Berget.ai data shows identical AI inference emits 8.5x more CO2 on Texas grids than Swedish grids

Berget.ai measured per-request CO2 output for the same model inference across regions. Texas grid intensity reached 0.85 kgCO2e per kWh while Sweden averaged 0.10 kgCO2e per kWh, producing the reported 8.5x multiplier. The analysis isolates grid mix as the dominant variable after holding model size, hardware, and utilization constant.

Strubell et al. 2019 documented that training and inference emissions scale directly with regional electricity sources. IEA 2024 data confirm data centers already account for 2 percent of global electricity demand with rapid growth projected. Regional variance remains underreported in standard ML benchmarks that optimize only for latency and accuracy.

Texas coal and gas dominance drives higher marginal emissions during peak loads. Sweden relies on hydro and nuclear baseload with near-zero marginal carbon. Deployment decisions that ignore grid location therefore embed avoidable emissions into production systems.

Operators can reduce footprint by routing inference to low-carbon regions or shifting workloads to off-peak renewable surplus. Without explicit carbon-aware scheduling, aggregate AI emissions will track grid intensity rather than efficiency gains alone.