Pith. sign in

REVIEW 3 major objections 4 minor 31 references

Routing LLM Inference to the Cleanest Grid in Real Time

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Carbon-aware placement of LLM inference requests, driven by marginal-emissions signals, reduces modeled GPU-attributable operational carbon by a central 50.9% versus round-robin in a year-long 19-region replay, and a live two-region test…

desk verdict Solid live feasibility result with an honestly-scoped modeled magnitude; the zero-MOER premise is the one load-bearing assumption that needs external validation before the headline number can be trusted. read the letter →

arxiv 2608.06188 v1 pith:PPX3RCYZ submitted 2026-08-06 cs.DC cs.PF

Aleks Bernhard , Arif Baran Yardimci This is my paper
classification cs.DCcs.PF
keywords carbon-awareroutingLLMinferencemarginaloperatingemissionsrate(MOER)geographicloadbalancingGPUenergytelemetryhistoricalreplayrenewablecurtailmentproductionrouteroverlay
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Large-language-model inference is a growing electricity load whose carbon cost varies sharply by grid region and hour, so deciding where each request runs is a no-retraining, no-hardware lever. This paper tries to establish two things: that the lever works in practice, and roughly how much carbon it can save. Live multi-region GPU testbeds ran a marginal-emissions signal as a strict, reversible overlay on a production pressure-based router and moved real traffic with zero observed dispatch failures, though with a small settled-emissions delta. To size the effect, a replay of a full year of hourly marginal-emissions data across 19 CONUS regions estimates a central 50.9% reduction in modeled GPU-attributable operational emissions versus round-robin, with a 95% block-bootstrap interval of 48.5–53.3%, under explicitly stated fleet, policy, workload, and zero-emissions-credit assumptions. The paper is careful that this is an upper-bound opportunity, because the replay dispatches on perfect foresight of historical emissions rather than forecasts, and that the headline depends materially on treating zero marginal-emissions readings as genuine renewable-curtailment intervals.

What carries the argument

The load-bearing mechanism is a multiplicative routing overlay: each cell's pressure weight is scaled by $1 - w \cdot \text{intensity}(\text{region})$, where $w=0$ restores the unmodified production router, making the carbon term a strict, reversible layer. The intensity input is derived from the absolute marginal operating emissions rate (MOER), not the region-normalized percentile signal-index, because the former answers the spatial “which region is cleaner” question. Energy attribution runs on measured GPU-telemetry concurrency curves from DCGM (per-token joules fall roughly 30× from low concurrency to saturation), and every request is settled against historical MOER at five-minute resolution after the fact. The magnitude arm is a capacity-capped cleanest-first dispatcher replayed over 8,760 hourly intervals across 19 CONUS regions, with a 5-day block bootstrap for confidence intervals and a zero-MOER-credit sensitivity sweep that reprices zero readings at a fraction of each region's median non-zero MOER.

What would settle it

Cross-check every timestamp with MOER equal to zero in the licensed historical series against independent operational curtailment records for the same region and hour; if a material share of zero readings occur when no curtailable renewable capacity is available, or if the shifted load exceeds the recorded curtailed megawatt volume, the zero-credit assumption fails. A simpler numerical check: re-run the replay with all zero readings repriced at 1% of each region's median non-zero MOER and see whether the reduction stays in the claimed 34–51% range.

Watch

Extended reading notes

Core claim

The central claim is that marginal-operating-emissions-rate (MOER) signals can steer live inference workloads across real multi-region GPU testbeds with no observed dispatch failures, as a strict and reversible overlay on the existing production pressure router; the magnitude of the effect, which the live runs cannot establish, is then characterized by a year-long historical replay. In the primary configuration — capacity-capped cleanest-first dispatch at a 50%-per-site cap with a modeled multi-turn session layer — carbon-aware placement reduces modeled GPU-attributable operational emissions by a central 50.9% versus round-robin (51.0% versus a modeled carbon-blind load balancer), with a 95% block-bootstrap CI of 48.5–53.3% conditional on the modeled fleet, policy, workload, and zero-MOER-credit assumptions. Because the replay dispatches against historical MOER rather than a forecast, the paper treats this as an upper bound. A practical finding runs alongside: cross-region routing should rank regions by absolute MOER, not by the within-region percentile signal-index, because equal percentiles can invert the absolute clean/dirty ordering, as demonstrated on the H100 testbed with ERCOT-NC versus SOCO.

