Beyond Energy: When Sustainability Dimensions Reshape LLM Serving Decisions

Tianyao Shi; Xipeng Shen; Yi Ding

cs.CYSep 28, 2026

Beyond Energy: When Sustainability Dimensions Reshape LLM Serving Decisions

Authors: Tianyao Shi, Xipeng Shen, Yi Ding

Organizations: Elmore Family School of Electrical and Computer Engineering, Purdue University, USA

Abstract

Large language model (LLM) serving has environmental impacts across energy consumption, carbon emission, water consumption, and biodiversity loss. Yet these dimensions are largely evaluated in isolation, leaving it unclear when and how they lead to different optimization decisions. We present PRISM, a unified framework for characterizing and optimizing LLM serving across energy, carbon, water, and biodiversity impacts. Our analysis reveals a fundamental distinction: computing configurations determine energy consumption, whereas where and when LLM serving is deployed determine its carbon, water, and biodiversity impacts. Under a fixed deployment choice and operational-only accounting, all dimensions preserve the same energy-based configuration ranking. Deployment rankings can diverge across dimensions, while embodied impacts can break configuration invariance when they exceed a lifecycle crossover boundary. PRISM identifies these conditions, quantifies cross-dimensional regrets, and balances the four dimensions. In regional-routing experiments, PRISM reduces median worst-case regret by 50.2% relative to the strongest baseline.

Figures & tables

Dimension	Effective Operational Intensity $\alpha_{m}(r,t)$	Embodied Component	Output
Energy	$\mathrm{PUE}(r,t)$	—	kWh
Carbon	$\mathrm{PUE}(r,t)\cdot\mathrm{CI}(r,t)$	$C_{\mathrm{emb}}(x,w)$	kg CO 2 e
Water	$[\mathrm{WUE}(r,t)+\mathrm{PUE}(r,t)\cdot\mathrm{EWIF}(r,t)]\cdot\mathrm{WSF}(r)$	$W_{\mathrm{emb}}(x,w)$	m 3 world-eq
Biodiversity	$\mathrm{PUE}(r,t)\cdot\mathrm{BIF}(r,t)+\mathrm{WUE}(r,t)\cdot\mathrm{WSF}(r)\cdot\mathrm{CF}_{B,W}(r)$	$B_{\mathrm{emb}}(x,w)$	species $\cdot$ year

Table 1: Unified operational and embodied accounting across four sustainability dimensions.

Figure 2: Impact of model family and size on per-request energy, carbon, water, and biodiversity. Solid and hatched bars denote operational and embodied impacts, respectively; stars mark the minimum-impact configuration for each dimension.

Figure 3: Summary of computing configuration and workload effects under a fixed deployment choice. (a) Median and interquartile range of impact variation when varying model choice, GPU platform, GPU count, or non-agentic workload. (b) Cross-dimensional rank agreement for the same comparisons. Each comparison varies only the indicated category and uses a same functional unit.

Figure 4: Regional variation in per-request sustainability impacts for Qwen3-235B-A22B-Instruct on H100 (TP8). Stars mark the lowest-impact region for each dimension among the locations shown.

Figure 5: Lifecycle-induced configuration ranking disagreement for RepoBench in Norway with edit-similarity

\geq 44.65

. Stars mark the minimum-impact configuration for each dimension.

Figure 6: Left: directional cross-dimensional regret among the dimension-optimal regions in Figure 4 . Cell $(i,j)$ reports the additional impact under dimension $j$ when selecting the region optimized for dimension $i$ . Right: hourly directional cross-regret across selected regions during June 24–26, 2024, using the spatiotemporal operational impact intensities in Figure 20 .

Figure 7: Environmental trade-offs in offline regional routing. Regret measures the relative increase over the best feasible value for each impact. (a) Median regret for each dimension across 20 regional-capacity seeds; the symlog axis is linear below 10%. (b) Median worst-case regret, with whiskers showing the full range across seeds. Lower is better.

