Hot AI in Cold Space: Thermal-Crosstalk-Aware Scheduling for Sustainable Orbital AI Clusters Shuyi Chen
Zhengchang Hua
Southern University of Science and Technology Shenzhen, China [email protected]
arXiv:2606.26150v1 [cs.DC] 23 Jun 2026
Nikos Tziritas
University of Thessaly Lamia, Greece [email protected]
Southern University of Science and Technology Shenzhen, China [email protected]
Georgios Theodoropoulos
Research Institute of Trustworthy Autonomous Systems Southern University of Science and Technology Shenzhen, China [email protected]
Abstract
Keywords
Terrestrial AI training faces an unsustainable energy and water crisis, positioning Orbital Data Centers (ODCs) as a "zero operational carbon" alternative. However, the sub-10𝜇s communication latency required for distributed Large Language Model (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 deepspace 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.
orbital data centers, sustainable computing, embodied carbon, thermalaware scheduling, large language models, satellite swarms
CCS Concepts • Computing methodologies → Distributed computing methodologies; Artificial intelligence; • Hardware → Thermal issues; Impact on the environment. Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from [email protected]. HotCarbon’26, Seattle, WA © 2018 Copyright held by the owner/author(s). Publication rights licensed to ACM. ACM ISBN 978-1-4503-XXXX-X/2018/06 https://doi.org/XXXXXXX.XXXXXXX
ACM Reference Format: Shuyi Chen, Zhengchang Hua, Nikos Tziritas, and Georgios Theodoropoulos. 2018. Hot AI in Cold Space: Thermal-Crosstalk-Aware Scheduling for Sustainable Orbital AI Clusters. In Proceedings of HotCarbon’26. ACM, New York, NY, USA, 6 pages. https://doi.org/XXXXXXX.XXXXXXX
1
Introduction
Terrestrial AI training faces an unsustainable energy and water crisis, making Orbital Data Centers (ODCs) an attractive "zero operational carbon" alternative [4, 19]. However, to truly offset the massive embodied carbon of rocket launches [20], ODCs must maximize their lifetime Model Flops Utilization (MFU) and avoid premature hardware failures. This presents a severe challenge for the most demanding workloads: distributed Large Language Model (LLM) training. LLMs, particularly those utilizing Tensor Parallelism, demand sub-10𝜇s communication latencies [24]. This forces computing nodes into extreme high-density configurations (e.g., 3-meter physical proximity), triggering a critical Proximity-Thermal Paradox: the spatial compactness required for low-latency synchronization inevitably leads to intense thermal crosstalk. Furthermore, distributed LLM training exhibits unique "All-orNothing" synchronization patterns (e.g., All-Reduce operations). A single thermally congested "straggler" node stalls the entire gigawatt-scale cluster [8, 13]. If unmitigated, this persistent heat stagnation not only degrades training throughput but also induces severe thermal fatigue, which drastically shortens hardware lifespans and generates premature "space e-waste." Scaling these high-density ODCs leads to two distinct architectural paradigms, each presenting unique multi-scale thermal interferences. The first paradigm is Monolithic Structures: massive self-assembled units that rely on centralized fluid loops. The unique challenge here is Thermal-Fluid Crosstalk, where heat from highdensity GPU tiles accumulates along the shared cooling flow path, creating downstream heat traps [2]. The second paradigm is Proximity Swarms: free-flying nodes (e.g., 20kW design power) with independent cooling maintained in tight formations [26]. These suffer from Thermal-Radiative Crosstalk, where geometric shadowing
HotCarbon’26, July 16-17th, 2026, Seattle, WA
Chen et al.
