Direct wireless large language model inference from high-altitude platform stations under payload constraints

High-altitude platform stations (HAPS) could extend 6G non-terrestrial networks from wireless access to onboard AI inference, but serving capacity is jointly constrained by computing, thermal management , radio circuitry and finite payload resources. We develop a cross-layer framework combining a 110-day LLM workload trace, concurrency-dependent latency, stochastic user activity, thermal simulation , radio-power modelling and reported HAPS payload limits. Long-context requests account for only 1.35% of elapsed-time occupancy yet consume 48.6% of the serving budget for the evaluated 70B configuration. Under a 5% model-defined compute-overload criterion, the 8B/1×H100 and 70B/2×H100 configurations support 5,876 and 991 user equivalents. Integrating computing, thermal, communication, power and mass constraints yields platform-level capacities up to 7,245 and 23,144 user equivalents for Sunglider-class and Stratobus-class platforms. These results show that long-context demand, heat-rejection capability and payload power—not accelerator throughput alone—govern HAPS-based LLM inference feasibility in 6G NTN.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-20
DOI
https://doi.org/10.21203/rs.3.rs-11385664/v1
Primary Topic
Satellite Communication Systems
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Direct wireless large language model inference from high-altitude platform stations under payload constraints

Byeongheon Lee, Young‐Chai Ko
3 links
Research Square
Satellite Communication Systems
preprint

Direct wireless large language model inference from high-altitude platform stations under payload constraints

Byeongheon Lee, Young‐Chai Ko
preprint en

Abstract

High-altitude platform stations (HAPS) could extend 6G non-terrestrial networks from wireless access to onboard AI inference, but serving capacity is jointly constrained by computing, thermal management , radio circuitry and finite payload resources. We develop a cross-layer framework combining a 110-day LLM workload trace, concurrency-dependent latency, stochastic user activity, thermal simulation , radio-power modelling and reported HAPS payload limits. Long-context requests account for only 1.35% of elapsed-time occupancy yet consume 48.6% of the serving budget for the evaluated 70B configuration. Under a 5% model-defined compute-overload criterion, the 8B/1×H100 and 70B/2×H100 configurations support 5,876 and 991 user equivalents. Integrating computing, thermal, communication, power and mass constraints yields platform-level capacities up to 7,245 and 23,144 user equivalents for Sunglider-class and Stratobus-class platforms. These results show that long-context demand, heat-rejection capability and payload power—not accelerator throughput alone—govern HAPS-based LLM inference feasibility in 6G NTN.

Research Square
Korea University (KR)
Industry, innovation and infrastructure
Satellite Communication Systems
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon