Feb 9, 2026

The Thermal Wall: Why Orbital Data Centers Keep Hitting the Cooling Problem

February 2026 marked an unprecedented race to secure orbital data center approvals. Between January 21 and February 3, three companies filed proposals with the Federal Communications Commission: Blue Origin’s TeraWave constellation (5,408 satellites delivering 6 Tbps connectivity), SpaceX’s million-satellite orbital computing network (targeting 100 GW annual AI capacity), and Starcloud’s 88,000-satellite GPU cluster for machine learning workloads.

The filings painted an optimistic picture. SpaceX projects cost parity with terrestrial data centers by 2028. Blue Origin emphasizes continuous solar power and elimination of cooling infrastructure. Starcloud already has an operational H100 GPU satellite training neural networks in orbit.

On February 6, Voyager Space Technologies CEO Dylan Taylor provided a different perspective in an interview with CNBC: “Two-year time frame for data centers in space would be aggressive.” The reason? Thermal management. Taylor explained: “It’s counterintuitive, but it’s hard to actually cool things in space because there’s no medium to transmit hot to cold. All heat dissipation has to happen via radiation, which means you need to have a radiator pointing away from the Sun.”

While companies compete for FCC approval, the engineering constraint remains unchanged. No satellite operator has demonstrated thermal management at the megawatt scale these proposals require.

The Physics of Space Cooling

The assumption that “space is cold” leads to underestimating thermal challenges. The vacuum of space eliminates the two primary cooling mechanisms used on Earth: convection (air or liquid cooling) and conduction (heat sinks). Only radiative heat transfer remains.

The Stefan-Boltzmann law governs radiative cooling: Q = εσA(T⁴ - T_env⁴), where ε represents emissivity (0.85-0.95 for optimized coatings), σ is the Stefan-Boltzmann constant, A is radiator surface area, and T is temperature. The T⁴ dependence creates a critical constraint. Small temperature changes require massive radiator area increases.

A single Nvidia H100 GPU consumes approximately 350W in terrestrial data centers. Dissipating this power in space requires roughly 1.1 m² of radiator surface area. A DGX H100 system with eight GPUs needs over 16 m² of radiators, exceeding typical satellite body dimensions (1-2 m³ volume for smallsats).

Radiator mass compounds the problem. Industry estimates suggest 1 kg radiator mass per 10-50W heat rejection capacity. A 10 MW orbital data center could require 200,000 to 1,000,000 kg in radiators alone. At $100/kg launch costs (optimistic 2030s projection), this translates to $20-100 million just for thermal management infrastructure.

Thermal cycling presents an additional challenge. Low Earth Orbit satellites experience temperature swings from +120°C in direct sunlight to -150°C in Earth’s shadow every 90 minutes. Hardware must reject heat without freezing components during eclipse. Thermal stress accumulates over five to ten-year mission lifespans, degrading performance and increasing failure rates.

Current spacecraft thermal systems manage kilowatt-scale power budgets for communications satellites. Edge AI satellites like D-Orbit’s AIX constellation and STAR.VISION operate successfully at 100-500W. AI training clusters require 10-50 megawatts. The gap represents a 100-1,000× scaling challenge with no proven solutions.

What Current Projects Are Doing

Several projects are testing orbital computing thermal management, each at different power scales and maturity levels.

China’s Three-Body Computing Constellation

China’s Three-Body constellation represents the most operationally mature orbital computing infrastructure. Twelve satellites have been operational since May 2025, each delivering approximately 744 TOPS of computing power with an estimated 500W thermal load.

The constellation uses passive radiators sized for sub-kilowatt heat rejection. Five POPS (Parallel Orbit Processing Satellites) handle distributed computing tasks. The system works because power levels remain within proven spacecraft thermal management capabilities.

China targets 2,800 satellites with 1,000 POPS by 2030, maintaining the distributed, lower-power node approach rather than concentrating megawatt-scale computing on individual satellites.

Google Project Suncatcher

Google’s Project Suncatcher plans to launch two prototype satellites in 2027 equipped with radiation-hardened Trillium TPU v6e processors. The company selected a dawn-dusk sun-synchronous orbit specifically to minimize thermal cycling.

Google has not publicly disclosed radiator specifications or thermal management strategies. The 80-satellite tight constellation concept suggests power budgets remain in the single to low double-digit kilowatt range per satellite, orders of magnitude below SpaceX’s proposals.

The 2027 launch represents thermal validation testing. Google projects cost parity with terrestrial data centers in the “mid-2030s,” a more conservative timeline than SpaceX’s 2028 claim.