Appendix figures & tables36 assets

Supplementary material from the paper’s appendix.

Appendix

Notation	Meaning	Notation	Meaning
$E,C,W,B$	Energy, carbon, water, and biodiversity dimensions.	$\alpha_{m}(r,t)$	Operational impact intensity of dimension $m$ per unit IT energy.
$m$	Sustainability-dimension index, $m\in\{E,C,W,B\}$ .	$\mathrm{PUE}(r,t)$	Power Usage Effectiveness.
$w$	Workload context and functional unit.	$\mathrm{CI}(r,t)$	Grid carbon intensity.
$x$	Computing configuration, including the model, GPU type, and parallelism.	$\mathrm{WUE}(r,t)$	Direct datacenter water use per unit IT energy.
$r$	Deployment-region index.	$\mathrm{EWIF}(r,t)$	Electricity water-intensity factor.
$t$	Execution-time or routing-interval index.	$\mathrm{WSF}(\ell)$	Water stress factor at location $\ell$ ; $\mathrm{WSF}(r)$ is the datacenter-region value.

Appendix

Table 2: Notation used throughout the paper. Indices denote the corresponding workload, configuration, region, time, impact dimension, or request group unless stated otherwise.

Notation	Meaning	Notation	Meaning
$y_{ir}$	Fraction of request group $i$ assigned to region $r$ .	$q_{i},g_{i},t_{i}$	Request-equivalent count, GPU-seconds per request, and arrival hour of request group $i$ .
$z$	Maximum-regret variable in the offline routing Linear Programming.	$H$	Rolling-horizon forecast length in hours.
$y_{ir}^{(x)}$	Fraction of request group $i$ assigned to configuration $x$ in region $r$ .	$\mathcal{A}$	Set of agentic request groups.
$\tau_{x}$	Mean completion time of successful agentic tasks under configuration $x$ .	$\omega_{i}$	Weight of agentic request group $i$ in the completion-time objective.
$L(s),L^{*},$ $R_{L}(s)$	Mean completion time, its feasible optimum, and completion-time regret.	$I_{m}(s),I_{m}^{*},$ $R_{m}(s)$	Environmental impact, independent optimum, and regret for dimension $m$ in the agentic extension.
$\Omega_{\mathcal{A}}$	Total agentic-group weight, $\sum_{i\in\mathcal{A}}\omega_{i}$ .	$\mathcal{R}$	Candidate deployment-region set

Appendix

Table 3: Notation used throughout the paper (continued).

Dimension	Effective Operational Intensity $\alpha_{m}(r,t)$	Embodied Component	Output
Energy	$\mathrm{PUE}(r,t)$	—	kWh
Carbon	$\mathrm{PUE}(r,t)\cdot\mathrm{CI}(r,t)$	$C_{\mathrm{emb}}(x,w)$	kg CO 2 e
Water	$[\mathrm{WUE}(r,t)+\mathrm{PUE}(r,t)\cdot\mathrm{EWIF}(r,t)]\cdot\mathrm{WSF}(r)$	$W_{\mathrm{emb}}(x,w)$	m 3 world-eq
Biodiversity	$\mathrm{PUE}(r,t)\cdot\mathrm{BIF}(r,t)+\mathrm{WUE}(r,t)\cdot\mathrm{WSF}(r)\cdot\mathrm{CF}_{B,W}(r)$	$B_{\mathrm{emb}}(x,w)$	species $\cdot$ year

Appendix

Table 4: Unified accounting of energy, carbon, water, and biodiversity. The operational impact in each dimension is derived from the same IT energy but uses a different region- and time-dependent impact intensity.

Figure 8: Overview of PRISM . PRISM characterizes LLM serving across energy, carbon, water, and biodiversity, analyzes configuration and deployment rankings, identifies their disagreement, and selects sustainability-aware deployment decisions subject to service requirements.

