Carbon Cost of Intelligence

The crisis

• AI inference is now a measurable contributor to global carbon emissions with no standard for reporting or comparing cost
• Data centers consumed ~200 TWh in 2023; AI workloads are the fastest-growing segment
• Organizations deploying AI systems have no per-query carbon cost visibility decisions that carry real emissions are made blind
• Climate targets require measurable reduction in compute carbon intensity, not just renewable pledges
• Without a cost framework, carbon optimization in AI is impossible

About this research

AI systems consume energy at query time, but this cost is invisible by default. Organizations running AI pipelines make model selection decisions based on capability and price rarely on carbon. This paper establishes a per-query carbon cost framework for frontier LLMs, benchmarked across 5 MMLU domains. The core finding is a 4.3× energy differential between model endpoints meaning that model selection is a carbon decision as much as a capability decision. A 12.1% cross-domain bias in energy allocation shows that domain context, not just model choice, affects carbon cost. The CCI metric is now integrated as a live service in the ClinicalSearch pipeline, providing real-time carbon cost per query alongside clinical evidence retrieval.

Research question

What is the per-query carbon cost differential across frontier LLMs on standardized benchmarks, and can a general-purpose CCI framework enable carbon-aware model selection in production AI pipelines?

Methodology

Empirical benchmarking across frontier LLMs on 5 MMLU domain subsets; per-query energy measurement using inference provider APIs and hardware-level monitoring; carbon intensity calculation using regional grid carbon factors; cross-domain bias analysis across subject areas; CCI metric derivation as a composable service for downstream pipeline integration; validation against GPT-4 as benchmark baseline.

Key findings

• 4.3× energy differential between most and least efficient frontier model endpoints
• 12.1% cross-domain bias energy cost varies by domain even within the same model
• CCI framework is composable: integrated as a live service in the ClinicalSearch pipeline
• Model selection is a carbon decision: equivalent capability endpoints can differ 4× in emissions

References

• IEA (2025) Data Centre and Data Transmission Networks energy consumption
• Patterson et al. (2021) Carbon emissions and large neural network training, Communications of the ACM
• Lottick et al. (2019) Energy usage reports: environmental awareness as part of algorithmic efficiency
• Schwartz et al. (2020) Green AI, Communications of the ACM

Related project