Reduced-Mass Orbital AI Inference via Integrated Solar, Compute, and Radiator Panels

This paper proposes a distributed compute architecture for Sun-Synchronous Orbit satellites that integrates solar, compute, and radiator functions into small panels to achieve high specific power and thermal efficiency, enabling a single Starship-launched satellite to deliver over 100 kW of compute power per metric ton and support thousands of simultaneous large language model inferences.

The Gist

Imagine you want to build a supercomputer in space to run Artificial Intelligence. Usually, when engineers design these space computers, they treat the three main parts like separate rooms in a house:

1. The Solar Panels (the power plant) go on the roof.
2. The Computer Chips (the brain) go in the basement.
3. The Radiators (the air conditioning) go on the side to dump heat.

This paper proposes a radical new idea: Stop building separate rooms. Instead, build a single, giant, flexible sheet where the power plant, the brain, and the air conditioner are all baked into the same layer of material.

The authors call this the ISCR (Integrated Solar, Compute, and Radiator) architecture. Here is how it works, explained with simple analogies.

1. The "All-in-One" Sandwich

Think of a standard solar panel as a thick, heavy sandwich. It has a glass top, a heavy frame, and a backing. It's bulky.

The ISCR panel is like a high-tech tortilla. It is incredibly thin and light.

• The Top Layer: Ultra-thin solar cells (like a skin) that catch sunlight.
• The Middle Layer: The computer chips (the "brain") are sandwiched right underneath the solar cells.
• The Bottom Layer: A "Vapor Chamber" (the "air conditioner").

Why is this cool?
In a normal satellite, the computer chips get hot, and that heat has to travel through wires and metal to get to the radiator. In this design, the chips are sitting on top of the radiator. It's like putting a hot pizza directly on a cooling rack instead of on a plate. The heat escapes instantly.

2. The "Magic Cooling Blanket" (Vapor Chamber)

The secret sauce here is the Vapor Chamber. Imagine a flat, sealed metal sheet filled with a tiny bit of water.

• When the computer chip gets hot, the water underneath it boils into steam.
• The steam rushes to the cooler edges of the panel.
• It turns back into water and drips back down.

This cycle moves heat incredibly fast. Because the computer is so close to this "magic blanket," the chips stay cool (around 40°C or 104°F).

• Why does cool matter? Hot computer chips get sluggish and leak energy. Cool chips are faster, use less electricity, and last longer. It's the difference between a runner sweating in a heavy coat versus running in a breeze.

3. The "Space Blanket" Satellite

The authors imagine building a satellite that isn't a box, but a giant, rolled-up carpet.

• The Launch: You fold this carpet up tight and stuff it into a SpaceX Starship rocket. It fits easily.
• The Deployment: Once in space, you blow up air-filled tubes along the edges (like inflating a raft). The carpet unrolls itself into a massive sheet, stretching 2 kilometers long and 20 meters wide.
• The Scale: This single sheet contains 16,000 of those "tortilla" panels.

4. What Can It Do?

This isn't just a calculator; it's an Orbital AI Data Center.

• The Power: It generates over 16 Megawatts of computing power. That's enough to run thousands of complex AI conversations at the same time.
• The Efficiency: Because the chips are cool and the solar cells are super-light, this satellite produces 5 times more computing power per ton of weight than current satellites.
• The Analogy: If a normal satellite is a heavy, fuel-guzzling truck, this ISCR satellite is a lightweight, electric sports car that carries the same cargo but uses a fraction of the fuel.

5. Why Do This? (The "Why Bother?" Factor)

You might ask, "Why put a data center in space?"

• Sunlight is Free: In a specific orbit (Dawn-Dusk), the sun never sets. The satellite gets 24/7 solar power, unlike Earth data centers that need batteries or backup generators at night.
• No Water Needed: Earth data centers use massive amounts of water to cool their servers. Space has no water, but it has the vacuum of space, which is the ultimate heat sink.
• Scalability: If you need more power, you don't build a bigger building; you just roll out a bigger sheet of "tortillas."

The Bottom Line

This paper suggests we stop thinking of space satellites as "boxes with solar panels on top." Instead, we should think of them as giant, intelligent solar sails.

By merging the power source, the computer, and the cooling system into one thin, flexible sheet, we can launch a supercomputer that is lighter, cheaper, and faster than anything we can build on Earth. It turns the harsh environment of space into the perfect home for the next generation of Artificial Intelligence.

Technical Summary

1. Problem Statement

The paper addresses the limitations of terrestrial and current orbital data centers for Artificial Intelligence (AI) workloads.

• Terrestrial Constraints: Earth-based data centers face constraints on power and water resources, environmental opposition, and permitting delays.
• Current Orbital Limitations: Existing concepts for orbital data centers (ODCs) typically separate solar power generation, compute modules, and thermal radiators into distinct structures. This separation leads to:
  • High Mass: Dedicated structural substrates for solar panels and radiators increase the total launch mass.
  • Thermal Inefficiency: High-temperature radiator designs force compute chips (ICs) to operate at high junction temperatures (often >85°C), leading to increased leakage currents, reduced clock speeds, and lower energy efficiency.
  • Scalability Issues: Monolithic designs struggle to scale to the massive power requirements (multi-MW to GW) needed for large-scale AI inference without prohibitive launch costs.

The authors propose a paradigm shift to achieve >100 kW of compute power per launched metric ton, enabling single-launch satellites capable of massive AI inference tasks.

