Hot AI in Cold Space: Thermal-Crosstalk-Aware Scheduling for Sustainable Orbital AI Clusters

Abstract: Terrestrial AI training faces an unsustainable energy and water crisis, positioning Orbital Data Centers (ODCs) as a "zero operational carbon" alternative. However, the sub-$10μ\text{s}$ communication latency required for distributed LLM training forces ODCs into extreme physical density, triggering a critical "Proximity-Thermal Paradox." As these high-density systems scale into Monolithic Structures or Proximity Swarms, they suffer from intense thermal-fluid crosstalk (heat traps in shared cooling loops) and thermal-radiative crosstalk (mutual heating that blocks deep-space cooling radiators). If left unmitigated, this persistent heat stagnation not only triggers severe thermal throttling that degrades training throughput, but also induces severe thermal fatigue, drastically shortening hardware lifespans and generating premature space e-waste. To make orbital AI truly sustainable, this position paper challenges traditional uniform load-sharing. We propose the Thermal-Aware Heterogeneity Thesis, which treats spatial cooling variances as a primary resource management dimension. Building on this, we introduce Thermal-Load Balancing (TLB), a software framework that dynamically migrates LLM workloads to the coolest available units based on instantaneous fluid temperatures or absorbed radiation. Our analysis demonstrates that TLB resolves thermal bottlenecks to restore Model Flops Utilization (MFU), while simultaneously reducing physical thermal stress. Extending the operational lifespan of orbital hardware is crucial to amortize the massive embodied carbon of rocket launches, outlining a necessary pathway to scale orbital AI without accelerating e-waste.

• The paper introduces Thermal-Load Balancing (TLB), a closed-loop scheduler that uses thermal telemetry to assign asymmetric workloads across liquid-cooled and free-flying orbital AI clusters.
• The paper’s co-simulation shows TLB increases monolithic cluster MFU from 75.1% to 82.7%, reduces peak temperatures, and improves swarm core mean time to failure by 6.15%.
• The paper argues that extending hardware lifespan is more important than modest throughput gains for offsetting rocket-launch emissions, while highlighting network-thermal co-design and lifecycle assessment as open challenges.

Motivation and the Proximity-Thermal Paradox

This position paper addresses the sustainability of Orbital Data Centers (ODCs) for distributed LLM training. While ODCs offer effectively zero operational carbon due to solar power, their environmental viability hinges on amortizing the massive embodied carbon of rocket launches, which requires maximizing lifetime Model Flops Utilization (MFU) and avoiding premature hardware failure. The central obstacle the authors identify is the "Proximity-Thermal Paradox": Tensor Parallelism and synchronous Data Parallelism demand sub-$10\mu\text{s}$ communication latencies, forcing nodes into extreme physical density (e.g., 3-meter proximity), which in turn induces intense thermal crosstalk. Because a global training step's latency is governed by a max over per-node computation times ($t_{step} = \max_i(t_{comp}(i) + t_{comm})$), a single thermally throttled straggler stalls the entire cluster.

The paper distinguishes two architectural paradigms with distinct crosstalk mechanisms. Monolithic Structures rely on centralized liquid cooling loops in which downstream nodes receive coolant pre-heated by upstream nodes, creating inherent thermal imbalance along each flow path. Proximity Swarms of free-flying satellites with independent radiative cooling suffer from mutual geometric shadowing: core nodes have severely reduced Effective View Factors $\rho_i$, blocking deep-space heat rejection via the Stefan-Boltzmann mechanism. In both cases, uniform load-sharing—the default in frameworks like PyTorch DDP—exacerbates congestion at thermally disadvantaged nodes.

The Thermal-Aware Heterogeneity Thesis and TLB Framework

The paper's core argument is that spatial and temporal cooling variance should be treated as a primary, software-schedulable resource rather than a hardware limitation. The proposed Thermal-Load Balancing (TLB) framework is a closed-loop orchestration system with three phases: thermal telemetry aggregation, capability profiling into scores $w_i$, and asymmetric workload slicing. The framework is deliberately policy-agnostic—decision engines ranging from convex solvers to DRL agents can be plugged in without modifying distributed AI communication backends.