Load-bearing premise

The magnitude result assumes the signal provider's readings of exactly zero marginal emissions are real renewable-curtailment intervals with enough addressable curtailed volume to absorb the shifted load; if those zero readings are artifacts, or the curtailment volume runs out, the reported 50.9% reduction is not supported and falls to roughly 34% or 24.5% under the paper's own sensitivities.

Editorial extensions

If this is right

  • Production inference fabrics can adopt carbon-aware placement as a default-off, reversible overlay without retraining models or changing hardware; at $w=0$ the router is unchanged.
  • The 50.9% figure is an upper bound under perfect foresight; operationally realized savings will be lower to the extent forecasts miss, so deployed systems should be sized against the roughly 23% soft-multiplier policy as a conservative reference.
  • Routing decisions that compare regions should consume absolute MOER, because the percentile signal-index can send load to the absolutely dirtier grid; the paper shows this inversion concretely on ERCOT-NC versus SOCO.
  • Because per-token energy falls roughly 30× with concurrency, an underloaded clean region can emit more per token than a saturated dirtier one, so utilization and placement must be optimized together.
  • The result depends on crediting zero-MOER intervals as genuine curtailment; if those zeros are repriced at each region's median non-zero MOER the reduction falls to roughly 34%, and dropping the zero timestamps yields roughly 24.5%.
  • Beyond the paper, a live deep-curtailment test, such as CAISO versus ERCOT during spring solar-peak hours, would check whether the modeled roughly 50% opportunity appears in settled live carbon on hardware.
  • The 50% per-site cap lets one region absorb roughly 9.5 times its baseline share, so real fleets with queue, memory, and latency limits will likely realize less than the headline; the paper's roughly 23% soft-multiplier arm is probably closer to operational practice.
  • The same overlay logic could steer other latency-tolerant, geographically flexible workloads such as batch training jobs, which would need to consume the forecast MOER over the job's horizon rather than the instantaneous value.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper, a live deep-curtailment test (e.g. CAISO versus ERCOT during spring solar-peak hours) would check whether the modeled roughly 50% opportunity appears in settled live carbon on hardware.
  • The 50% per-site cap lets one region absorb roughly 9.5 times its baseline share, so real fleets with queue, memory, and latency limits will likely realize less than the headline; the paper's roughly 23% soft-multiplier arm is probably closer to operational practice.
  • The same overlay logic could steer other latency-tolerant, geographically flexible workloads such as batch training jobs, which would need to consume the forecast MOER over the job's horizon rather than the instantaneous value.
  • Because cross-region dispatch breaks prefix-cache locality, the unmodeled prefill-energy cost could shrink the net benefit; an energy-, cache-, and carbon-aware placement policy is a natural next step.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper reports a live systems study of carbon-aware routing for LLM inference. Its central feasibility claim is that a WattTime MOER signal can steer live inference workloads across multi-region GPU testbeds as a strict, reversible overlay on a production pressure-based router, with no observed dispatch failures; the live A/B runs also show the signal moving load and settled emissions in the intended direction, although the 1.45% settled-emissions delta is explicitly not powered. To characterize magnitude, the authors replay a year of hourly MOER across 19 CONUS regions with a synthetic workload, reporting that a capacity-capped cleanest-first policy reduces modeled GPU-attributable operational emissions by 50.9% versus round-robin (95% block-bootstrap CI 48.5–53.3%, conditional on modeled assumptions), with a 34–51% range under alternative treatments of zero-MOER intervals and about 23% for a softer multiplier policy matching the live overlay's form. The paper also documents a practical observation: cross-region comparisons should rank by absolute MOER rather than the region-normalized percentile signal-index.