and mutual heating block deep-space cooling windows, particularly for nodes trapped in the cluster’s core [18]. Monolithic Coolant Inlet (Cold)
Swarm Coolant Outlet (Hot) High Radiative Cooling
High Radiative Cooling
Thermal Trap
Coolant Inlet (Cold)
Coolant Outlet (Hot)
High Radiative Cooling
Cold Temperature
High Radiative Cooling
Hot
Figure 1: Two ODC architectural paradigms: Monolithic Structures with centralized cooling loops and Proximity Swarms of free-flying nodes. In both paradigms, traditional load-balancing frameworks force structural and operational uniformity, which inevitably leads to synchronous bottlenecks. To make orbital AI physically and environmentally sustainable, we introduce the Thermal-Aware Heterogeneity Thesis. We argue that ODCs must reject uniform loadsharing and instead treat spatial and temporal cooling variance as a primary, schedulable computational resource. Building on this, we propose the Thermal-Load Balancing (TLB) paradigm. TLB is a software orchestration framework that dynamically migrates LLM computation blocks to the coolest available units based on instantaneous fluid temperatures or absorbed radiation. By proactively shifting tasks, TLB actively resolves heat traps, prevents the emergence of thermal stragglers, and mitigates cyclical thermal stress on the silicon. This paper outlines a pathway to scale orbital AI without accelerating space e-waste. Specifically, we make the following contributions: • We define the Proximity-Thermal Paradox in ODCs, detailing how extreme density induces severe Thermal-Fluid and Thermal-Radiative Crosstalk. • We introduce the Thermal-Aware Heterogeneity thesis, establishing that exploiting cooling variance is essential to preventing thermal fatigue and straggler-induced MFU degradation. • We propose the Thermal-Load Balancing framework, providing workload migration strategies to resolve heat traps in both monolithic flow paths and swarm radiative fields. • We formulate an evaluation methodology, modeling a dense multi-node mesh, to demonstrate that TLB restores MFU and extends hardware lifespans by reducing cyclic thermal stress, thereby amortizing the embodied carbon of ODCs.
2
Related Work
To contextualize the novelty of the TLB paradigm, we contrast it with the blind spots in two established research domains: Space-Based Computing and Thermal Management. Recent studies have proposed ODCs for on-orbit edge computing
[3, 4, 19, 22, 25], primarily targeting asynchronous tasks like remote sensing offloading [6, 7, 11]. However, supporting synchronous LLM training requires extremely dense spatial topologies to emulate the sub-10𝜇s latency of terrestrial Optical Circuit Switches [10, 23], introducing unprecedented thermal congestion. Concurrently, spacecraft thermal management relies on radiative and fluid systems [14–16]. Yet, current designs assume isolated satellites or loosely formed constellations [5, 21]. They fail to address the multiscale thermal crosstalk inherent to GW-scale monoliths or densely packed swarms, nor do they account for the rapid thermal fatigue induced by dynamic AI workloads. Terrestrial AI Load Balancing. In terrestrial data centers, distributed AI load balancing optimizes for computational throughput or cooling energy expenditure (i.e., operational carbon) [1, 8, 12, 17, 27]. These models rely on active convective air-cooling and cannot be applied to the vacuum of space. Crucially, because ODCs are powered by solar arrays, operations are effectively zero-carbon. Thus, the sustainability objective must pivot from minimizing electricity usage to maximizing hardware lifespan to amortize the massive embodied carbon of space launches [20]. In summary, existing literature treats spatial topology, thermal dissipation, and AI scheduling as isolated domains. When conventional symmetric load balancers distribute synchronous AI tasks across ODCs, they ignore spatial cooling variances. Thermally congested nodes quickly hit temperature thresholds, triggering hardware throttling that stalls the entire cluster. This creates a catastrophic "straggler" effect [9] while simultaneously inducing severe thermal fatigue that shortens hardware lifespan. Our work bridges this exact gap by treating thermal heterogeneity not as a hardware limitation to be mitigated, but as a fundamental, softwareschedulable resource.
3 System model 3.1 Orbital Data Center Model We model ODCs as high-density satellite clusters and consider two distinct architectural paradigms. Let N = {1, . . . , 𝑁 } be the set of satellite nodes. Each node 𝑖 is characterized by its thermal capacitance 𝐶𝑖 , dynamic power limits (𝑃𝑖𝑑𝑙𝑒 and peak 𝑃𝑚𝑎𝑥 ), and maximum processing speed 𝑆𝑚𝑎𝑥 . Heat dissipation properties differ fundamentally depending on the architecture: Monolithic Structures: The monolithic ODC features a rigid, large-scale (e.g., MW-to-GW scale) macro-structure shaped like a massive unified container, similar to conceptual tether-based orbital space stations (Starcloud-4) [2]. Power generation and heat dissipation for the entire ODC are managed globally by the primary structure via externally mounted large-area heat radiators and solar arrays. Standalone compute units are housed within this structure and managed by a centralized liquid cooling system. As shown in the left part of Figure 1, multiple coolant pipes are routed through the data center, pumping coolant through the compute units before returning it to the heat exchanger. Along the pathway of each coolant loop, upstream nodes receive cold coolant, while downstream nodes receive pre-heated coolant, thereby creating an inherent thermal imbalance across the compute units. Proximity Swarms: The proximity swarm ODC architecture consists of multiple independent satellite computing nodes flying
Hot AI in Cold Space: Thermal-Crosstalk-Aware Scheduling for Sustainable Orbital AI Clusters
HotCarbon’26, July 16-17th, 2026, Seattle, WA
in a tight formation, similar to Google’s Project Suncatcher concept [26]. These nodes form a dense constellation with inter-node distances typically ranging from hundreds of meters to a few kilometers, communicating via high-bandwidth FSO links, as shown in the right part of Figure 1. Each satellite in the swarm is equipped with its own solar panels and heat radiators. Although each node utilizes independent passive radiative cooling, the nodes within the constellation are physically close enough to thermally influence one another. Due to view factor occlusion caused by neighboring nodes, their radiators experience varying effective radiating areas. Furthermore, warmer nodes may radiate waste heat directly toward their neighbors, establishing a dynamic thermal imbalance across the distributed swarm.