The proof-of-concept uses a greedy proportional heuristic under Data Parallelism, assigning micro-batch sizes proportional to $w_i$ while guaranteeing every node at least one unit of work to preserve gradient synchronization participation. For monolithic architectures, upstream nodes receiving coldest fluid receive higher scores; for swarms, peripheral nodes with larger view factors are favored. Practical integration details include space-grade BMC telemetry, a Thermal-Aware Data Sampler interfacing with DDP/Megatron-LM, dynamic-shape compilation or pre-compiled graph buckets to avoid XLA recompilation overhead, and a hybrid triggering policy combining epoch-boundary redistribution for slow orbital thermal dynamics with event-driven evacuation as a safety net near $T_{soft}$.

Evaluation Results

Validation uses a time-stepped thermal-compute co-simulator: 64 nodes across 8 daisy-chained liquid cooling pipes (monolithic) and a $6\times6$ planar satellite grid (swarm). Key quantitative findings:

Metric	Baseline	TLB
Monolithic MFU	75.1%	82.7%
Swarm MFU	90.0%	90.2%
Monolithic peak temp	354.4 K (81.3°C)	353.3 K (80.2°C)
Swarm core temp	357.2 K (84.1°C)	353.9 K (80.8°C)
Swarm edge temp	344.5 K (71.4°C)	351.9 K (78.8°C)
MTTF gain (swarm core)	—	+6.15%
MTTF gain (monolithic outlet)	—	+1.71%

TLB flattens the spatial thermal gradient—raising edge temperatures slightly while lowering core temperatures—and compresses tail latency variance, eliminating synchronization wait times in the swarm case. Lifespan extension is computed via the Arrhenius model with activation energy $E_a = 0.685$ eV; because MTTF depends exponentially on temperature, even modest peak-temperature reductions yield disproportionate durability gains. The authors argue this lifespan extension, not raw throughput, is the primary pathway to amortizing embodied carbon in a zero-operational-carbon regime.

Notably, the paper concedes that absolute MFU gains are modest (7.6% monolithic, 0.2% swarm) and explicitly frames its static greedy heuristic as a lower bound, ignoring multi-hop optical routing latency and transient fluid lag. The evaluation's stated purpose is to validate the existence of the paradox and the solvability of spaceborne heat traps, not to present an optimal scheduler.

Limitations and Open Questions

Several limitations bear directly on the results. The simulation models only idealized crosstalk (linear fluid pre-heating, static view factor geometry); real orbital dynamics—including eclipse cycles, attitude changes, and transient fluid lag—are abstracted away. The heuristic assumes thermal telemetry is accurate and low-latency, and that dynamic batch-size adjustment does not incur prohibitive recompilation costs, an assumption contingent on maturing dynamic-shape compiler tooling. Three open challenges are identified: (1) network-thermal co-design, since reconfiguring FSO links alters node orientation and hence radiative view factors—an unsolved coupled optimization problem; (2) the e-waste versus carbon trade-off, since orbital accelerators cannot be repaired and rigorous Life Cycle Assessments are needed to determine whether zero operational carbon justifies effectively disposable high-end silicon; and (3) federated thermal telemetry standards for multi-vendor swarms that share thermal margins without exposing proprietary silicon layouts.

Conclusion

This paper reframes orbital AI sustainability around thermal heterogeneity management rather than energy efficiency. Its contributions—a formalized Proximity-Thermal Paradox, the Thermal-Aware Heterogeneity thesis, the TLB orchestration framework, and a co-simulation methodology demonstrating restored MFU and extended MTTF—establish that asymmetric workload slicing is both necessary and feasible for dense ODCs. Whether more sophisticated schedulers, network-thermal co-design, and lifecycle analysis can make orbital LLM training genuinely sustainable remains the substantive question the paper leaves open.