2. Methodology: The ISCR Architecture

The authors propose the Integrated Solar, Compute, and Radiator (ISCR) architecture. Instead of separate subsystems, the ISCR integrates three functions into a single, distributed array of thin panels.

Core Design Principles

• Panel Integration: Each panel (1–4 m²) integrates:
  1. Solar Cells: On the front face (using thin-film Perovskite/Si Tandem or Triple-Junction GaAs).
  2. Compute: Integrated circuits (ICs) sandwiched between the solar layer and the radiator.
  3. Radiator: A vapor chamber backplane that serves as the sole mechanical substrate for the solar cells, eliminating the need for a separate solar panel structure.
• Thermal Management:
  • The system operates in a Dawn-Dusk Sun-Synchronous Orbit (SSO) for continuous solar illumination.
  • Vapor Chambers: Each panel uses a water-based vapor chamber to spread heat from the compute ICs to the entire back surface.
  • Low-Temperature Operation: The radiator operates at ~25–30°C (back face), allowing IC junction temperatures to remain near 40°C. This is significantly lower than the 85–105°C typical of liquid-cooled terrestrial or high-temp orbital systems.
• Deployment Mechanism:
  • Panels are rolled around a central bus cylinder (similar to ROSA solar arrays).
  • Pneumatic Deployment: Argon gas inflates Kevlar/polyimide tubes along the panel edges to unroll and stiffen the array without motors or pyrotechnics.
  • Scale: A full satellite fits within a single SpaceX Starship payload volume (approx. 22m x 2200m deployed array).

Technical Specifications

• Power Density: Targeting 1 kW per panel.
• Mass Density: Estimated at 3.1 kg/m² total (including solar, compute, and radiator).
• Specific Power: ~506 W/kg for the solar array alone; ~112 kW/ton for the total satellite system (including propulsion and structure).
• Communication: Wired inter-panel links (copper) providing 100 GB/s duplex bandwidth; fiber optic links to a central hub.

3. Key Contributions

1. Structural Integration: The novel use of the vapor chamber radiator as the mechanical support for solar cells eliminates the dedicated solar panel structure, drastically reducing mass.
2. Thermal-Electrical Co-Optimization: By maintaining low junction temperatures (~40°C), the design allows for:
   • Lower supply voltages.
   • Higher clock rates (up to 25% faster than 60°C liquid cooling).
   • Significantly reduced static leakage current.
   • Result: ~30% improvement in energy per token compared to high-temperature designs.
3. Distributed Compute Architecture: The paper demonstrates how Large Language Model (LLM) inference can be distributed across thousands of panels using Pipeline Parallelism (dividing attention blocks across the array) and Tensor Parallelism (within small 2x2 panel clusters), mimicking terrestrial rack architectures but at a planetary scale.
4. Scalability Analysis: The design scales linearly with launch mass. A single 150-ton Starship launch can deploy a 16 MW computational satellite, capable of running thousands of simultaneous LLM inference sessions.

4. Results and Performance Analysis

The authors modeled a full-scale satellite (150 tons, 16 MW compute power) and simulated LLM inference performance.

• LLM Inference Capability:
  • A 512-panel subarray can run a representative LLM with a 500,000 token context window and 128 attention blocks.
  • Throughput: 553 tokens/sec/session across 256 simultaneous sessions.
  • Total Capacity: A full 16 MW satellite can support >7,900 concurrent inferences.
• Thermal Performance:
  • At 1000 km SSO, the back-face radiator temperature remains below 24°C.
  • IC junction temperatures are maintained near 40°C, enabling the use of Low-Voltage Threshold (LVT) transistors for better performance.
• Mass and Power Metrics:
  • Specific Power: 112.5 kW/ton (system level), exceeding the SpaceX goal of 100–150 kW/ton.
  • Solar Specific Power: 506 W/kg (5x higher than conventional ~100 W/kg).
• Cost Efficiency:
  • While launch costs are a variable, the reduction in radiator mass and the elimination of separate structures make the ISCR competitive even if launch costs remain high.
  • Lower energy-per-token reduces the required solar array size for a given compute output.

5. Significance and Future Outlook

• Economic Impact: If realized, ISCR could dominate the AI computation market, potentially generating trillions of dollars in value by providing scalable, continuous, and efficient compute resources.
• Technological Shift: The paper argues that inference (which is latency-tolerant for asynchronous queries) is the immediate opportunity for space computing, whereas training requires tighter synchronization that is harder to achieve across distributed panels.
• Manufacturing: The design relies on mass-producible, continuous-line manufacturing (similar to flexible solar modules and PCBs), which is critical for the economics of deploying tens of thousands of panels.
• Future Work: The authors identify key areas for further research, including:
  • 3D thermal simulations and structural dynamics (torsional stability) of the 2km array.
  • Radiation qualification for thin-film cells and custom ASICs in LEO.
  • Optimization of custom ASICs (NRE costs) vs. merchant GPUs.
  • Detailed trajectory analysis for station-keeping and orbit transfer.

Conclusion:
The ISCR architecture represents a fundamental rethinking of orbital data center design. By merging solar, compute, and thermal functions into a single, low-mass, distributed panel array, it overcomes the mass and thermal bottlenecks of current designs. This enables a new class of satellites capable of delivering massive AI inference capabilities with unprecedented energy efficiency and specific power.