For distributed LLM training via Data Parallelism, the global batch (size 𝐵𝑔𝑙𝑜𝑏𝑎𝑙 ) is partitioned such that each node processes a micro-batch of 𝑏𝑖 tokens. The computation time is directly proportional to 𝑏𝑖 and inversely proportional to the thermally throttled speed 𝑆𝑖 (𝑇 ). Crucially, a global training step strictly requires gradient synchronization (e.g., All-Reduce). This means the total step latency is dictated entirely by the single slowest node in the cluster: 𝑡𝑠𝑡𝑒𝑝 = max𝑖 (𝑡𝑐𝑜𝑚𝑝 (𝑖) + 𝑡𝑐𝑜𝑚𝑚 ). This mathematical "max" operation establishes the severe consequence of the Proximity-Thermal Paradox: a single node suffering from intense thermal crosstalk will experience degraded 𝑆𝑖 (𝑇 ), thereby stalling the entire ODC cluster and causing a global collapse in MFU.
3.2
4
Heat Dissipation and Thermal Crosstalk Models
Let 𝑃𝑖𝑛 and 𝑇𝑖𝑛𝑠𝑡𝑎𝑛𝑡 denote the instantaneous chip power and temperature. We link computational load to heat generation non-linearly: 𝑃𝑖𝑛 = 𝑃𝑖𝑑𝑙𝑒 + (𝑃𝑚𝑎𝑥 − 𝑃𝑖𝑑𝑙𝑒 )(𝑆𝑖𝑛𝑠𝑡𝑎𝑛𝑡 /𝑆𝑚𝑎𝑥 )𝛾 . The transient thermal state is dictated by the net heat flow 𝑃𝑖𝑛 − 𝑃𝑜𝑢𝑡 , where heat dissipation 𝑃𝑜𝑢𝑡 is architecturally dependent: Thermal-Fluid Crosstalk (Monolithic): Downstream nodes ) are cooled by fluid pre-heated by upstream components (𝑇𝑓(𝑖𝑙𝑢𝑖𝑑 = −1) (𝑖 −1) 𝑇𝑓(𝑖𝑙𝑢𝑖𝑑 + 𝜂𝑃𝑜𝑢𝑡 ). Consequently, the downstream convective heat (𝑖 ) (𝑖 ) ) dissipation rate (𝑃𝑜𝑢𝑡 ∝ 𝑇𝑖𝑛𝑠𝑡𝑎𝑛𝑡 − 𝑇𝑓(𝑖𝑙𝑢𝑖𝑑 ) degrades, creating downstream heat traps. Thermal-Radiative Crosstalk (Swarms): Dense formations rely on deep-space radiation governed by the Stefan-Boltzmann law 4 4 ). Mutual geometric shadowing severely re(𝑃𝑜𝑢𝑡 ∝ 𝑇𝑖𝑛𝑠𝑡𝑎𝑛𝑡 − 𝑇𝑎𝑚𝑏 stricts this, which we model via an Effective View Factor 𝜌𝑖 ∈ (0, 1]. Peripheral nodes retain larger effective radiating areas (higher 𝜌𝑖 ), while core nodes suffer severe occlusion (significantly lower 𝜌𝑖 ), triggering immediate heat traps.
