acceptodds
Under review as a conference paper at ICLR 2027

LLMSpace: Carbon Footprint Modeling for LLM Inference on LEO Satellites

Abstract

Large language model (LLM) inference imposes growing energy and carbon costs. Solar-powered low Earth orbit (LEO) platforms can reduce reliance on grid electricity, but manufacturing, launch, peripheral, and radiation-hardening requirements introduce substantial embodied carbon. We present LLMSpace, a component-level framework for modeling the carbon and performance of LEO-based LLM inference, including compute hardware, solar arrays, batteries, radiative cooling, launch emissions, and radiation-hardening overheads. Across the evaluated scenarios, launch carbon intensity is the primary determinant of the orbital–terrestrial carbon crossover, while radiation-hardening power overhead is an important secondary factor. Workload and GPU evaluations further show that prompt and output lengths, batch size, request rate, and hardware throughput substantially affect inference energy and amortized embodied carbon per token. These results show that LEO inference is not inherently lower-carbon; its carbon benefit depends on launch accounting, mission lifetime, terrestrial grid intensity, hardware efficiency, workload characteristics, and system utilization.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM, Space Computing, Carbon Estimation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_ynpgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

LLMSpace estimates the carbon footprint of running LLM inference on low Earth orbit satellites, including emissions from hardware, launch, solar panels, batteries, and cooling. It compares orbital and terrestrial computing and studies how assumptions about launch emissions, hardware, lifetime, and workload affect the results.

Critical strengths
  • LLMSpace brings several carbon sources into one analysis, including power and cooling equipment that simpler models did not consider.
  • The paper acknowledges its uncertain assumptions and shows that changing them can reverse the comparison between orbital and terrestrial computing.
Critical weaknesses
  • The connection to LLMs is weak. LLM experiments mostly show familiar effects of batching, throughput, and utilization without clearly leading to new insights about orbital computing. The main findings follow from general carbon accounting.
Questions for authors
  1. Which findings actually depend on LLM behavior? If you replaced the LLM experiments with a generic workload described by power, throughput, and utilization, what important insight would be lost?

  2. How do the LLM experiments change the satellite design or deployment decision? Can you show a case where prompt length, output length, or batching changes the preferred satellite configuration or whether orbital computing is worthwhile?

  3. When would orbital inference be preferable to low-carbon computing on Earth? Your low-carbon terrestrial scenario performs best. Can you identify a realistic situation where space offers a benefit while meeting the same workload, response-time, and availability requirements?

1 comment

Thank you for asking which design decisions change when we model the LLM explicitly.

Q1. What depends on LLM behavior? The launch accounting and basic utilization effects apply to many workloads. An average power and throughput profile can reproduce carbon for a fixed deployment. It loses the ability to predict how a different prompt and output mix changes the feasible batch size, memory requirement, and node count. Prompt tokens drive prefill and populate the key-value cache. Output tokens extend that cache and require sequential decode iterations. Batching can increase aggregate throughput while increasing queueing or changing per-request response times. We will distinguish time to first token, time per output token, and aggregate throughput and clarify that Equation 14 counts both input and generated tokens.

Q2. Can context length change the satellite choice? Yes, in an analytical memory-sizing example for the Qwen2.5-32B model. With a 16-bit key-value cache, each resident token takes 256 KiB. At batch 8, going from 4,096 to 32,768 resident tokens per sequence raises cache memory from 8 to 64 GiB. The long-context cache plus the model weights exceeds the commercial H100's 80 GB memory budget and fits within the B200's 180 GB budget. For 64 concurrently resident requests, independent replicas, and the evaluated batch sizes, the memory calculation gives one H100 node at short context and two H100 nodes at batch 4 or one B200 node at batch 8 at long context. Under Figure 3's one-node-per-platform inventory, the lower-carbon choice among these configurations switches from one H100 platform to one B200 platform. At long context, one B200 node has a modeled power budget of 30 kW versus 40 kW for two H100 nodes, with corresponding array, battery, and radiator requirements. The calculation assumes commercial-equivalent memory capacities and includes weights and cache. Flight latency and reliability require separate measurements. Figure 8 gives a second example. At the aggregate decode target of 250 output tokens/s, the model requires three A100 or two H100 GPU replicas, versus one B200.

Q3. When could orbit be preferable to low-carbon Earth? For the ground-originating inference evaluated in the paper, the 3 gCO₂e/kWh terrestrial case remains the lowest-carbon option within the modeled boundary. The submission establishes no orbital carbon advantage at matched response times and availability for that service. A plausible setting is generating a short event report from data already collected on an Earth-observation satellite, where raw data would otherwise have to be downlinked before the same report could be generated on Earth. ESA's Φsat-2 (https://www.esa.int/articles/introducing-phsat-2) illustrates how onboard processing can reduce downlink volume. For an LLM deployment, both placements would need the same input, report quality, delivery deadline, and on-time availability target. An orbital carbon advantage would require avoided downlink and terrestrial emissions to exceed the incremental onboard manufacturing, launch, and operating emissions, including any redundancy needed to meet the service target. Our current evaluation leaves that end-to-end comparison open.

Reviewer_w6yCOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents LLMSpace, a framework for modeling the carbon footprint of LLM inference on solar-powered LEO satellites, broken down by hardware component. It jointly models embodied carbon from launch, solar arrays, batteries, radiative cooling, and radiation-hardened compute and memory hardware, alongside workload-dependent operational carbon tied to prompt length, generated tokens, batch size, and request rate. Table 1 positions the framework against three prior carbon models for LEO computing, each missing at least one of peripheral breakdown, radiation-hardening modeling, or LLM-specific workload modeling. The paper runs a crossover analysis across three uncertain parameters (launch carbon intensity, radiation-hardening manufacturing factor, radiation-hardening power factor) against terrestrial data center models, and a workload study across nine LLM tasks and three GPU generations. The stated conclusion is that launch carbon intensity is the dominant determinant of whether orbital inference beats terrestrial inference, and that LEO inference does not inherently have a lower carbon footprint.

Critical strengths
  1. Figure 8's real hardware measurements show the ranking of GPU generations by embodied carbon per token reversing between demand regimes, with the B200 lowest under sustained high-throughput demand and highest once intermittent demand forces replication across older GPUs. This is the paper's one verifiable result that depends on its own accounting, not on which input parameter carried the widest assumed range.

  2. Table 1 shows each of the three closest prior frameworks omitting peripheral breakdown, radiation-hardened hardware modeling, or LLM-specific workload characteristics, a gap LLMSpace fills jointly. The calculation of Falcon 9 launch intensity and the DGX H100 component inventory totaling 2.172 tCO2e both follow exactly from their stated inputs.

Critical weaknesses
  1. Section 6.1 tests launch carbon intensity over a 48x range against 4x and 2.7x ranges for the two radiation-hardening factors, a width imbalance that mechanically drives which parameter dominates any sensitivity ranking. Section 5 states its evidence "does not validate... radiation-hardening overheads," yet the Abstract reports the finding on launch intensity without that qualification.

  2. Table 6 checks LLMSpace's own modeling coefficients against independent external values for every row except launch intensity, where the "External" value (12.85 kgCO2e/kg) is itself LLMSpace's own Appendix A derivation and the Reference scenario input. The "LLMSpace" column value (14.5) is a separate literature citation never used as a modeling input, so the row compares two external numbers instead of validating the parameter the crossover analysis identifies as dominant.

Questions for authors
  1. Section 6.1 tests launch carbon intensity over a 48x range against 4x and 2.7x ranges for the two radiation-hardening factors. What is the basis for these different range widths, and does "launch intensity dominates" still hold if the radiation-hardening factors were given a comparably wide range?

  2. Was the 14.5 kgCO2e/kg value on page 3 ever used as a modeling input anywhere in the paper? If not, does Table 6's row for launch intensity compare two literature citations instead of checking the Reference scenario value of 12.85 against an independent source?

  3. Section 2 scopes the target hardware as GPUs and TPUs, and the radiation-hardening cost model is built around COTS CMOS logic. Given that the paper's own motivation is a LEO deployment with tight power constraints, was an ultra-low-power alternative such as neuromorphic or analog compute considered, and would the embodied carbon framework need to change to cover it?

1 comment

Thank you for identifying the sensitivity-range and launch-provenance issues. We will revise the abstract and conclusion to state that the crossover depends jointly on the launch accounting boundary and the assumed hardening design.

Q1. Why are the ranges different, and does the ranking survive equal ranges?