Significance. If the claims hold, the paper makes a useful systems contribution: a live validation on real GPUs with a production pressure router as the blind control, per-request energy from DCGM telemetry via measured concurrency curves, settlement against historical MOER, and a clearly disclosed perfect-foresight upper bound with conditional confidence intervals. The percentile-versus-absolute MOER observation is practical and empirically grounded. The zero-MOER sensitivity analysis is also honest and well documented. However, the headline 50.9% magnitude rests on a specific curtailment interpretation that is attested only by a personal communication and a release note, and the paper's own sensitivity analysis shows that this interpretation is decisive: repricing zeros at median non-zero MOER drops the reduction to roughly 34% and dropping zero timestamps yields roughly 24.5%. As it stands, the magnitude contribution is provisional; the feasibility contribution is solid and independently valuable.

major comments (3)
  1. [§6, Curtailment (zero-MOER) treatment and Zero-MOER-credit sensitivity; Table 3] The 50.9% headline is computed under the assumption that every numeric 0.0 co2_moer value is a genuine renewable-curtailment interval whose addressable volume exceeds the shifted load. The only support cited is a personal communication from WattTime (2026-07) and a release note, and the licensed MOER series is not redistributed. The paper's own α-sweep shows that repricing each zero at that region's median non-zero MOER reduces the central figure to about 34%, and dropping zero-valued timestamps yields about 24.5%; the zeros also drive the spring (63.7%) and midday (69.4%) peaks. Because this assumption is load-bearing for the magnitude claim, please either provide an independent, auditable validation (for example, comparison with regional curtailment-MW data or published renewable-curtailment statistics over the same intervals) or reframe the primary result as the α-sweep range with the zero-credit interpretation treated as a first-class uncertainty rather than as a single primary-configuration point estimate.
  2. [§6, Energy attribution vs. §5.4; Table 3] The replay uses a fixed near-saturation per-token energy of about 0.104 J/token, while §5.4 reports a roughly 30× spread in per-token energy across the concurrency curve. The paper acknowledges that routing-induced changes in regional concurrency and batching efficiency are "potentially material" but does not bound them. Since the policy concentrates traffic in low-MOER regions and reduces load elsewhere, the percentage reduction could shift materially in either direction. Please add a sensitivity analysis with pessimistic and optimistic concurrency curves, or explicitly downgrade the magnitude claim to the fixed-energy special case.
  3. [§6, Configuration and assumptions; Table 3] The 50%-per-site capacity cap admits up to roughly 9.5× baseline traffic per region, and the text itself states this is "more concentration than most operational fleets could accept." The paper nonetheless reports the capacity-capped cleanest-first result as the "primary configuration," while the gentler soft-multiplier policy matching the live overlay's form gives about 23%. For the magnitude claim to be meaningful, the cap needs a physical or operational justification, or the primary configuration should be the conservative soft policy with the cleanest-first number reported strictly as an upper bound of the modeled frontier.
minor comments (4)
  1. [§1, Contributions] The numbered contribution list contains two items labeled "1." and then continues with "2." and "3."; the list should be renumbered 1–4.
  2. [§6, Zero-MOER-credit sensitivity] The explanation of α=1 is confusing: describing it as "curtailment credited as worth nothing" while also repricing each zero at the region's median non-zero MOER is internally inconsistent. Please clarify that α=1 values curtailment intervals at the region's median non-zero MOER.
  3. [§3.1 and §5.2] The paper recommends disabling the overlay when intensity is unknown or stale, but it does not state whether the live testbeds used this safer behavior or the neutral multiplier-of-1 default. Please specify which behavior was used in the live runs.
  4. [§5.2] The live latency comparison (p95 18.7 s vs 20.9 s) is reported as an 11.7% increase; because the authors correctly describe the run as an observational, shared-resource comparison, a sentence explicitly noting that this latency difference is not statistically testable would help prevent readers from treating it as a measured treatment effect.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the feasibility claim is a live measurement against a real production router, and the replay magnitude is a conditioned simulation whose decisive zero-MOER premise is disclosed and stress-tested, not a fitted input.

full rationale

The central feasibility claim rests on a direct live experiment: the MOER overlay moved real inference load across multi-region GPU testbeds with zero dispatch failures, compared against the authors' own production pressure router. That baseline is a concrete deployed system described in the authors' prior work [22], not a fitted model, and the contrast was measured rather than derived, so the self-citation is not load-bearing in a circular sense. The magnitude claim is an explicitly conditioned historical replay: capacity-capped cleanest-first dispatch against a year of WattTime MOER, with a block-bootstrap CI and a long list of stated assumptions. No parameter is fitted to the headline outcome and then reported as a prediction; the per-token energy scale is a telemetry measurement, and the static-versus-dynamic decomposition is an allocation of the reduction between two uses of the same external data. The most assumption-sensitive element, the treatment of numeric 0.0 MOER values as genuine curtailment intervals, is supported by WattTime release notes and personal communication rather than by the paper's own derivation, and the paper discloses both an alpha-sweep (50.9% down to ~34%) and a drop-zeros sensitivity (~24.5%). Dependence on an external data-model interpretation is a correctness or validation risk, not circularity. No equation or claimed result reduces to its own inputs by construction, no self-citation is invoked to forbid alternatives, and no known result is renamed as new. The paper is self-contained as an empirical study and honestly separates feasibility from modeled magnitude.