3.3
Throttling and Lifespan Degradation Models
To avoid hardware damage, nodes enforce a soft temperature threshold (𝑇𝑠𝑜 𝑓 𝑡 ) and a hard threshold (𝑇ℎ𝑎𝑟𝑑 ). When 𝑇𝑖𝑛𝑠𝑡𝑎𝑛𝑡 exceeds 𝑇𝑠𝑜 𝑓 𝑡 , the processing speed 𝑆𝑖𝑛𝑠𝑡𝑎𝑛𝑡 scales down linearly from 𝑆𝑚𝑎𝑥 , capping at a minimum safe speed 𝑆𝑚𝑖𝑛 once 𝑇ℎ𝑎𝑟𝑑 is reached. Furthermore, we model thermal-induced hardware degradation using the standard Arrhenius equation, where the Mean Time To Failure (MTTF) is exponentially dependent on temperature: 𝑀𝑇𝑇 𝐹 ∝ exp(𝐸𝑎 /𝑘𝐵𝑇 ). Here, 𝐸𝑎 is the activation energy (e.g., 0.685 eV) and 𝑘𝐵 is the Boltzmann constant. Because of this exponential relationship, even minor reductions in peak operating temperatures yield disproportionately large extensions in hardware lifespan.
3.4
Workload Model with Asymmetric Data Parallelism
To accelerate training when memory is sufficient, we apply Data Parallelism. Assuming every computing node holds a full replica of the model, the training data batch is split across different nodes. Each node processes a certain volume of data, determined by the configured micro-batch size.
The Thermal-Load Balancing (TLB) Framework 4.1 A Generalized Orchestration Architecture TLB is designed as a generalized, closed-loop orchestration framework, fundamentally decoupling the load-balancing mechanism from any specific scheduling policy. It continuously operates through a three-phase control loop: 1) Thermal Telemetry: Aggregating real-time physical metrics, such as instantaneous silicon temperatures, local fluid pre-heating levels, and dynamic view factor occlusions; 2) Capability Profiling: Translating raw thermal data into an abstract capability score (𝑤𝑖 ) that quantifies each node’s real-time thermal margin; and 3) Asymmetric Workload Slicing: Dynamically adjusting the distribution of the AI workload (e.g., micro-batch sizes in Data Parallelism) to match the profiled capabilities. This abstraction allows ODC operators to plug in arbitrarily complex decision engines, ranging from convex optimization solvers to Deep Reinforcement Learning (DRL) agents, without altering the underlying distributed AI communication backends.
4.2
Proof-of-Concept: Proportional Heuristic
For this position paper, rather than deploying a computationally expensive global optimization solver, we implement a lightweight, greedy heuristic to demonstrate the viability of the TLB paradigm. Standard Data Parallel frameworks enforce a uniform batch distribution, allocating 𝐵𝑔𝑙𝑜𝑏𝑎𝑙 /|N | tokens to each node. This severely penalizes the entire cluster when nodes with degraded cooling drop their processing speed 𝑆𝑖 (𝑇 ). TLB breaks this uniformity by calculating the thermal capability score 𝑤𝑖 . The workload 𝑏𝑖 assigned to node 𝑖 is strictly proportional to its score: 𝑤𝑖 𝑏𝑖 = 1 + (𝐵𝑔𝑙𝑜𝑏𝑎𝑙 − |N |) · Í (1) 𝑘 ∈ N 𝑤𝑘 where every node receives at least one baseline unit of work to maintain gradient synchronization participation, and any fractional rounding remainders are allocated to the highest-scoring nodes. We define an aggressiveness multiplier 𝛼 and compute the capability scores based on the architectural paradigm: Monolithic Flow-Path Priority: We prioritize upstream nodes that receive the coldest fluid. Let 𝐿𝑝𝑖𝑝𝑒 be the total number of nodes on a cooling pipe, and 𝑖𝑑𝑥𝑖 be the zero-based index of node 𝑖 along
HotCarbon’26, July 16-17th, 2026, Seattle, WA
Chen et al.
that pipe. The capability score is: 𝑤𝑖 = 1.0 + 𝛼 · (𝐿𝑝𝑖𝑝𝑒 − 𝑖𝑑𝑥𝑖 )
5.1 (2)
Swarm Radiative Priority: We prioritize nodes on the periphery of the cluster with an unobstructed view of deep space. By directly utilizing the Effective View Factor (𝜌𝑖 ) introduced in our thermal model, which inherently accounts for geometric shadowing, the capability score is: 𝑤𝑖 = 1.0 + 𝛼 · 𝜌𝑖
(3)
By proactively shifting the heaviest computational burdens to nodes with high 𝑤𝑖 , TLB inherently limits the heat generation of central or downstream nodes, preventing them from hitting 𝑇𝑠𝑜 𝑓 𝑡 and triggering the straggler effect.