The launch sweep spans a direct-combustion estimate for Falcon 9 (12.85, rounded to 13), an exploratory intermediate setting (450), and a high stress setting (630) motivated by a broader process assessment of a different launch vehicle. The values mix accounting boundaries and launch assumptions. The 6--24x logic-manufacturing and 1.5--4x system-power sweeps are exploratory because matched commercial and hardened H100-class data are unavailable. Appendix B's coefficient-based illustration gives a 3.3--6.6x logic-manufacturing factor. Its power references motivate a starting point near 1.5x, with 2x and 4x used as stress settings. The ranking depends on the comparison point. From the rounded Table 7 inventory, doubling launch intensity from 13 to 26 adds about 0.8 tCO2e/year, while doubling the nominal power factor from 2 to 4 adds 2.4 t/year by increasing peripheral manufacturing and launch mass. At a launch intensity of 450, the respective increases are about 28.2 and 16.8 t/year. Doubling the logic factor adds about 0.45 t/year under the Appendix D inventory convention. An intentionally wide 48.5× increase from each reference setting also changes the ranking. Starting at (IL,SRHA,SRHP)=(13,12,2)(I_L,S_{\rm RHA},S_{\rm RHP})=(13,12,2) and changing one input at a time, the annual increases are roughly 38.7 t for launch, 21.1 t for logic manufacturing, and 112.7 t for power. The latter multipliers are mathematical extrapolations well outside our hardware scenarios.

Q2. Was 14.5 used? Yes, the modeled columns of Table 7 imply 14.5. For the 130.45 kg DGX node, 14.5 gives 1.89 tCO$_2$e, the modeled launch entry. The constructed reference lists 1.68 tCO$_2$e, corresponding to 12.85. The latter is derived in Appendix A from published propellant and emissions inputs and is the basis for the rounded 13 in the Figure 4 sweep. Table 6 juxtaposes two literature-based estimates, with our Appendix A derivation in the external column. It overstates the independence of the comparison. We will document both estimates in Appendix A and apply a single launch input across Table 7.

Q3. Could the framework evaluate analog or neuromorphic hardware? Our numerical evaluation covers GPUs. An analog or neuromorphic accelerator can use the same spacecraft sizing and launch equations given its total power, heat, and mass. It needs its own manufacturing inventory and workload profile, including memory, packaging, and peripheral circuits. For analog in-memory compute, power and performance should include data conversion, calibration, and residual digital processing. A comparable evaluation also requires task quality, delivered throughput, and radiation tolerance for the mapped workload.

Reviewer_W5spOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper builds LLMSpace, a component-level carbon model for running LLM inference on solar-powered low Earth orbit satellites. It covers launch emissions, solar arrays, batteries, radiators, radiation-hardening overheads, and workload-dependent throughput and energy.

Critical strengths
  1. A transparent, open carbon model for orbital compute could help counter one-sided claims that space data centers are carbon-free.
  2. Figure 1 and Section 2 give clear background on inference phases, satellite peripherals, and radiation effects for an ML audience.
Critical weaknesses
  1. The novelty is limited. The contribution is mainly an engineering and life-cycle accounting model. The ML-specific part uses standard measurements and a known simulator, and yields generic conclusions.
  2. The main carbon comparisons do not fully connect the assumed mission lifetime to degradation, availability, or delivered service. Equations (4), (7), and (15)-(16) provide useful sizing relations, but the main evaluation does not show how charging losses, declining solar output, or unavailable compute time affect the ten-year token denominator and crossover.
Questions for authors
  1. Could the authors clarify the methodological novelty beyond the engineering and life-cycle accounting model?
  2. How does the assumed ten-year mission lifetime account for degradation, availability, and delivered compute service in the main carbon comparisons? Although Equations (4), (7), and (15)–(16) provide useful sizing relations, could the authors quantify how charging losses, declining solar output, and unavailable compute time affect the ten-year token denominator and the carbon crossover point?
1 comment

Thank you for identifying the gap between the assumed ten-year mission and the inference service it actually delivers.