ESA ASCEND

The European Space Agency’s ASCEND (Advanced Space Cloud for European Net zero emission and Data sovereignty) program allocated €300 million through 2027 for orbital data center research. A demonstration mission is planned for 2026.

ASCEND pursues a modular architecture where robotic assembly distributes thermal load across multiple units. This approach acknowledges that concentrating compute power on single satellites creates thermal bottlenecks.

The program explicitly lists thermal management as a core research objective. The conservative European approach prioritizes validation before scaling, reflecting awareness that thermal challenges remain unsolved.

D-Orbit AIX and STAR.VISION

Operational edge AI satellites demonstrate what works today. D-Orbit’s AIX constellation and STAR.VISION’s computing platforms run at 100-500W power budgets with proven passive thermal management. STAR.VISION’s STRING unit delivers 300 TOPS while remaining within current spacecraft thermal capabilities.

These systems perform inference workloads (running pre-trained models) rather than training. AI training requires 10-100× more power than inference. The operational success of edge AI satellites proves that orbital computing is viable at specific power scales. It does not validate megawatt-scale proposals.

Engineering Solutions and Scaling Limits

Four primary approaches exist for spacecraft thermal management. Each faces fundamental scaling limitations.

Passive Radiators

Passive radiators offer maximum reliability with no moving parts. Aluminum honeycomb or carbon composite structures coated with high-emissivity materials (ε = 0.85-0.95, approaching perfect blackbody radiation) emit infrared radiation to deep space.

The limitation is surface area. Radiators must be large relative to heat rejection requirements. Mass penalties restrict economically viable power levels. A 10 MW data center requiring 200,000-1,000,000 kg of radiators becomes economically unviable even with optimistic launch cost reductions.

Heat Pipes

Heat pipes transfer thermal energy from processors to radiator panels through passive phase-change cycles. Working fluids like ammonia or water provide high thermal conductivity (20 W/cm² maximum flux density) without pumps.

Heat pipes excel for distributing heat across spacecraft structures but do not eliminate radiator area requirements. They move heat efficiently but cannot reduce the total radiator surface needed for a given power level.

Active Cooling Loops

Pumped fluid loops provide higher heat transport capacity than passive systems. The trade-off is mechanical complexity. Pumps create failure points, require power, and reduce long-term reliability.

Multi-year mission life with mechanical components remains challenging. The satellite industry generally avoids active cooling for precisely this reason. Increasing capability while decreasing reliability runs counter to spacecraft design principles.

Advanced Materials

Graphene thermal conductors, carbon nanotube composites, and phase-change materials for thermal buffering exist at Technology Readiness Level 3-5. Laboratory demonstrations show promise. Space qualification requires extensive testing.

Carnegie Mellon’s radiation-hardened neuromorphic chips targeting 2026 CubeSat deployment demonstrate progress in power-efficient processors, but even optimized chips require proportional thermal management.

The scaling problem persists across all approaches. Radiative cooling follows physical laws that cannot be engineered away. Launching hundreds of square meters of radiator panels creates structural challenges independent of material advances.

The Timeline Problem

Industry timelines for orbital data center viability vary significantly depending on who is making the projection.

SpaceX-xAI Claims: 2028 Cost Parity

SpaceX’s orbital data center filing projects that space-based AI will become cheaper than terrestrial computing by 2028. This timeline assumes Starship launch costs drop to $100-500/kg and thermal management solutions scale linearly with increasing satellite power budgets.

The 2028 timeline depends on two independent variables: launch economics (showing progress) and thermal physics (showing no equivalent progress). Linear scaling of thermal solutions has not been demonstrated.

Financial Analysis: 2030s Reality

Deutsche Bank analysis cited in industry coverage projects orbital data centers will “reach close to parity” with terrestrial facilities “well into the 2030s.” Voyager Technologies CEO Dylan Taylor stated on February 6 that even a two-year timeline would be “aggressive.”

The discrepancy centers on thermal R&D timelines. Launch costs are declining on a demonstrated trajectory. Radiation hardening shows measurable progress with NASA’s 22nm processor programs and commercial chip testing. Thermal management at megawatt scale remains at TRL 3-4, Technology Readiness Level 3-4 (proof-of-concept, not validated in relevant environment at scale).

Required Development Path

Three phases must complete before megawatt-scale orbital computing becomes viable:

2026-2027: Prototype Demonstrations ESA ASCEND’s 2026 demo mission, Google Suncatcher’s 2027 prototypes, and Starcloud-2’s planned October 2026 multi-GPU satellite will test thermal management at smallsat scales. Success proves feasibility at kilowatt levels, not megawatt levels.