Dataset	Description	Prompt Length			Response Length
		P50	P90	P95	P50	P90	P95
ShareGPT (2022)	Open-ended chatbot conversations sampled from real user-LLM interactions.	31	701	1,377	243	568	703


RepoBench ( Liu et al., 2024 )	Repository-level code completion with relevant in-repository context included in the prompt.	691	5,366	6,732	3	7	9
LongBench ( Bai et al., 2023 )	Long-document summarization of government reports.	8,432	17,316	21,185	655	876	934

Appendix

Table 5: Selected LLM serving workloads and their prompt/response length statistics encoded using the Qwen3 tokenizer. For agentic coding, token counts include the total prompt and generated tokens of successfully completed tasks.

Workload	TTFT	TPOT	Execution Time
ShareGPT	1000 ms	150 ms	—
RepoBench	5000 ms	75 ms	—
LongBench	$\min(45,11.2\times\frac{\mathrm{PromptLength}}{1000})$ s	150 ms	—
SWE-Bench Verified	—	—	30 min

Appendix

Table 6: Consolidated latency SLOs used, reported as p90 TTFT, p90 TPOT, and end-to-end execution time.

Family	Evaluated Models
Llama 3.1 ( Grattafiori et al., 2024 )	8B-Instruct, 70B-Instruct
GPT-OSS ( OpenAI, 2025 )	20B, 120B
Qwen3 ( Qwen Team, 2025 )	4B-Instruct / Thinking, 8B, 14B, 30B-A3B-Instruct / Thinking, 32B, 235B-A22B-Instruct / Thinking
Gemma 4 ( Google DeepMind, 2026 )	E4B-it, *26B-A4B-it, 31B-it

Appendix

Table 7: Evaluated model families and model variants.

Testbed	Accelerator	Accelerator Memory	CPU / VM	DRAM / Storage	Network
L40	4 $\times$ NVIDIA L40 ( NVIDIA, 2022b ) 300 W TDP	48 GB GDDR6/GPU 864 GB/s	AMD EPYC 7443 ( AMD, ) 24 cores, 200 W TDP 6 CPU cores/GPU	128 GB DRAM/GPU 1 TB HDD	Intel X550-AT2 ( Intel, ) 10 GbE, single-port operation 6.1 W typical
A100	4 $\times$ NVIDIA A100 SXM ( NVIDIA, 2020 ) 400 W TDP	40 GB HBM2/GPU 1,555 GB/s	AMD EPYC 7763 ( AMD, ) 64 cores, 280 W TDP 16 CPU cores/GPU	128 GB DRAM/GPU 1 TB SSD	NVIDIA ConnectX-6 ( NVIDIA, ) 100 Gb/s link 19.58 W typical
H100	8 $\times$ NVIDIA H100 SXM ( NVIDIA, 2022a ) up to 700 W TDP	80 GB HBM3/GPU 3.35 TB/s	Intel Xeon Platinum 8480+ ( Intel, ) 56 cores, 350 W TDP 7 CPU cores/GPU	128 GB DRAM/GPU 1 TB SSD	10 $\times$ NVIDIA ConnectX-7 ( NVIDIA, ) 400 Gb/s 24.9 W typical
Agent sandbox	–	–	GCP e2-highmem-16 ( Google Cloud, ) 16 vCPUs Intel Xeon E5-2699 v4 ( Intel, ) 22 cores / 44 threads, 145 W TDP	128 GB DRAM 100 GB SSD 1 TB HDD	GCP virtual network

Appendix

Table 8: Hardware specifications of the LLM-serving and agent-sandbox testbeds.