Assumptions & free parameters 6 free parameters · 7 assumptions · 0 invented entities

The replay and live analysis rest on a set of domain assumptions about marginal-emissions accounting and WattTime's data, plus several hand-set modeling parameters (capacity cap, workload shape, session stickiness, alpha sensitivity). None of these are fitted to make the headline come out; the paper reports sensitivity to the most consequential ones. No new physical entities are introduced.

free parameters (6)
  • Carbon weight w = 0.25 / 0.50 / 1.00 (eco-low / eco-med / eco-high)
    The carbon weight is a hand-set policy knob defining the overlay's strength; the reported live contrast uses w=1 versus the off arm at w=0.
  • Per-site capacity cap = 50 percent of instantaneous fleet demand per region
    The hand-chosen cap acts as a provisional capacity proxy; the paper admits it permits about 9.5 times baseline concentration per region.
  • Session pinning fraction and lifetime = 40 percent of load multi-turn, mean life 2 hours
    This modeled session layer pins a fraction of load to origin, lagging carbon-optimal placement; the paper states this makes the headline conservative.
  • Workload demand model parameters = diurnal amplitude 0.4, peak 15:00 local; weekday scale 0.75; seasonal peak 1.15 in mid-July; near-uniform origin skew
    The illustrative synthetic workload is used in the replay and is not fitted to a production trace.
  • Zero-MOER alpha sensitivity values = 0 and 1
    The sensitivity parameters reprice each zero-MOER interval at alpha times the region's median non-zero MOER; alpha=0 gives about 51 percent reduction and alpha=1 about 34 percent.
  • Bootstrap block length = 5 days primary, with 1/7/14-day sensitivity
    The block length is chosen as the primary specification; the point estimate is unchanged and intervals widen modestly with longer blocks.
assumptions (7)
  • domain assumption MOER is the correct basis for consequential carbon accounting for load-shifting decisions.
    Invoked in §2.1 with citations [1,19,28,29]; the entire routing and settlement methodology presupposes marginal rather than average emissions factors.
  • domain assumption WattTime numeric 0.0 MOER values represent modeled renewable-curtailment intervals and are distinguishable from missing data.
    Stated in §6 Curtailment treatment, supported by personal communication with WattTime (2026-07) and model version 2026-03-01 release notes; essential for the zero-MOER credit in the replay.
  • domain assumption The volume of addressable curtailed renewable energy exceeds the volume of load shifted.
    Explicit assumption in §6 zero-MOER sensitivity paragraph; required for the zero rate to apply for shifted load.
  • ad hoc to paper A 50 percent per-site capacity cap is an adequate provisional proxy for physical capacity and latency constraints.
    Introduced in §6 configuration as the dispatch-concentration limit; the paper admits it is permissive because one region can absorb about 9.5 times its baseline share.
  • ad hoc to paper The synthetic workload model (diurnal, weekday, seasonal, multi-turn sessions, near-uniform origin skew) is representative enough for magnitude estimation.
    Defined in §6 workload configuration; not derived from a production trace, so the replay result is conditional on this load model.
  • ad hoc to paper A fixed near-saturation per-token energy of about 0.104 J/token is adequate for the replay; concurrency-dependent energy shifts are not modeled.
    Stated in §6 energy/hardware row and §5.4; the paper notes this could change the percentage reduction if efficiency varies with placement.
  • standard math Block bootstrap with a 5-day block length preserves the relevant temporal autocorrelation in the MOER series.
    Used in §6 with Efron and Tibshirani [3]; the paper reports sensitivity from 1-day to 14-day blocks showing modest CI widening.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Routing LLM Inference to the Cleanest Grid in Real Time." pith.science (2026). https://pith.science/paper/PPX3RCYZ