4.3
Integration with the Distributed AI Stack
To transition TLB from a theoretical model to a deployable system, it must bridge the gap between hardware telemetry and the AI application layer. At the hardware level, TLB leverages standard space-grade baseboard management controllers (BMCs) to poll instantaneous silicon temperatures and coolant states. At the application layer, TLB interfaces with distributed AI frameworks (e.g., PyTorch DistributedDataParallel (DDP) or Megatron-LM). Instead of relying on a static DataLoader that evenly shards the dataset, TLB implements a Thermal-Aware Data Sampler. When the profiling engine updates the capability scores 𝑤𝑖 , the sampler dynamically adjusts the micro-batch size 𝑏𝑖 for the upcoming training steps. A key practical challenge in dynamic workload slicing is memory pre-allocation. Terrestrial LLM training heavily relies on static computational graphs (e.g., XLA compilation) to maximize hardware utilization. Dynamically changing batch sizes mid-flight can trigger expensive graph recompilations. To mitigate this, TLB assumes the use of emerging dynamic-shape compilers (e.g., PyTorch 2.0 torch.compile) or pre-compiled computational graph buckets corresponding to a discrete set of allowable batch sizes. This ensures that thermal-induced workload shifts do not introduce prohibitive software overhead. Triggering Policies and Overhead Mitigation. Executing the TLB control loop at every training step would introduce unacceptable synchronization delays. Therefore, TLB employs a hybrid policy: workload redistribution occurs at coarse-grained intervals (e.g., epoch boundaries) to handle slow-moving thermal dynamics like orbital eclipses. Conversely, an event-driven safety net triggers immediate asynchronous task evacuation if hardware telemetry detects a node rapidly approaching 𝑇𝑠𝑜 𝑓 𝑡 , guaranteeing thermal safety while bounding orchestration overhead.
5
Performance Evaluation
To validate the Thermal-Aware Heterogeneity thesis, we implement a proof-of-concept simulation of the TLB framework. Since this paper focuses on the necessity of a paradigm shift rather than proposing an optimal algorithmic solver, we benchmark a greedy variation of TLB against a standard homogeneous load balancer.
Simulation Methodology
To model the multi-scale thermal dynamics of ODCs, we developed a time-stepped thermal-compute co-simulator that evaluates distributed LLM training workloads executing via Data Parallelism. For the Monolithic architecture, 64 nodes are distributed across 8 liquid cooling pipes (8 nodes daisy-chained per pipe). Downstream nodes process coolant that has been pre-heated by upstream nodes. For the Proximity Swarm architecture, 36 independent satellites fly in a dense 6 × 6 planar grid formation. We implement a radiative view factor model where nodes dynamically cast thermal shadows on each other; edge nodes retain larger effective radiating areas (higher 𝜌𝑖 ), while the core nodes suffer severe radiative occlusion (significantly lower 𝜌𝑖 ). We compare two scheduling paradigms: 1) Baseline (Uniform): Shards the global micro-batch evenly across all nodes, mimicking standard PyTorch DDP behavior; 2) TLB (Greedy): Implements the proportional allocation heuristic. We track instantaneous hardware temperatures, dynamic processing speeds (throttling), and global step synchronization latency.
5.2
Mitigating Heat Traps (Spatial Analysis)
Figure 2: Average Node Temperatures Over the Iteration (Baseline vs. TLB) The spatial thermal distribution reveals the fundamental flaw of homogeneous scheduling in ODCs. Figure 2 shows the average node chip temperatures in the Monolithic and Proximity Swarm architectures. Under the Baseline policy, the Proximity-Thermal Paradox manifests rapidly. In the Monolithic configuration, heat progressively accumulates along the cooling loops, pushing the downstream nodes (Row 7) into thermal saturation (up to 354.4 K / 81.3◦ C). In the Swarm configuration, the core nodes become severe
Hot AI in Cold Space: Thermal-Crosstalk-Aware Scheduling for Sustainable Orbital AI Clusters
heat traps reaching 357.2 K (84.1◦ C) due to high radiative view factor occlusion, while the periphery satellites remain significantly cooler (344.5 K / 71.4◦ C). TLB dynamically maps the thermal topology and deliberately shifts the heaviest computational burdens to the upstream Monolithic nodes and the periphery Swarm nodes. This asymmetric workload slicing neutralizes the cooling variance. For the Monolithic setup, the maximum temperature drops to 353.3 K (80.2◦ C). For the Swarm, the extreme thermal gradient is flattened, with core temperatures reducing to 353.9 K (80.8◦ C) and edge temperatures rising to 351.9 K (78.8◦ C), effectively eliminating spatial heat traps.
