he global AI industry consumed an estimated 85 terawatt-hours of electricity in 2025, roughly equivalent to Belgium's entire annual energy consumption. Every GPU hour you rent, every model you train, every inference request you serve contributes to this footprint. By 2028, AI energy consumption is projected to triple.
But energy efficiency is not just an environmental concern. It is a cost driver. Data center electricity bills flow through to GPU rental prices. The most efficient facilities offer the same hardware at lower prices because their operating costs are lower. This is a key advantage of io.net's decentralized model: aggregating capacity from efficient data centers worldwide passes those savings to users.
Power Consumption by GPU Generation
| GPU | TDP (Watts) | Typical (Watts) | Annual Energy (kWh) | Annual Cost ($0.10/kWh) |
|---|---|---|---|---|
| A100 80GB SXM | 400W | 350W | 3,066 | $307 |
| H100 80GB SXM | 700W | 600W | 5,256 | $526 |
| B200 | 1,000W | 850W | 7,446 | $745 |
| GB300 | 1,400W | 1,200W | 10,512 | $1,051 |
A single H100 running 24/7 uses as much electricity as an average US household. A 1,000-GPU cluster draws as much as a small town.
Where Energy Goes in an AI Data Center
| Component | Share of Total Power | Optimization Potential |
|---|---|---|
| GPU compute | 50-65% | Use newer, efficient GPUs |
| Cooling | 15-35% | Liquid cooling (biggest lever) |
| Networking | 5-8% | Efficient switches |
| Storage | 3-5% | NVMe vs spinning disk |
| Power conversion losses | 5-10% | High-efficiency PSUs |
PUE: The Key Metric
Power Usage Effectiveness = Total Facility Power / IT Equipment Power
| PUE | Rating | Cooling Method | Energy Overhead |
|---|---|---|---|
| 1.60+ | Poor | Outdated air | 60%+ waste |
| 1.30-1.60 | Average | Standard air | 30-60% waste |
| 1.10-1.30 | Good | Hot/cold aisle, free cooling | 10-30% waste |
| 1.05-1.10 | Excellent | Direct liquid cooling | 5-10% waste |
| 1.01-1.05 | World-class | Immersion cooling | 1-5% waste |
Strategies for Energy-Efficient AI
Strategy 1: Use Newer GPUs
Each GPU generation improves performance per watt by 50-80%:
| GPU | TFLOPS (BF16) | Power | TFLOPS/Watt |
|---|---|---|---|
| A100 | 624 | 400W | 1.56 |
| H100 | 1,979 | 700W | 2.83 (1.8x better) |
| B200 | ~4,500 | 1,000W | 4.50 (2.9x better) |
Counterintuitively, newer higher-wattage GPUs use less total energy per workload because they finish faster.
Strategy 2: Optimize Utilization
An idle H100 draws 200-300W. Push utilization above 70% to minimize energy waste per useful computation.
Strategy 3: Use Efficient Precision
| Precision | Energy per Op (relative) | Throughput |
|---|---|---|
| FP32 | 1.0x | 1.0x |
| BF16 | 0.5x | 2.0x |
| FP8 | 0.25x | 4.0x |
| INT4 | 0.12x | 8.0x |
INT4 inference uses ~85% less energy per token than FP32.
Strategy 4: Choose Efficient Regions
| Region | Grid Carbon (gCO2/kWh) | Electricity Cost | Climate |
|---|---|---|---|
| Nordics | 20-50 | $0.05-0.08/kWh | Cold (free cooling) |
| Quebec | 10-30 | $0.04-0.07/kWh | Cold, hydro |
| Pacific Northwest | 80-150 | $0.06-0.09/kWh | Moderate |
| US Midwest | 400-600 | $0.07-0.10/kWh | Variable |
| Singapore | 350-450 | $0.12-0.18/kWh | Hot (heavy cooling) |
Strategy 5: Liquid Cooling
Moving from air cooling (PUE 1.4) to direct liquid cooling (PUE 1.1) saves 25-30% of total facility energy. For a 10 MW facility, that is $2.5-3M annually.
Try io.net GPU Cloud
io.net offers 40-60% savings over hyperscalers on identical GPU hardware. H100 80GB at $2.49/hr with NVLink connectivity.
Measuring Your AI Carbon Footprint
Calculation Formula
CO2 = GPU_hours x GPU_power_kW x PUE x Grid_carbon_intensity_kg/kWh
Example: 64x H100, 14-day training run - GPU energy: 64 x 0.7 kW x 336 hours = 15,053 kWh - With PUE 1.2: 18,064 kWh total - US average grid (0.4 kg CO2/kWh): 7.2 metric tons - Nordic grid (0.03 kg CO2/kWh): 0.54 metric tons
Reporting Frameworks
- Scope 2 emissions: Direct electricity use
- Science Based Targets (SBTi): Requires AI compute reduction targets
- EU CSRD: Mandatory reporting for EU companies in 2026
- GHG Protocol: Standard framework for carbon accounting
The Economics of Efficiency
Energy Cost as Fraction of GPU Price
| GPU | io.net Price/hr | Energy Cost/hr (PUE 1.2) | Energy % |
|---|---|---|---|
| A100 80GB | $1.89 | $0.048 | 2.5% |
| H100 80GB | $2.49 | $0.084 | 3.4% |
Energy is a small fraction of rental price but a huge fraction of data center operating costs. Efficient facilities pass savings through to customers.
Future Trends (2026-2030)
- Liquid cooling universal by 2028: All new AI facilities
- Waste heat recapture: Selling GPU heat for district heating
- Dedicated renewables: Providers building solar/wind for data centers
- Carbon-aware scheduling: Workload routing based on grid carbon intensity
- Nuclear for AI: Multiple hyperscalers exploring nuclear partnerships
- Efficiency mandates: EU regulations setting minimum PUE requirements
Frequently Asked Questions
How much energy does AI training use?
GPT-4-class training: 50-100 GWh. Fine-tuning 70B for a week on 64 GPUs: 7-10 MWh. Individual inference requests: fractions of a Wh each.
Does io.net reduce my carbon footprint?
Depends on specific facilities. io.net aggregates from modern, efficient data centers worldwide. The decentralized model naturally selects for cost-efficient (and therefore energy-efficient) facilities.
What is the most impactful efficiency improvement?
Use newer GPUs. An H100 completes work 3-4x faster than A100, using less total energy despite higher per-GPU wattage.
Does quantization save energy?
Yes. INT4 inference uses ~85% less energy per token than FP32. BF16 training uses 50% less than FP32.
What PUE should I target?
Below 1.2 is good, below 1.1 is excellent. Avoid facilities above 1.4.
How do I track energy use on io.net?
Track GPU hours and types. Multiply by TDP and PUE for energy estimates. Tools like CodeCarbon automate this for training runs.
Conclusion
Energy efficiency is simultaneously an environmental imperative, a cost lever, and a regulatory requirement. The most energy-efficient approaches are also the most cost-effective: newer GPUs, higher utilization, lower precision, efficient cooling.
io.net's decentralized network includes facilities across the efficiency spectrum, with the most efficient offering the lowest prices. By optimizing your workloads and choosing efficient infrastructure, you reduce both your cloud bill and your carbon footprint.
Deploy energy-efficient AI on io.net. Get started with H100 GPUs at $2.49/hr.