Q1. Methodological contribution. LLMSpace links workload-dependent accelerator capacity to spacecraft provisioning. Prompt length, generated length, and batch size change the throughput and energy profiles used to size the compute allocation. The resulting node power sets the array, eclipse battery, and radiator requirements. Their manufacturing emissions and launched mass then enter carbon per delivered token. Figures 7–8 show a consequence. The lower-power A100 has lower embodied carbon per token under sparse demand, while B200 can be preferable when an aggregate decoding target requires additional A100 or H100 replicas. The profiling procedure and simulator are established. The contribution is the link between inference capacity, spacecraft size, and the delivered-service denominator.

Q2. How much do losses, degradation, and downtime matter? The nominal comparison assumes continuous service over ten years. We use Appendix D’s 95-minute orbit with 60 minutes of sunlight and 35 minutes of eclipse. The nominal array collects enough energy during sunlight to power the 20 kW load for 95 minutes under lossless conditions. We power the node directly during sunlight and use the surplus to recharge the battery for eclipse operation.

Delivered tokens. At 90% battery round-trip efficiency, the surplus supports 35×0.90=31.535\times0.90=31.5 minutes of eclipse operation, giving (60+31.5)/95=96.3%(60+31.5)/95=96.3\% of ideal service. With solar output declining linearly to 80% by year ten, the array supplies an average of 95×0.90=85.595\times0.90=85.5 minutes of energy at the design load. After 60 minutes of direct operation, the battery delivers (85.5−60)×0.90=22.95(85.5-60)\times0.90=22.95 minutes in eclipse, giving 87.3% of ideal service. Applying 90% availability during power-feasible periods reduces this to 78.6%. These time fractions equal the delivered-token fractions under sustained demand and constant throughput while operating. The fixed-array scenarios assume sufficient usable battery capacity.

Carbon crossover. For this sensitivity calculation, the rounded Table 7 inventory gives 29.0 tCO₂e in orbital manufacturing emissions and approximately 627 kg of payload. The continuously operating 10 kW terrestrial node has a power usage effectiveness of 1.1 and a grid intensity of 348 gCO₂e/kWh, yielding 337.5 tCO₂e over ten years including hardware manufacturing. We assume both nodes serve the same request mix at the same full-rate throughput. If orbit delivers a fraction ff of Earth’s tokens, the break-even launch intensity is IL∗=(fCEarth−Cmfg)/mI_L^*=(fC_{\rm Earth}-C_{\rm mfg})/m, with carbon totals in kgCO₂e and payload mass mm in kg.

Ten-year scenarioTokens delivered (% of ideal)Break-even launch intensity (kgCO₂e/kg)
Ideal continuous service100.0492
Charging and discharging losses96.3472
Losses and solar degradation87.3424
Losses, solar degradation, and downtime78.6377
Array and battery sized for year ten, with downtime90.0394

With the original array, a 21.4% reduction in delivered tokens raises orbital carbon per token by 27% at a fixed launch intensity. In this sensitivity calculation, the 450 kgC$O_2$e/kg scenario shifts from favoring orbit under continuous service to favoring Earth once losses, degradation, and downtime are included.

Sizing for year ten. The array grows by 30.1% to power the node in sunlight and recharge the battery at 80% of its initial solar output, using the factor (60+35/0.90)/(95×0.80)(60+35/0.90)/(95\times0.80). Equation 16 and Appendix D already apply the 80% depth-of-discharge limit. A 35-minute eclipse at 20 kW requires 11.67 kWh, giving a nominal battery capacity of 11.67/0.80=14.5811.67/0.80=14.58 kWh. With 80% of nameplate capacity retained at year ten, the initial capacity rises to 11.67/(0.80×0.80)=18.2311.67/(0.80\times0.80)=18.23 kWh, a 25% increase at the same depth of discharge. This calculation uses the load-side usable-energy convention and applies round-trip efficiency in the solar recharge budget. After accounting for the added manufacturing emissions and payload mass, sizing for year ten raises the break-even launch intensity from 377 to 394 kgCO₂e/kg.

We will add the delivered-token denominator and these illustrative assumptions to the revision. Battery cycle life, outage timing, and link availability require mission-specific inputs to evaluate latency and uptime targets.

