LLMSpace: Carbon Footprint Modeling and Sustainability Analysis of Large Models Running on Low Earth Orbit Satellites

LLMSpace is the first carbon modeling framework for Large Language Model (LLM) inference on Low Earth Orbit (LEO) satellites. It comprehensively considers key factors such as operational carbon, embodied carbon, and radiation-hardened hardware, revealing core trade-offs between carbon footprint, inference latency, hardware design, and operational lifespan, and providing a systematic tool for sustainability assessment of space AI deployment.

The rapid development of large language models has brought severe energy and carbon crises, with data center carbon emissions becoming a global focus. Running LLM inference on solar-powered LEO satellites is considered a potential solution, but existing analyses often ignore key carbon footprint dimensions:

• Launch emissions: Rocket launches produce large amounts of greenhouse gases
• Satellite manufacturing: Carbon footprint across all supply chain links
• Radiation-hardened hardware: Special treatment required for space environments increases cost and carbon footprint
• Limited lifespan: LEO satellites need replacement every 5-10 years, repeatedly incurring the above costs

Is this solution more environmentally friendly than ground data centers? A comprehensive assessment of the life-cycle carbon footprint is needed.

Modeling Scope

LLMSpace covers a broader range of carbon footprint dimensions:

• Operational carbon: LLM inference energy consumption (prefill/decode phases), thermal management, communication, attitude control, etc.
• Embodied carbon: Satellite platform, radiation-hardened AI accelerators/memory, launch services, etc.
• Peripheral subsystems: Ground stations, data relay satellites, etc.

LLM-specific Workload Characteristics

• Prefill-Decode behavior: Significant energy consumption differences between compute-intensive prefill and memory-intensive decode phases
• Token generation mode: Large energy consumption span from short queries to long document generation
• Batch processing dynamics: Need to balance throughput and latency under resource constraints

LLMSpace is the first carbon modeling framework specifically for LLM inference on LEO satellites.

LLMSpace integrates multidisciplinary knowledge to build the model:

• Life Cycle Assessment (LCA): Tracks carbon emissions from raw material acquisition to disposal, focusing on rare material mining, aerospace component manufacturing, and launch emission factors
• Orbital mechanics and energy budget: Simulates solar energy collection efficiency, discharge during Earth shadow periods, and propellant consumption for orbit maintenance
• AI workload simulation: Based on real LLM inference traces, evaluates the energy consumption impact of model size, quantization strategies, and batch processing optimization
• Radiation effect modeling: Analyzes single-event effects, total dose effects, and the protective effect of radiation-hardened designs

These methods ensure the accuracy and comprehensiveness of the model.

Based on real configurations, LLMSpace reveals core trade-offs:

• Carbon footprint vs. inference latency: Reducing latency requires more powerful hardware and higher power consumption, but increasing utilization can amortize costs
• Hardware design vs. operational lifespan: Highly radiation-hardened hardware extends lifespan but increases launch costs
• Orbital altitude vs. coverage efficiency: Low orbits have low latency but require frequent launches; high orbits have longer lifespans but higher energy consumption
• Satellite scale vs. service capacity: High-capability satellites serve more requests but have higher manufacturing and launch costs

The framework helps find the optimal balance point for each trade-off.

Open Source Value

The LLMSpace source code has been open-sourced, supporting:

• Verification review and improved result credibility
• Parameter sensitivity analysis
• Scenario expansion (other orbits/AI workloads)
• Policy-making support

Practical Implications

• Scenario dependence: Global coverage/low-latency services (e.g., emergency communications) may be more optimal, while fixed-batch tasks are more environmentally friendly on the ground
• Scale effect: Expanding constellations amortizes embodied carbon
• Technological progress: Reusable rockets and efficient AI chips reduce carbon footprint
• Energy structure: Space solar energy is clean, but manufacturing and launch still produce emissions

The framework provides an objective assessment tool.

Limitations

• Data uncertainty: Limited data on some emission factors
• Dynamic factors: Difficult to predict changes in space environment and ground demands
• Multi-objective optimization: Need to integrate carbon footprint, cost, latency, etc.

Future Directions

• Integrate real-time space weather data to improve radiation prediction accuracy
• Develop multi-objective optimization algorithms to find Pareto frontiers
• Expand the model to applications such as remote sensing processing and on-board autonomous decision-making

LLMSpace is an important first step and needs continuous improvement.

LLMSpace provides the first systematic framework for sustainability assessment of space AI, reminding us that technological innovation must consider environmental impacts. As AI and space technologies develop, such tools will become the core of responsible innovation, helping to find a balance between technological progress and the protection of Earth's environment.