@misc{pith2026260806188,
  author       = {Pith},
  title        = {Pith review of: Routing LLM Inference to the Cleanest Grid in Real Time},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PPX3RCYZ}},
  note         = {Machine review of arXiv:2608.06188}
}
read the original abstract

Large-language-model inference is a fast-growing electricity load whose marginal carbon intensity varies by more than an order of magnitude across grid regions and across the day, making request placement an attractive lever: no retraining, no hardware change. We report a live validation of carbon-aware inference routing on multi-region GPU testbeds driven by marginal operating emissions rate (MOER) signals, with three properties uncommon in prior work: a blind baseline that is an actual production pressure-based router rather than uniform placement; per-request energy attributed from GPU telemetry (NVIDIA DCGM) via measured concurrency curves rather than nameplate TDP; and carbon settlement of every request against historical MOER, not only the forecast that drove the decision. The central live result is feasibility: a MOER signal steered inference across regions with no observed dispatch failures, as a strict and reversible overlay on the production router. To size the effect, we replay a year of hourly MOER across a grid-diverse CONUS fleet. In the primary modeled configuration, carbon-aware placement reduces modeled GPU-attributable operational emissions by 50.9% versus round-robin (95% block-bootstrap CI 48.5-53.3%). Because the replay dispatches against historical MOER rather than a forecast, this is an upper bound under that configuration; forecast error would reduce operationally realized savings. Before session pinning, hourly lowest-MOER routing contributes about 22.4 percentage points, roughly 40% of the 54.0% placement reduction, beyond a static annual-mean-MOER policy. These are modeled results for one fleet and historical year, not a universal estimate. We also record a practical observation: when comparing regions, rank by absolute MOER rather than the percentile signal-index, which is normalized within each region and answers a temporal, not a spatial, question.

Figures

Figures reproduced from arXiv: 2608.06188 by the authors.

Figure 1
Figure 1. Carbon-aware routing as a default-off, reversible overlay. Forecast MOER scales each cell's pressure weight (top); energy and carbon are settled post-hoc against published historical MOER (bottom). At w = 0 the multiplier is 1 and the production router is unchanged [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Measured energy per output token at the two ends of the Phase-0 concurrency sweep — low concurrency vs saturation (NVIDIA DCGM). Continuous batching amortizes fixed overhead, cutting per-token energy ∼28–32×, so utilization — not grid intensity alone — governs per-request carbon. 5 [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Why cross-region comparisons should use absolute MOER, not the percentile signal-index. On a shared absolute scale ERCOT-NC (1253 lbs/MWh) is cleaner than SOCO (1330), yet sits at a higher percentile within its own wide-swing distribution- so a percentile-driven router inverts the ranking and picks the dirtier grid. 5.4 Energy characterization The Phase-0 curves (§3.2) show energy-per-output-token falling ∼30× from … view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Decomposition of the primary 50.9% reduction (vs round-robin). Static annual-mean allocation and the dynamic real-time signal supply ≈31.6 and ≈22.4 percentage points of the 54.0% freely-placeable placement reduction; the modeled session layer costs ≈3 points. Allocati…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 15 canonical work pages

  1. [1]

    Marginal Emissions Factors for the U.S. Electricity System

    Siler-Evans, K., Azevedo, I.L., Morgan, M.G. "Marginal Emissions Factors for the U.S. Electricity System." Environmental Science & Technology, 46(9), 4742–4748, 2012. doi:10.1021/es300145v

    work page doi:10.1021/es300145v 2012

  2. [2]

    Marginal Operating Emissions Rate (MOER): methodology and validation

    WattTime. "Marginal Operating Emissions Rate (MOER): methodology and validation." WattTime technical documentation, 2022. https://watttime.org/data-science/methodology- validation/ (accessed 2026-07)

    work page 2022

  3. [3]

    An Introduction to the Bootstrap

    Efron, B., Tibshirani, R.J. An Introduction to the Bootstrap. Chapman & Hall/CRC, 1993

    work page 1993

  4. [4]

    Carbon-Aware Computing for Datacenters

    Radovanović, A., et al. "Carbon-Aware Computing for Datacenters." IEEE Transactions on Power Systems, 38(2), 1270–1280, 2023. doi:10.1109/TPWRS.2022.3173250

    arXiv 2023

  5. [5]