Reviewer_FHkoOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
1
Soundness
2
Read review
Paper summary

The authors present a modeling framework for understanding embodied and operational costs of LLM inference in solar-powered low Earth orbit (LEO) configurations. They note that assumptions about inference workload, hardware used, and whether that hardware was radiation-hardened can be consequential for overall carbon footprint associated with LLM inference serving, such that LEO-based LLM inference is neither necessarily higher or lower in combined operational and embodied carbon intensity than traditional terrestrial LLM inference. For embodied carbon totals, the authors consider peripherals, launch emissions, and radiation-hardening overhead.They consider operational emissions for multiple workload settings (from different HELM tasks) using a Qwen2.5-32B-Instruct-GPTQ-Int8 model, under multiple throughput assumptions

Critical strengths
  1. Compelling motivation and problem statement that is of substantial interest to a portion of the ML research community and beyond
  2. Interesting conclusions if sound
Critical weaknesses

Non-exhaustively:

  1. Despite the relative clarity of the abstract and intro, the manuscript is extremely hard to follow at times. examples of issues: density of acronyms non-standard for ML (most defined only once, some of which are undefined altogether, e.g. EIR, NE), use of informal terms like "checks" and "matched" that make certain sections read like experimental logs (Section 5 is especially hard to read), two blocks of equations that are at times multiple pages away from their references (these would be much clearer inline!), figures without anchors to what exactly is being measured (a single datacenter? if so, of what kind? total global demand?). (Some more minor copy editing is needed, too, e.g. line 045, but this is less consequential to my overall assessment)
  2. Though the extensive focus on embodied emissions associated with LEO-based LLM inference appears detailed (tbh I have less authority to evaluate this part), the operational emissions analysis seems relatively thin. Given the venue, it seems especially important to make sure inference settings are realistic, and that limitations and underlying assumptions are clearly stated. Authors state they "run Qwen2.5-32B-Instruct-GPTQ-Int8 on a single NVIDIA A100 (Falk et al., 2026), H100 (NVIDIA Corporation, 2025b), or B200 (NVIDIA Corporation, 2025a) GPU with batch sizes of 1–8." Single-gpu inference with a 32b model feels disjoint from the hypothetical radiation-hardened node assumed for the embodied hardware calculations
Questions for authors
  1. what are practical takeaways about "orbital–terrestrial carbon crossover" that authors can describe?
  2. I'm not sure I really understood Section 5's "limited forms of evidence for LLMSpace"-- what exactly is being compared? how exactly does LLMSpace differ in methodology from the other references/info used to arrive at different estimates? How should we read the tables? are we trying to arrive at similar estimates, or are we trying to show that LLMSpace models and captures something materially different from existing estimates or models?
  3. What exactly constitutes "embodied" carbon? Including carbon emissions from launching into orbit in "embodied" carbon calculations feels non-standard
  4. Minor clarification question, Casey et al (cited by the authors) themselves note that "Ultimately, when single-event upsets (SEUs) or SEFIs are observed, there are negligible changes in the measured precision or accuracy of the algorithm, and rarely do the errors result in incorrect identification or classification" -- do the authors agree with this assessment? is radiation hardening necessary primarily because of the other two factors mentioned (total ionizing dose, displacement damage)?
3 comments

Thank you for pointing out where the manuscript obscures the scope and practical meaning of the comparison.

Q1. What does the crossover imply for deployment? The crossover is the point at which the orbital platform and terrestrial node emit the same carbon per delivered token for a fixed workload mix. For the nominal ten-year hardened configuration, we reconstruct the payload from Table 7 and obtain parity at approximately 490 kgCO$_2$e per kg of payload against a terrestrial grid at 348 gCO$_2$e/kWh, assuming continuous, equal-throughput service. With the 13 kgCO$_2$e/kg direct-launch scenario, Earth has lower modeled carbon below a grid intensity of about 36 gCO$_2$e/kWh, which includes our 3 gCO$_2$e/kWh case. A heavier power or thermal subsystem and a shortfall in delivered orbital service each lower the allowable launch intensity. These are one-node scenario calculations. A deployment comparison also requires matching workload, response times, and availability.

Q2. What do Tables 6 and 7 establish? Table 6 compares selected manufacturing inputs with published coefficients, such as 120 versus 118 kgCO$_2$e/m2m^2 for silicon panels. We will move the two launch estimates into Appendix A and document their sources and roles. Table 7 compares component coverage under a common hardware configuration. The implementation of Ohs et al. (2025) in that table omits the radiator term, and the implementation of Aili et al. (2025) omits a separately allocated solar-array term. The reference is a constructed component sum. It shares the node inventory with LLMSpace and uses some independently selected coefficients. The totals reflect both component coverage and coefficient choices.