Model	ShareGPT	RepoBench		LongBench	SWE-Bench
		ES	EM	Summarization	Verified
Llama-3.1-8B-Instruct	13.8	43.5	8.2	36.6	0.20
Llama-3.1-70B-Instruct	17.8	56.0	24.8	37.7	0.20
GPT-OSS-20B	73.7	–	–	30.9	2.03
GPT-OSS-120B	112.8	–	–	28.8	4.46
Qwen3-4B-Instruct	50.5	44.6	9.2	30.9	3.65

Appendix

Table 9: Output-quality scores used to define quality requirements. All values are percentages. The best scores are in bold for each benchmark. ShareGPT reports the relative chat-quality score

q_{\mathrm{chat}}

from pairwise LLM judging; RepoBench reports edit similarity (ES) and exact match (EM); LongBench reports ROUGE-L F1 for long-output summarization tasks in the GovReport subset; SWE-Bench Verified reports the resolution rate. Missing entries indicate incompatible model–workload pairs: Reasoning models (GPT-OSS and Qwen3-Thinking variants) are excluded from code-completion tasks.

Location	Provider	Region code
Vienna, Austria	Azure	austriaeast
Brussels, Belgium	GCP	europe-west1
Hamina, Finland	GCP	europe-north1
Paris, France	AWS	eu-west-3
Frankfurt, Germany	GCP	europe-west3
Milan, Italy	AWS	eu-south-1

Appendix

Table 10: Supported modeled deployment locations and representative cloud-region mappings based on the official region documentation from Amazon Web Services () , Google Cloud () , and Microsoft () . Cities denote modeling proxies for direct WUE.

Figure 9: Impact of GPU type and tensor parallelism on per-request energy, carbon, water, and biodiversity for Gemma models.

Figure 10: ShareGPT impact characterization across GPU types and tensor-parallel settings for Llama models. L40 and A100 support up to TP4; Llama 3.1 70B configurations exceeding their memory capacity are omitted.

Figure 11: ShareGPT absolute impact characterization varying GPU type and tensor parallelism when focusing on GPT-OSS models.

Figure 12: ShareGPT absolute impact characterization varying GPU type and tensor parallelism when focusing on Qwen models.

Figure 13: Per-request energy, carbon, water, and biodiversity impact across model families and sizes for RepoBench code completion.

Figure 14: Per-request energy, carbon, water, and biodiversity impact across model families and sizes for LongBench long-output summarization.

Figure 15: Per-successful-task energy, carbon, water, and biodiversity impact across model families and sizes for SWE-bench Verified.

Figure 16: Impact variation across workloads for Gemma models on H100. Each model–workload pair uses its minimum-IT-energy TP setting. Non-agentic workloads are reported per request, while SWE-bench Verified is reported per successfully completed task.

Figure 17: Impact variation across workloads for Qwen models on H100. Each model–workload pair uses its most energy-efficient tensor-parallel configuration. Non-agentic workloads are reported per request, while SWE-bench Verified is reported per successfully completed task. Crosses mark excluded configurations.

Figure 18: Impact variation across workloads for GPT-OSS models on H100. Each model–workload pair uses its most energy-efficient tensor-parallel configuration. Non-agentic workloads are reported per request, while SWE-bench Verified is reported per successfully completed task.

Figure 19: Impact variation across workloads for Llama models on H100. Each model–workload pair uses its most energy-efficient tensor-parallel configuration. Non-agentic workloads are reported per request, while SWE-bench Verified is reported per successfully completed task.

Figure 20: Temporal variation in per-request carbon, water, and biodiversity impact across selected deployment regions under the same serving setting as Figure 4 . Left: monthly averages in 2024; right: hourly variation during June 24–26, 2024. Carbon, water, and biodiversity are reported in g CO 2 e/request, mL world-eq/request, and

10^{-13}

species

\cdot

yr/request, respectively.

Figure 21: Daily variation in per-request carbon, water, and biodiversity impact across selected deployment regions in March, June, September, and December 2024 under the same serving setting as Figure 4 .

Figure 22: Annual-average PUE values for the modeled regions in Table 10 , based on public disclosures from Amazon Web Services () , Google () , and Microsoft Azure () .

Figure 23: AWARE-2.0 water-scarcity factors for the modeled regions ( Seitfudem et al., 2025 ) .