    Let’ s Wait Awhile: How Temporal Workload Shifting Can Reduce Carbon Emissions in the Cloud

    Wiesner, P., Behnke, I., Scheinert, D., Gontarska, K., Thamsen, L. "Let’ s Wait Awhile: How Temporal Workload Shifting Can Reduce Carbon Emissions in the Cloud." ACM/IFIP Middleware, 260–272, 2021. doi:10.1145/3464298.3493399

    arXiv 2021

  6. [7]

    Chasing Carbon: The Elusive Environmental Footprint of Computing

    Gupta, U., et al. "Chasing Carbon: The Elusive Environmental Footprint of Computing." IEEE HPCA, 2021. Extended version: IEEE Micro, 42(4), 37–47, 2022. doi:10.1109/MM.2022.3163226

    arXiv 2021

  7. [8]

    Ecovisor: A Virtual Energy System for Carbon-Efficient Applications

    Souza, A., et al. "Ecovisor: A Virtual Energy System for Carbon-Efficient Applications." ASPLOS, 252–265, 2023. doi:10.1145/3575693.3575709

    arXiv 2023

  8. [9]

    CarbonScaler: Leveraging Cloud Workload Elasticity for Optimizing Carbon-Efficiency

    Hanafy, W.A., et al. "CarbonScaler: Leveraging Cloud Workload Elasticity for Optimizing Carbon-Efficiency." Proc. ACM on Measurement and Analysis of Computing Systems (PO- MACS), 7(3), Article 57, 2023. doi:10.1145/3626788

    doi:10.1145/3626788 2023

Show all 31 references
  1. [10]

    Greening Geographical Load Balancing

    Liu, Z., Lin, M., Wierman, A., Low, S.H., Andrew, L.L.H. "Greening Geographical Load Balancing." ACM SIGMETRICS, 233–244, 2011. Extended version: IEEE/ACM Transactions on Networking, 23(2), 657–671, 2015. doi:10.1109/TNET.2014.2308295

    2011

  2. [11]

    Real-Time Carbon Accounting Method for the European Electricity Markets

    Tranberg, B., Corradi, O., Lajoie, B., Gibon, T., Staffell, I., Andresen, G.B. "Real-Time Carbon Accounting Method for the European Electricity Markets." Energy Strategy Reviews, 26, 100367, 2019. doi:10.1016/j.esr.2019.100367

    2019

  3. [12]

    Power Hungry Processing: Watts Driving the Cost of AI Deployment?

    Luccioni, A.S., Jernite, Y., Strubell, E. "Power Hungry Processing: Watts Driving the Cost of AI Deployment?" ACM FAccT, 85–99, 2024. doi:10.1145/3630106.3658542 16

    2024

  4. [13]

    Clover: Toward Sustainable AI with Carbon-Aware Machine Learning Inference Service

    Li, B., Samsi, S., Gadepally, V., Tiwari, D. "Clover: Toward Sustainable AI with Carbon-Aware Machine Learning Inference Service." ACM/IEEE SC, Article 20, 2023. doi:10.1145/3581784.3607034

    2023

  5. [14]

    Sprout: Green Generative AI with Carbon-Efficient LLM Inference

    Li, B., Jiang, Y., Gadepally, V., Tiwari, D. "Sprout: Green Generative AI with Carbon-Efficient LLM Inference." EMNLP, 21799–21813, 2024. doi:10.18653/v1/2024.emnlp-main.1215

    2024 doi

  6. [15]

    DynamoLLM: Designing LLM Inference Clusters for Performance and Energy Efficiency

    Stojkovic, J., Zhang, C., Goiri, Í., Torrellas, J., Choukse, E. "DynamoLLM: Designing LLM Inference Clusters for Performance and Energy Efficiency." IEEE HPCA, 1348–1362, 2025. Preprint doi:10.48550/arXiv.2408.00741

    2025 doi

  7. [16]

    Towards Environmentally Equitable AI via Geographical Load Balancing

    Li, P., Yang, J., Wierman, A., Ren, S. "Towards Environmentally Equitable AI via Geographical Load Balancing." ACM e-Energy, 291–307, 2024. doi:10.1145/3632775.3661938

    2024

  8. [17]