Q3. What is included in embodied carbon? We used the term for upfront emissions attributed to deployment. We will give hardware manufacturing, launch and deployment, and in-service operation their own line items. Manufacturing covers the modeled compute hardware, satellite bus and communications hardware, solar arrays, batteries, and radiators. The DGX inventory covers the components listed in Table 10. The reference launch estimate covers direct combustion CO$_2$ allocated by payload mass, with upstream rocket production, non-CO$_2$ exhaust effects, and re-entry outside that boundary. The orbital operational scenario assumes solar electricity and tracks its energy use. Ground-network energy and satellite end-of-life are outside the reference scenario.

Q4. How do we interpret Casey et al.? (https://ieeexplore.ieee.org/document/10153700) We agree with their assessment for the tested edge accelerators and classification tasks. Small changes in measured classification accuracy can coexist with functional interrupts and device hangs, which matter to service availability. Total ionizing dose and displacement damage motivate component-specific lifetime analysis. Single-event effects also matter for recovery and uptime. Which mechanism limits a mission depends on the orbit, shielding, hardware, and service target. We will describe radiation hardening as a design scenario and discuss lighter mitigation options such as error correction and restart.

Single-GPU profiles and manuscript clarity. The nine HELM tasks use Qwen2.5-32B-Instruct-GPTQ-Int8 on one commercial A100, H100, or B200. We measure throughput and estimate GPU energy from power telemetry at batch sizes 1, 4, and 8, with three runs per setting. Those measurements calibrate Vidur-Energy for unmeasured settings. The eight-GPU inventory and its 10 kW commercial or illustrative 20 kW hardened power budget are node-level inputs. In the future revision, we will state the active replica count and normalize the node inventory by the tokens those replicas deliver. Scaling the single-GPU rates assumes independent replicas without shared-resource contention. Hardened-system throughput, memory performance, and end-to-end latency remain unmeasured. We will replace EIR and NE with author-year labels, place the equations beside their first use, and identify the system and denominator in each figure caption.

Thank you for the response. I truly would have liked to be able to increase my score, but I found the rebuttal unconvincing and do not think I can support acceptance.

One note, I was not looking for a literal definition of "embodied carbon" in general as much as some rationale behind including the arguably "operational" cost of launching into orbit in the "embodied" cost.

I didn't expect full coverage, but at the very least I some acknowledgement of (or refutation or additional work to address) 1. the serious limitation posed by the single model that the LLM workload experiments are run on vs the many different models that one could think to run real workloads on, and 2. the unreliable extrapolation between single-gpu inference measurements and 8-gpu node-level hardware modeling.

In the future revision, we will state the active replica count and normalize the node inventory by the tokens those replicas deliver. Scaling the single-GPU rates assumes independent replicas without shared-resource contention.

I would like to understand what the authors exactly mean by this.

Thank you for clarifying your concerns. Our earlier response gave insufficient attention to both the single-model limitation and the assumptions connecting GPU measurements to the node inventory.

Rationale for allocating launch emissions. Launch is an operational activity for the launch provider. From the perspective of the inference service, it is an upfront deployment activity required to make the computing platform available. We grouped its emissions with manufacturing because both are incurred before service begins and are allocated over the tokens subsequently delivered. This grouping reflects the inference service’s accounting boundary. Reporting manufacturing, launch and deployment, and in-service operation separately gives the same total carbon and makes that rationale clearer. We agree that our previous response should have explained this allocation.

Limitation of evaluating one model. We agree that this is a substantial limitation. The nine HELM tasks vary prompt and output characteristics, but all experiments use Qwen2.5-32B-Instruct-GPTQ-Int8. They establish workload effects for that model and precision setting. Generalization to other model sizes, architectures, and precisions remains untested. Those choices can change memory requirements, achievable batch sizes, inference energy, and the number of GPUs required, potentially changing the preferred hardware configuration and carbon crossover. Applying LLMSpace to another model requires appropriate performance and energy profiles. We should have stated this limitation explicitly in our initial response.

What we meant by independent replicas. We meant a hypothetical deployment in which each active GPU holds a complete model copy and processes a different set of requests. The model used in our measurements fits on one GPU. Requests can therefore be distributed among multiple copies within the eight-GPU node.