Figure 24: Monthly CI, EWIF, WUE, and

\mathrm{BIF}

for selected modeled regions in 2024, together with the region-specific WSF used for water-scarcity adjustment. CI, EWIF, and

\mathrm{BIF}

follow the observed electricity-grid mix, WUE reflects cooling conditions, while WSF captures regional water scarcity.

Figure 25: Daily CI, EWIF, WUE, and

\mathrm{BIF}

for selected modeled regions in March, June, September, and December 2024.

Figure 26: Hourly CI, EWIF, WUE, and

\mathrm{BIF}

for selected modeled regions during June 24–26, 2024.

Figure	Low (%)	High (%)	Figure	Low (%)	High (%)
Figure 2	1.35	3.72	Figure 14	2.18	3.90
Figure 4	2.23	2.64	Figure 16	1.54	5.30
Figure 5	1.83	2.51	Figure 17	1.36	6.44
Figure 9	1.54	3.83	Figure 18	1.43	4.12
Figure 10	1.48	2.09	Figure 19	1.35	5.89
Figure 11	1.43	2.23	Figure 27	3.83	6.93

Appendix

Table 11: Ranges of embodied energy contribution to lifecycle energy impact across the evaluated settings in the characterization and disagreement analyses.

Figure 27: ShareGPT optimal configuration disagreement between dimensions

Figure 28: General model-ranking disagreement for ShareGPT across Norway, Switzerland, and France without imposing a quality constraint.

Figure 29: Lifecycle impact disagreement under controlled system choices. ① Qwen3-4B-Thinking on ShareGPT in Norway, comparing H100, A100, and L40 at TP1. ② Gemma 4 26B on ShareGPT in France, comparing H100 TP1 and TP4. Stars mark the preferred choice within each comparison.

Workload / configuration pair	Dimension	$\bm{\Delta E_{\mathrm{IT}}}$	$\bm{\Delta I_{m,\mathrm{emb}}/}$	$\bm{\alpha_{m}(r,t)}$	$\bm{\rho_{m}}$	Critical lifetime
( $x_{1}\!\rightarrow\!x_{2}$ )		(J/req.)	$\bm{\Delta E_{\mathrm{IT}}}$			(yr)
RepoBench ( Figure 5 ) Qwen3 4B-I $\rightarrow$ Qwen3 30B-I	Carbon	9.422	5.342	2.924	1.827	10.96
ShareGPT ( Figure 27 ) Qwen3 30B-I $\rightarrow$ Gemma 4 26B	Carbon	0.341	5.201	2.924	1.778	10.67
Figure 29 ① H100 TP1 $\rightarrow$ A100 TP1	Carbon	4.482	8.566	2.924	2.929	17.58
	Water	4.482	0.106	7.016	0.015	0.09
	Biodiversity	4.482	0.288	1.047	0.275	1.65

Appendix

Table 12: Numerical validation of the lifecycle crossover boundary. Bold

\rho_{m}>1

indicates a predicted lifecycle ranking reversal.

Region set	WaterWise	PRISM	Median reduction
Original 6	43.5% [28.8, 111.2]	37.2% [25.5, 53.5]	19.1%
All 12	157.8% [55.1, 212.3]	75.1% [41.4, 104.0]	50.2%
Concentrated 6	101.7% [57.7, 208.5]	68.1% [48.5, 107.4]	37.9%

Appendix

Table 13: Sensitivity of worst-case regret to geographic scope and capacity seed. Values report median [range] across 20 seeds.

	Oracle forecast		20% forecast error
Horizon	Max. regret	Gap	Max. regret	Gap
1 h	88.014%	0.697%	88.014%	0.697%
6 h	87.482%	0.088%	87.532%	0.146%
12 h	87.415%	0.012%	87.517%	0.129%
24 h	87.405%	$<0.001$ %	87.525%	0.138%
Offline PRISM	87.400%	–	–	–

Appendix

Table 14: Rolling-horizon PRISM under oracle and noisy environmental forecasts. Gap is the relative increase in maximum regret over full-day offline PRISM .

