Training a single 8×A100 workload for 100 hours can emit anywhere from 7.74 kg to 272.00 kg of CO₂ depending solely on which cloud region you pick. That is a 97.2% reduction between worst and best case, and the worst case is not some exotic outlier - it comes from the same cloud providers everyone uses.
What Drives the 35-Fold Spread
Green AI Carbon Optimizer scores regions by combining three factors: local grid carbon intensity, renewable energy share, and data center Power Usage Effectiveness (PUE). Ablation tests reveal a trap: ranking by renewable share alone can actually select regions with higher total CO₂ than rankings that also include grid carbon intensity. In other words, a region with 100% renewable contracts might still sit on a dirty backup grid - renewables alone don't guarantee clean training.
The scoring model covers 100+ regions across the major cloud providers. For the reference workload (8 NVIDIA A100s, 100 hours), the optimizer recommends specific regions that cut emissions by an order of magnitude. This is not a theoretical exercise; the data is already available from public carbon and PUE sources.
Forecasting 2030: From 7 TWh to 1,436 TWh
The second contribution is a power-law forecasting pipeline that relates model parameter count to training energy using 26 anchor models. Combining that fit with scenario assumptions about model growth, hardware efficiency gains, and training frequency, the authors project a global AI energy demand range from 7 TWh to 1,436 TWh in 2030. The spread is enormous because it depends on decisions that engineers make today: deployment choices, model scaling discipline, and transparent energy reporting.
What keeps the lower bound plausible is aggressive model efficiency and limited training frequency, not magical carbon offset schemes. The upper bound assumes business as usual with rapid model scaling. Neither number includes inference energy, which amplifies the uncertainty further.
Why This Matters Right Now
Most ML engineers still treat cloud region selection as a latency or cost optimization problem. This paper gives a concrete method to fold carbon outcomes into that decision without waiting for industry regulation. The takeaway is simple: if you are about to launch a training run on a major cloud provider, check the carbon optimizer's scoring for your region choice first. The 97.2% reduction is sitting on the table, unmined.
Source: Green AI Carbon Optimizer: Carbon-Efficient Training Location Recommendation and Global AI Energy Demand Forecasting
Domain: arxiv.org
Comments load interactively on the live page.