change of working fluid (commonly ammonia). Beyond their effective transport distances, mechanically pumped fluid loops (MPFLs) are required, adding mass, power draw, and leak-based failure modes. — spacecomputer-cooling

28. Spacecraft constant conductance heat pipes (CCHPs) can transport thermal energy several meters in microgravity; practical designs reach up to ~15 feet (~4.6 m). Loop heat pipes extend transport distances further, with spacecraft LHP transport lines exceeding 5 m. The NASA TFAWS 2015 heat pipe course confirms heat pipes transport "high rates: up to several kilowatts, over long distances: up to several meters." — act-cchps-space

29. Large constellation operators (Starlink, OneWeb) self-insure rather than purchasing per-satellite insurance, relying on strength-in-numbers. The space insurance market has contracted as mega-constellations don't buy coverage. — leo-insurance-market

30. ESA's EROSS IOD program targets autonomous assembly demonstrations. Full servicing flow demonstrated at DLR in December 2025. Full autonomous assembly missions explicitly targeted "after 2035." — cordis-eross-iod

31. NASA's OSAM-1 (On-orbit Servicing, Assembly, and Manufacturing 1) was cancelled in February 2024 due to "continued technical, cost, and schedule challenges" after costs grew significantly beyond initial estimates. — nasa-osam-1

32. DARPA NOM4D Phase 3 includes two orbital demonstrations in 2026: Caltech autonomous gantry robot constructing a 1.4 m truss from composite tubes, and University of Illinois carbon fiber polymerization on ISS. — darpa-nom4d