Compliance
GPU Carbon Footprint Calculator: EU AI Act Compliance Guide
The EU AI Act (Article 13) requires GPAI model providers to report energy consumption during training. No production tool does this yet. Here's how to calculate GPU carbon footprint and automate compliance reporting.
What the EU AI Act Requires
The EU AI Act entered into force in August 2024, with GPAI provisions applying from August 2025. Article 13 mandates that providers of general-purpose AI models must:
- Document the energy consumption of the AI model during training
- Where known, document energy consumption during fine-tuning and inference
- Report using standardized measurement methodologies
- Make this information available to downstream deployers and national authorities
For systemic risk models (models trained with >10²⁵ FLOPs), additional obligations apply including ongoing energy efficiency reporting.
Who does this affect?
Any company that develops or fine-tunes a GPAI model and deploys it within the EU, or makes it available to EU-based deployers. This includes US companies serving EU customers.
The Carbon Footprint Formula
GPU carbon footprint is calculated as:
CO₂ (kg) = Energy (kWh) × Grid Intensity (gCO₂/kWh) / 1000
Where:
Energy (kWh) = (GPU_TDP_W × Utilization × GPU_Count × Hours) / 1000
Example: 64x H100 for 30 days at 92% utilization, US grid
= (700 × 0.92 × 64 × 720) / 1000 = 29,675 kWh
= 29,675 × 385 / 1000 = 11,425 kg CO₂
PUE (Power Usage Effectiveness) should also be factored in. Data center PUE typically ranges from 1.1 (hyperscale) to 1.8 (older facilities). Multiply the GPU energy by PUE to get total facility energy:
Total Energy = GPU Energy × PUE
29,675 kWh × 1.2 PUE = 35,610 kWh total
Grid Carbon Intensity by Region
The same training run produces vastly different emissions depending on where it runs. A model trained in France (nuclear grid, 56g/kWh) has 15x lower carbon footprint than one trained in West Virginia (coal grid, 860g/kWh).
| Region | gCO₂/kWh | Grid Mix | CO₂ for 10,000 kWh |
|---|---|---|---|
| France | 56 | Nuclear-heavy grid | 560 kg |
| Sweden | 41 | Hydro + nuclear | 410 kg |
| Germany | 385 | Coal + gas mix | 3,850 kg |
| UK | 230 | Wind + gas mix | 2,300 kg |
| US Average | 385 | Varies widely by state | 3,850 kg |
| US (California) | 220 | Solar + gas | 2,200 kg |
| US (West Virginia) | 860 | Coal-heavy | 8,600 kg |
| India | 700 | Coal-dominant | 7,000 kg |
| China | 550 | Coal + hydro | 5,500 kg |
| Iceland | 15 | Geothermal + hydro | 150 kg |
Real-World Training Energy Consumption
Published energy consumption data for major model training runs provides useful benchmarks:
| Model | GPUs | Duration | Energy (kWh) | CO₂ (tonnes) | Source |
|---|---|---|---|---|---|
| GPT-3 (175B) | ~1,000 V100s | ~34 days | 1,287,000 | 502 | Strubell et al. + Patterson et al. |
| Llama 2 70B | 2,048 A100s | ~25 days | 291,000 | 112 | Meta Technical Report |
| Llama 3 405B | 16,384 H100s | ~54 days | ~11,000,000 | ~4,200 | Meta estimate |
| BLOOM 176B | 384 A100s | ~118 days | 433,196 | 24.7 | BigScience (low-carbon grid) |
Note BLOOM's remarkably low carbon footprint (24.7 tonnes) despite being one of the larger models — trained on a low-carbon French grid. Where you train matters as much as how much you train.
Carbon-Aware Scheduling
If your training run can tolerate a few hours of scheduling flexibility, carbon-aware scheduling can reduce emissions by 20-40% by starting jobs when grid carbon intensity is lowest:
# Find the lowest-carbon window for a 4-hour job
nemulai carbon-schedule --duration 4h --zone US-CAL-CISO
# Output: Start at 2026-06-01 14:00 UTC (solar peak)
# Carbon intensity: 120 gCO₂/kWh vs. 320 gCO₂/kWh at midnight
# Savings: 62% lower carbon footprint
NemulAI integrates with the Electricity Maps API to provide real-time grid carbon intensity data for carbon-aware scheduling decisions.
Compliance Reporting with NemulAI
NemulAI automates EU AI Act energy reporting with three capabilities:
1. Automated Energy Metering
The agent measures real GPU power draw (not TDP estimates) with 5-second resolution across NVIDIA, AMD, Intel Gaudi, Intel Arc, Apple Silicon, and CPU-only platforms. Per-job attribution ensures energy is mapped to specific training runs, not just GPU-hours.
2. Carbon Footprint Calculation
Real-time grid carbon intensity from Electricity Maps, multiplied by measured energy consumption. Supports per-region calculation for distributed training across multiple data centers.
3. Compliance-Ready Reports
# Generate a compliance report with energy + carbon data
nemulai report --format json --with-carbon
# Output includes:
# - Total energy (kWh) per training run
# - CO₂ emissions (kg) with grid intensity source
# - Per-GPU breakdown
# - Timestamp range and methodology description
Beyond Compliance: The Business Case
Energy reporting isn't just a regulatory checkbox. Teams that measure energy consumption consistently find optimization opportunities:
- Right-size GPU allocation: Many training jobs run on 8 GPUs when 4 would suffice, doubling energy waste
- Power cap idle GPUs: An idle H100 draws 105W. Power-capping idle GPUs to 40% saves energy without affecting active workloads
- Schedule off-peak: Carbon-aware scheduling reduces emissions 20-40% with minimal impact on job completion time
- Choose efficient hardware: AMD MI300X delivers more TFLOPS per watt than H100 for memory-bound workloads
Getting Started
NemulAI is open-source (Apache-2.0) and free for up to 4 GPUs. Start measuring energy consumption today:
pip install nemulai
nemulai
Use our GPU Cost Calculator to estimate energy costs and carbon footprint for your specific GPU configuration.
EU AI Act compliance, automated
Measure real energy consumption, calculate carbon footprint, and generate compliance-ready reports.