    Energy and Policy Considerations for Deep Learning in NLP

    Strubell, E., Ganesh, A., McCallum, A. "Energy and Policy Considerations for Deep Learning in NLP." ACL, 3645–3650, 2019. doi:10.18653/v1/P19-1355

    2019 doi

  9. [18]

    Measuring the Carbon Intensity of AI in Cloud Instances

    Dodge, J., et al. "Measuring the Carbon Intensity of AI in Cloud Instances." ACM FAccT, 1877–1894, 2022. doi:10.1145/3531146.3533234

    2022

  10. [19]

    Estimating Marginal CO2 Emissions Rates for National Electricity Systems

    Hawkes, A.D. "Estimating Marginal CO2 Emissions Rates for National Electricity Systems." Energy Policy, 38(10), 5977–5987, 2010. doi:10.1016/j.enpol.2010.05.053

    2010 doi

  11. [20]

    Carbon Emissions and Large Neural Network Training

    Patterson, D., et al. "Carbon Emissions and Large Neural Network Training." arXiv:2104.10350,

    arXiv

  12. [21]

    TowardstheSystematic Reporting of the Energy and Carbon Footprints of Machine Learning

    Henderson, P., Hu, J., Romoff, J., Brunskill, E., Jurafsky, D., Pineau, J."TowardstheSystematic Reporting of the Energy and Carbon Footprints of Machine Learning." Journal of Machine Learning Research, 21(1), 10039–10081, 2020. https://jmlr.org/papers/v21/ (open access)

    2020

  13. [22]

    Solyx AI Grid: Hardware-Telemetry-Aware Routing Across Geographically Distributed GPU Clusters

    Bernhard, A., Katla, N. (Solyx AI). "Solyx AI Grid: Hardware-Telemetry-Aware Routing Across Geographically Distributed GPU Clusters." arXiv:2606.15050, 2026. doi:10.48550/arXiv.2606.15050

    2026 doi

  14. [23]

    EcoServe: Designing Carbon-Aware AI Inference Systems

    Li, Y., et al. "EcoServe: Designing Carbon-Aware AI Inference Systems." arXiv:2502.05043,

    arXiv

  15. [24]

    Serving Models, Fast and Slow: Optimizing Heterogeneous LLM Inferencing Workloads at Scale (SAGESERVE)

    Jaiswal, S., et al. "Serving Models, Fast and Slow: Optimizing Heterogeneous LLM Inferencing Workloads at Scale (SAGESERVE)." arXiv:2502.14617, 2025. doi:10.48550/arXiv.2502.14617

    2025 doi

  16. [25]

    GAR: Carbon-Aware Routing for LLM Inference via Constrained Optimization

    Sheshanarayana, D., Pal, R.S., Sinha, M., Dasgupta, T. "GAR: Carbon-Aware Routing for LLM Inference via Constrained Optimization." arXiv:2605.11603, 2026. doi:10.48550/arXiv.2605.11603

    2026 doi arXiv

  17. [26]

    v3 API documentation

    WattTime. "v3 API documentation." WattTime, 2024–2026. https://docs.watttime.org/ (ac- cessed 2026-07)

    2024

  18. [27]

    North-America MOER model version 2026-03-01 release notes (expanded renewable- curtailment detection; zero-MOER values during modeled curtailment)

    WattTime. "North-America MOER model version 2026-03-01 release notes (expanded renewable- curtailment detection; zero-MOER values during modeled curtailment)." WattTime, 2026. https://watttime.org/data-science/release-notes/ (accessed 2026-07)

    2026

  19. [28]

    The GHG Protocol for Project Accounting

    World Resources Institute & World Business Council for Sustainable Development. "The GHG Protocol for Project Accounting." WRI/WBCSD, 2005. https://ghgprotocol.org/project- protocol (accessed 2026-07). 17

    2005

  20. [29]

    Guidelines for Quantifying GHG Reductions from Grid-Connected Electricity Projects

    World Resources Institute & World Business Council for Sustainable Development. "Guidelines for Quantifying GHG Reductions from Grid-Connected Electricity Projects." WRI/WBCSD,

  21. [2007]

    https://ghgprotocol.org/guidelines-grid-connected-electricity-projects (accessed 2026-07). 18

    2026

  22. [2021]

    doi:10.48550/arXiv.2104.10350

    doi arXiv

  23. [2025]

    doi:10.48550/arXiv.2502.05043

    doi arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.