Let q1q_1 denote the measured throughput of one GPU and n≤8n\leq8 the number of active replicas. The assumed aggregate throughput is

qnode=nq1q_{\mathrm{node}}=nq_1.

This expression assumes sufficient demand, the same workload mix and batching on each GPU, and negligible interference through shared host resources. For a service duration TT, expressed in seconds, the corresponding embodied carbon per token would be

cemb=Cemb,node+satelliteTqnodec_{\mathrm{emb}}=\frac{C_{\mathrm{emb,node+satellite}}}{Tq_{\mathrm{node}}}.

The numerator includes the complete modeled node and its allocated satellite infrastructure. If only one GPU serves requests, the denominator uses that GPU’s delivered tokens while the numerator still includes the full eight-GPU installation. Eight active replicas justify using 8q18q_1 only under the stated scaling assumption.

Evidence supporting that extrapolation. Our measurements characterize individual commercial GPUs. We have no eight-GPU measurements validating linear throughput scaling or the resulting node-level energy consumption. Shared host processing, scheduling, input/output, and power or thermal constraints could reduce aggregate throughput. GPU energy telemetry also leaves CPU, memory, and other node energy unmeasured. The node power budget and radiation-hardening overheads are separate scenario inputs, and the performance of the hypothetical hardened node remains unverified.

Your concern therefore identifies a real gap in the empirical support for the node-level results. We will explicitly describe those results as conditional on the replica and power assumptions, state the single-model limitation, and specify the active GPU count and full-node carbon allocation in Section 6.2 and the relevant captions.

Scope and contribution of this exploration. Exhaustively measuring every combination of accelerator, model, workload, batch size, precision, and GPU count is infeasible within a single study. Eight-GPU measurements would strengthen the evidence for the scaling assumption. Further questions would remain about TPUs, other GPU generations, larger models, and different workload distributions. These are valuable extensions, and additional profiles can be incorporated into LLMSpace.

The purpose of this initial exploration is to make the carbon implications of LEO-based LLM inference explicit and testable by connecting inference characteristics with spacecraft manufacturing, power, thermal, and launch requirements. Such a framework can help the community identify promising configurations, consequential uncertainties, and priorities for further measurement. Encouraging those investigations is a central goal of this paper.

We recognize that validating central extrapolations matters to the strength of our conclusions. We respectfully hope the assessment will also consider the contribution of establishing this framework and research direction, alongside the acknowledged limits of the current profiling coverage.

Comments on the paper

1 comment

We thank Reviewers FHko, W5sp, w6yC, and ynpg for their careful reading and constructive feedback. We appreciate that all four reviews found the paper informative, either by identifying something useful or by gaining a deeper understanding of the problem.

Orbital LLM inference is an emerging research area. With LLMSpace, we aim to make the assumptions behind orbital and terrestrial carbon comparisons explicit and testable. We hope the framework will encourage follow-up work using measured flight-hardware data, realistic workloads, and comparable service requirements.

  • FHko. We clarified the carbon accounting boundaries, what Tables 6 and 7 establish, and how the single-GPU measurements relate to the modeled eight-GPU node. We will make these distinctions clearer in the manuscript and figure captions.
  • W5sp. We clarified how workload-dependent inference capacity affects spacecraft sizing and carbon per delivered token. We also quantified the effect of the ten-year service assumptions. In our illustrative fixed-array scenario, charging losses, solar degradation, and downtime reduce delivered tokens to 78.6% of the continuous-service baseline and raise carbon per delivered token by 27% at the same launch intensity.
  • w6yC. We clarified the provenance of the launch inputs and showed that the sensitivity ranking depends on the ranges and reference points used. We will qualify the corresponding claims in the abstract and conclusion.
  • ynpg. We provided a concrete example in which longer contexts increase key-value cache requirements enough to change the preferred GPU configuration. We also clarified that the low-carbon terrestrial case remains the lowest-carbon option for the ground-originating workload we evaluated.

We would be grateful if all four reviewers considered these responses in their final assessments. In particular, we respectfully ask Reviewer W5sp to reconsider the clear-rejection recommendation in light of the quantified mission-service analysis and the clarification of the paper’s contribution. If any soundness concerns remain, we would appreciate the opportunity to address them directly.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.