5.3
Restoring MFU and Hardware Lifespan
Flattening the temperature curve directly extends the physical durability of the hardware. In orbital computing, where the operational carbon footprint is zero, extending hardware lifespan is the primary mathematical pathway to amortizing the massive embodied carbon of rocket launches. Limitations and Evaluation Scope. While the absolute throughput improvements (e.g., 7.6% MFU gain in the Monolithic cluster, 0.2% in the Swarm) may appear modest, this evaluation intentionally employs a rudimentary static heuristic. By ignoring secondary bottlenecks like multi-hop optical routing latency and transient fluid lag, this heuristic serves merely as a lower-bound baseline. Rather than presenting a theoretically optimal scheduler, our primary objective is to empirically validate the existence of the ProximityThermal Paradox and the solvability of spaceborne heat traps. Ultimately, extending hardware lifespans by mitigating thermal cycling, even alongside marginal MFU gains, is the fundamental prerequisite for amortizing the immense embodied carbon of space launches.
6
Figure 3: Tail Latency & MFU of node computation times (Baseline vs. TLB). The resolution of spatial heat traps translates directly into measurable improvements in computational efficiency and hardware sustainability. Latency Variance and MFU: In the Baseline scenario, nodes trapped in the thermal core exceed temperature thresholds and trigger hardware throttling. Due to the synchronization requirement of LLM training, the entire cluster must wait for the slowest overheated nodes. Figure 3 demonstrates this straggler effect: the Baseline Monolithic architecture exhibits severe step-like latency variance, capping the MFU at 75.1%. TLB compresses this tail latency spread, improving the Monolithic MFU to 82.7%. In the Proximity Swarm, while the Baseline MFU is already high at 90.0%, TLB perfectly equalizes the computation time across all 36 nodes (indicated by the flat TLB latency distribution), raising the MFU to 90.2% and completely eliminating synchronization wait times. Extending Lifespan to Amortize Embodied Carbon: Using an apparent activation energy (𝐸𝑎 ) of 0.685 eV, we calculate the cumulative thermal degradation to determine the relative hardware lifespan extension provided by TLB. Our analysis demonstrates that the TLB policy extends the MTTF of the most thermally constrained nodes, yielding a 6.15% lifespan increase for the core node in the Proximity Swarm and a 1.71% increase for the downstream outlet node in the Monolithic cluster, compared to the Baseline. Under uniform loading, central nodes experience severe thermal fatigue due to relentless temperature oscillation and sustained peak temperatures (e.g., 357.2 K / 84.1◦ C). By smoothing out the thermal gradients and capping peak temperatures at safe operational limits, TLB prevents these nodes from entering cyclical throttling regimes.
HotCarbon’26, July 16-17th, 2026, Seattle, WA
Discussion and Conclusion
Terrestrial AI’s voracious appetite for energy and water is driving the industry toward ODCs. However, this transition fundamentally shifts the root cause of unsustainability: while operations become strictly zero-carbon, the environmental burden is entirely frontloaded into the massive embodied carbon of rocket launches and the impending accumulation of space e-waste. This paper highlights that the sub-10𝜇s latency demands of LLM training force ODCs into extreme physical density, triggering a Proximity-Thermal Paradox that causes premature hardware failure and collapses MFU. By introducing the Thermal-Aware Heterogeneity Thesis and the TLB framework, we demonstrated that resolving spatial heat traps through asymmetric software scheduling can restore synchronous throughput and mitigate cyclic thermal fatigue. Extending hardware lifespans via software orchestration is not merely a performance optimization; it is the indispensable prerequisite for amortizing the embodied carbon of space-based AI. To fully realize sustainable orbital computing, the sustainable computing community must confront several critical open challenges: 1. Dynamic Network-Thermal Co-design: Terrestrial LLM clusters rely on dynamic network reconfiguration (e.g., Optical Circuit Switches [23]) to optimize all-to-all communication. In space, reconfiguring FSO links within a proximity swarm requires mechanical or optical steering. This directly alters a node’s physical orientation, instantly changing its radiative view factors and thermal state. Co-optimizing dynamic network topologies with instantaneous orbital thermodynamics remains an entirely unsolved challenge. 2. The E-Waste vs. Carbon Trade-off: While software schedulers like TLB can extend hardware survival, orbital chips currently cannot be repaired. The community must conduct rigorous Life Cycle Assessments (LCAs) to answer a provocative question: does the "zero operational carbon" benefit of ODCs mathematically justify a paradigm of "disposable" high-end AI accelerators? Addressing this may eventually necessitate designing orbital hardware explicitly for in-space recycling. 3. Federated Thermal Telemetry for Heterogeneous Swarms: Future GW-scale ODCs will likely be multi-tenant constellations
HotCarbon’26, July 16-17th, 2026, Seattle, WA
composed of hardware from diverse vendors. Executing clusterwide thermal scheduling requires nodes to share their physical statuses. Developing open standards to broadcast "thermal margins" without exposing proprietary silicon thermal layouts or payload capabilities is essential to building cooperative, sustainable orbital ecosystems.
