AI training and deployment consume substantial amounts of electricity, yet carbon outcomes remain weakly integrated into routine model development decisions, according to a new paper on arXiv titled "Green AI Carbon Optimizer: Carbon-Efficient Training Location Recommendation and Global AI Energy Demand Forecasting." The paper, authored by Chen, Gao, and Zou, presents a system with two primary contributions: a carbon-aware cloud region recommendation method for training workloads, and a power law forecasting pipeline for global AI energy demand.
Cloud Region Recommendation for Lower Carbon Training
The location recommendation component combines three factors into a unified scoring model: regional grid carbon intensity, renewable energy share, and data center Power Usage Effectiveness (PUE). The model covers over 100 regions from major cloud providers. For a reference workload of 8 NVIDIA A100 GPUs running for 100 hours, estimated emissions across sampled regions range from 7.74 kg to 272.00 kg CO2. Selecting the best region instead of the worst corresponds to a 97.2% reduction relative to the worst case. An ablation study shows that ranking by renewable share alone can select regions with higher CO2 emissions than rankings that include grid carbon intensity, indicating the importance of a multi-factor approach.
| Factor | Impact on Recommendation |
|---|---|
| Grid carbon intensity | Directly penalizes fossil-heavy grids |
| Renewable share | Favors regions with high renewables |
| Data center PUE | Accounts for facility efficiency |
| Combined score | Selects regions with lowest total carbon footprint |
Forecasting Global AI Energy Demand
The forecasting contribution fits a power law relation between model parameter count and training energy using 26 anchor models. This fit is combined with scenario assumptions on model growth, hardware efficiency, and training frequency. The model also evaluates sensitivity to inference ratio and ecosystem scaling. Across multiple scenarios, projected 2030 demand ranges from 7 TWh to 1,436 TWh under the stated assumptions, highlighting the importance of deployment choices, model scaling discipline, and transparent energy reporting.
Implications for Enterprise Technology Decision-Makers
For CTOs and technology procurement leaders, the findings underscore that the choice of cloud region for AI training can dramatically affect carbon emissions—by as much as 97.2% in the worst-case comparison. The paper demonstrates that simplistic metrics like renewable share alone are insufficient; a holistic score including grid carbon intensity and PUE is necessary. The wide range in 2030 energy demand projections (7 TWh to 1,436 TWh) suggests that policy and industry decisions on model scaling and hardware efficiency will have outsized impact. Transparent energy reporting and disciplined model scaling are critical to managing future AI energy consumption and carbon footprint. The Green AI Carbon Optimizer provides a framework for integrating carbon outcomes into everyday model development choices, potentially guiding investments in data center locations and training scheduling.X