Figure 30: Joint computing-configuration and regional-routing optimization under the agent-heavy workload. (a) Maximum environmental regret versus agentic completion-time regret. (b) Configuration shares selected for the agentic workload.

Explore similar work

May 26, 2026q-bio.OT

BIRDS: Characterizing and Understanding Biodiversity Impact of Large Language Model Serving

Large language model (LLM) serving creates environmental impacts beyond carbon and water, including ecosystem damage through biodiversity-related pathways. We present BIRDS, a framework for Biodiversity Impact of Request-Driven LLM Serving. BIRDS defines request-level functional units, quantifies operational and embodied biodiversity impact, and introduces Quality-Normalized Biodiversity Impact (QNBI) to jointly analyze ecological impact and response quality. Across diverse workloads, models, GPUs, and regions, BIRDS reveals that biodiversity impact accumulates at scale and exposes actionable quality-aware serving tradeoffs.

Tianyao Shi, Yi Ding

Purdue University

Large Language Model Serving Biodiversity Monitoring Ecosystems Bird Animal Vocalizations

Ask PDF

Aug 13, 2026cs.DC

InFactPlanner: Planning Sustainable Geo-Distributed LLM Data Centers

The rapid growth of LLM inference is shifting sustainability concerns from one-time training to continuous serving, where infrastructure decisions shape energy use, carbon emissions, water consumption, and service quality. Yet operators often need to compare deployment alternatives before large-scale infrastructure is built, making direct measurement costly, slow, and sometimes infeasible. We present InFactPlanner, a trace-driven decision-support framework for what-if analysis of sustainable AI data center deployment for LLM inference across single and geo-distributed sites. InFactPlanner combines query traces, hardware-model profiles, candidate site configurations, PUE/WUE parameters, renewable generation models, and time-varying grid carbon intensity to estimate power, energy, carbon emissions, water use, latency, and server utilization. The framework abstracts low-level serving effects into configurable hardware-model profiles, enabling rapid comparison of site selection, capacity placement, hardware, model, renewable integration, and routing choices. We validate the energy accounting pipeline by reproducing reference LLM inference energy estimates with less than 10% deviation, evaluate scalability across multiple data centers and server counts, and demonstrate scenario-driven decision analyses for hardware selection, renewable placement, geographic deployment, and carbon-aware routing. Our results show that sustainability-optimal choices can differ from latency-optimal ones, and that the carbon value of deployment depends strongly on the local grid mix.

Nicoletta Tsiopani, Moysis Symeonides, George Pallis +1

Department of Computer Science University of Cyprus

Artificial Intelligence Infrastructure LLM Inference Optimization Planner Agent Planning

Ask PDF

May 13, 2026cs.DC

MARLIN: Multi-Agent Game-Theoretic Reinforcement Learning for Sustainable LLM Inference in Cloud Datacenters

Large Language Models (LLMs) have become increasingly prevalent in cloud-based platforms, propelled by the introduction of AI-based consumer and enterprise services. LLM inference requests in particular account for up to 90% of total LLM lifecycle energy use, dwarfing training energy costs. The rising volume of LLM inference requests is increasing environmental footprints, particularly carbon emissions and water consumption. To improve sustainability for LLM inference serving in cloud datacenter environments, we propose a novel multi-agent game-theoretic reinforcement learning framework called MARLIN to co-optimize time-to-first token (TTFT), carbon emissions, water usage, and energy costs associated with LLM inference. MARLIN demonstrates a reduction of at least 18% in TTFT, 33% in carbon emissions, 43% in water usage, and 11% in energy costs compared to state-of-the-art LLM inference management frameworks.

H. Moore, S. Qi, D. Milojicic +2

Colorado State University Fort Collins, CO, USA · Hewlett Packard Labs Milpitas, CA, USA

LLM Inference Optimization Data Centers Ai-Based Multi-Agent Reinforcement Learning Fish

Ask PDF