References [1] Rui Chen, Bo Liu, WeiWei Lin, JianPeng Lin, HuiWen Cheng, and KeQin Li. 2023. Power and thermal-aware virtual machine scheduling optimization in cloud data center. Future Generation Computer Systems 145 (2023), 578–589. [2] Ezra Feilden, Adi Oltean, and Philip Johnston. 2024. Why we should train AI in space. Lumen Orbit Inc (2024). [3] Ran Ginosar and David Steenari. 2025. Beyond Traditional Payload Data Handling: Micro-Datacenter in Space for Converged Software-Defined Storage and Payload Processing. In 2025 European Data Handling & Data Processing Conference (EDHPC). 1–7. [4] Carlos Guimarães, Alessio Netti, Markus Sauer, Florian Zeiger, Hans-Peter Huth, and Elizaveta Boriskova. 2026. A Survey on Satellite Computing: Connecting the Dots Between Networks and Applications. IEEE Communications Surveys & Tutorials 28 (2026), 567–592. doi:10.1109/COMST.2025.3579525 [5] Mohamad Hnayno, Ali Chehade, Henryk Klaba, Hadrien Bauduin, Guillaume Polidori, and Chadi Maalouf. 2022. Performance analysis of new liquid cooling topology and its impact on data centres. Applied Thermal Engineering 213 (2022), 118733. [6] Yifei Hu and Wenbin Gong. 2023. An On-Orbit Task-Offloading Strategy Based on Satellite Edge Computing. Sensors 23, 9 (2023). doi:10.3390/s23094271 [7] Qiangqiang Jiang, Lujie Zheng, Yu Zhou, Hao Liu, Qinglei Kong, Yamin Zhang, and Bo Chen. 2025. Efficient On-Orbit Remote Sensing Imagery Processing via Satellite Edge Computing Resource Scheduling Optimization. IEEE Transactions on Geoscience and Remote Sensing 63 (2025), 1–19. doi:10.1109/TGRS.2025.3528015 [8] Mingwei Li, Jilin Zhang, Jian Wan, Yongjian Ren, Li Zhou, Baofu Wu, Rui Yang, and Jue Wang. 2020. Distributed machine learning load balancing strategy in cloud computing services. Wireless Networks 26, 8 (2020), 5517–5533. [9] Shigang Li, Tal Ben-Nun, Salvatore Di Girolamo, Dan Alistarh, and Torsten Hoefler. 2020. Taming unbalanced training workloads in deep learning with partial collective operations. In Proceedings of the 25th ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming. 45–61. [10] Yingzhi Li, Baisong Chen, Haolun Du, Ziming Wang, Heming Hu, Xuetong Li, Huan Qu, Jie Li, Weipeng Wang, Min Tao, et al. 2026. Integrated Optical Wireless Communication Featured With Optical Phased Array Transceivers for Full-Duplex and NonLine-of-Sight Transmission. Laser & Photonics Reviews 20, 7 (2026), e00822. [11] Yuejin Li, Mi Wang, Kai Hwang, Zhengdao Li, and Tongkai Ji. 2023. LEO Satellite Constellation for Global-Scale Remote Sensing With On-Orbit Cloud AI Computing. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 16 (2023), 9369–9381. doi:10.1109/JSTARS.2023.3316298 [12] Jianpeng Lin, Weiwei Lin, Wentai Wu, Wenjun Lin, and Keqin Li. 2024. Energyaware virtual machine placement based on a holistic thermal model for cloud data centers. Future Generation Computer Systems 161 (2024), 302–314. [13] Rui Lu and Dan Wang. 2025. A Thermal-Aware Workload Scheduler for HighPerformance LLM Inference in Cooling-Regulated Datacenters. SIGENERGY Energy Inform. Rev. 5, 2 (Aug. 2025), 98–104. doi:10.1145/3757892.3757906 [14] Yi-Gao Lv, Yao-Ting Wang, Tong Meng, Qiu-Wang Wang, and Wen-Xiao Chu. 2024. Review on thermal management technologies for electronics in spacecraft environment. Energy Storage and Saving 3, 3 (2024), 153–189. doi:10.1016/j.enss. 2024.03.001 [15] Erdinç Mermer and Rahmi Ünal. 2023. Passive thermal control systems in spacecrafts. Journal of the Brazilian Society of Mechanical Sciences and Engineering 45, 3 (2023), 160. [16] Jianyin Miao, Qi Zhong, Qiwei Zhao, and Xin Zhao. 2021. Spacecraft thermal control technologies. Springer. [17] Sergio Moreno-Alvarez, Juan M Haut, Mercedes E Paoletti, Juan A Rico-Gallego, Juan C Diaz-Martin, and Javier Plaza. 