2028-2030: Multi-Year Thermal Validation Five-year mission life testing under accelerated thermal cycling conditions. Monitoring radiator degradation from micrometeorite impacts and atomic oxygen exposure in Low Earth Orbit. Measuring cumulative thermal stress on electronics. This phase cannot be compressed. Time-dependent failure modes require time to observe.

2030-2035: Economic Viability Window Achieving cost parity requires both thermal solutions scaling AND launch cost reductions. If thermal management cannot scale beyond single-digit kilowatts per satellite, orbital data centers remain limited to niche applications regardless of launch costs.

The real bottleneck is not launch economics. It is fundamental physics. Radiative cooling efficiency cannot be improved beyond Stefan-Boltzmann limits. Satellite surface area constrains radiator deployment. These are not engineering refinement problems. They are hard physical limits.

Neuromorphic Computing as Thermal Efficiency Alternative

Power efficiency directly determines thermal feasibility. Neuromorphic processors offer 10-100× better power efficiency than GPUs for specific workloads.

Power Comparison

An Nvidia H100 GPU consumes 350W TDP for approximately 2,000 TFLOPS FP8 performance (5.7 TFLOPS per watt). Intel’s Loihi 2 neuromorphic processor achieves 1 million neurons at approximately 1W power consumption with 10 billion synaptic operations per second.

The ArkSpace Exocortex Constellation targets 100 million neurons at 50-100W power budgets. This represents 10-100W for equivalent computational tasks (application-dependent), compared to 1,000+W for GPU clusters performing the same functions.

Thermal Advantages

Lower absolute power reduces radiator requirements proportionally. A 50-100W neuromorphic payload requires 0.5-1 m² radiator area versus 10+ m² for GPU clusters. This translates to 5-10 kg thermal system mass versus 100+ kg, making deployment economically viable within current launch costs.

Event-driven processing provides additional benefits. Neuromorphic circuits consume power only during active spiking events. Most neurons remain inactive most of the time. Dynamic power scaling means idle periods approach near-zero power consumption, reducing average thermal load.

Challenges and Limitations

Neuromorphic processors are not general-purpose computing platforms. They cannot run arbitrary code like CPUs or GPUs. Current applications center on pattern recognition, sensor processing, and autonomous control.

AI training workloads still favor GPUs. Neuromorphic training algorithms remain at research stages. Performance for large language models or generative AI is unproven.

The ArkSpace position is not that neuromorphic processors replace general-purpose orbital data centers. The thesis is that for specific applications (neural data processing, autonomous satellite operations, sensor fusion), neuromorphic architectures offer a path to thermally viable orbital computing where GPU approaches encounter physical limits.

Path Forward

Thermal management is the bottleneck, not radiation hardening or launch costs. Edge AI satellites at 100-500W power levels are operational today. D-Orbit, STAR.VISION, and China’s Three-Body constellation prove distributed orbital computing works within current thermal management capabilities.

General-purpose orbital data centers at megawatt scales remain unsolved at TRL 3-4. No demonstration exists.

2026: Validation Year

ESA ASCEND’s demonstration mission will provide data on modular thermal architectures. Starcloud-2’s multi-GPU satellite will test higher power budgets in the 500W-1kW range. These missions establish whether thermal management scales beyond current proven limits or encounters hard physical barriers.

2027: Prototype Testing

Google Suncatcher’s dual-satellite launch begins five-year thermal validation. This timeline cannot be compressed. Thermal stress accumulates over time. Single-orbit testing does not reveal multi-year degradation modes.

2028-2030: Viability Assessment

Multi-year thermal data determines whether orbital data centers can achieve the reliability required for commercial deployment. Economic models depend on this data. If satellites cannot maintain performance over five-year missions due to thermal cycling, the business case collapses regardless of launch costs.

Critical Engineering Question

Can radiator technology achieve less than 1 kg mass per 100W heat rejection? This ratio determines economic viability. Can active cooling systems survive 5-10 year missions without mechanical failure? Can thermal cycling be managed without degrading compute performance over time?

If the thermal wall proves insurmountable at megawatt scales, orbital computing follows a different path. Many distributed lower-power nodes (China’s Three-Body approach) succeed where concentrated high-power data centers fail. Specialized applications using power-efficient processors (neuromorphic, custom ASICs) become viable where general-purpose GPU clusters remain economically impossible.

While SpaceX, Blue Origin, and Starcloud race for FCC approval, Voyager’s CEO identified the actual constraint. It is not regulatory. It is physics. The thermal wall remains unbreached at megawatt scales. The 2026-2027 demonstration missions will reveal whether orbital data centers are thermally viable or thermally impossible. The announcements are easy. The cooling is hard.