2020. Training deep neural networks: a static load balancing approach: S. Moreno-Álvarez et al. The Journal of Supercomputing 76, 12 (2020), 9739–9754. [18] Yunus Murat, Hatice Mercan, Nedim Sözbir, and Ahmet Selim Dalkilic. 2025. Thermal design for a communications satellite payload module. Proceedings of the Institution of Mechanical Engineers, Part C: Journal of Mechanical Engineering Science 239, 18 (2025), 7629–7646. [19] Shimaa Naser, Maryam Tariq, Raneem Abdel-Rahim, De Mi, Azzam Mourad, Hadi Otrok, Mahmoud Al-Qutayri, Ayman Elnashar, and Sami Muhaidat. 2026. From Connectivity to Multi-Orbit Intelligence: Space-Based Data Center Architectures for 6G and Beyond. arXiv:2603.18601 [cs.ET] https://arxiv.org/abs/2603.18601
Chen et al.
[20] Robin Ohs, Gregory F. Stock, Andreas Schmidt, Juan A. Fraire, and Holger Hermanns. 2025. Dirty Bits in Low-Earth Orbit: The Carbon Footprint of Launching Computers. SIGENERGY Energy Inform. Rev. 5, 2 (Aug. 2025), 26–33. doi:10.1145/3757892.3757896 [21] Yeon-Kyu Park, Geuk-Nam Kim, and Sang-Young Park. 2021. Novel structure and thermal design and analysis for cubesats in formation flying. Aerospace 8, 6 (2021), 150. [22] Cong Peng, Yuanzhi He, Shanghong Zhao, Lingyang Song, and Boyu Deng. 2023. Integration of Data Center into the Distributed Satellite Cluster Networks: Challenges, Techniques, and Trends. IEEE Network 37, 3 (2023), 52–58. doi:10. 1109/MNET.105.2100614 [23] Leon Poutievski, Omid Mashayekhi, Joon Ong, Arjun Singh, Mukarram Tariq, Rui Wang, Jianan Zhang, Virginia Beauregard, Patrick Conner, Steve Gribble, et al. 2022. Jupiter evolving: transforming google’s datacenter network via optical circuit switches and software-defined networking. In Proceedings of the ACM SIGCOMM 2022 Conference. 66–85. [24] Jesmin Jahan Tithi, Hanjiang Wu, Avishaii Abuhatzera, and Fabrizio Petrini. 2025. Scaling intelligence: Designing data centers for next-gen language models. arXiv preprint arXiv:2506.15006 (2025). [25] Guoping Wang, Gang Wan, Zhijuan Su, Yang Wang, Yutong Jia, Gong Li, and Shi Liang. 2025. High-Performance On-Orbit Intelligent Computing and Real-Time Services for Remote Sensing Satellites Based on Large-Scale Computing Power in Space. IEEE Access 13 (2025), 92114–92133. doi:10.1109/ACCESS.2025.3573932 [26] Blaise Agüera y Arcas, Travis Beals, Maria Biggs, Jessica V Bloom, Thomas Fischbacher, Konstantin Gromov, Urs Köster, Rishiraj Pravahan, and James Manyika. 2025. Towards a future space-based, highly scalable AI infrastructure system design. arXiv preprint arXiv:2511.19468 4 (2025). [27] Qing Ye, Yuhao Zhou, Mingjia Shi, Yanan Sun, and Jiancheng Lv. 2022. DLB: A dynamic load balance strategy for distributed training of deep neural networks. IEEE Transactions on Emerging Topics in Computational Intelligence 7, 4 (2022), 1217–1227.