AI energy consumption data: the numbers are bigger than you think
I compiled power consumption data for AI training and inference from every source I could find. A single GPT-4 query uses about 10x the energy of a Google search. At current growth rates, AI could consume 3% of US electricity by 2028.
I spent two weeks compiling energy consumption data for AI. Training costs, inference costs, data center power, cooling overhead. Every source I could find.
The numbers are larger than the public conversation suggests.
Energy per query
| Task | Energy per query | Equivalent to | |------|-----------------|--------------| | Google search | ~0.3 Wh | Turning on a lightbulb for 1 second | | ChatGPT query (GPT-4o) | ~3 Wh | Lightbulb for 10 seconds | | Complex reasoning (o3) | ~15 Wh | Lightbulb for 54 seconds | | Image generation (DALL-E 3) | ~4 Wh | Lightbulb for 14 seconds | | Full agent task (20 tool calls) | ~50 Wh | Running a laptop for 30 minutes |
Sources: IEA estimates, SemiAnalysis analysis, NVIDIA power specs, my calculations based on GPU utilization data.
A ChatGPT query uses roughly 10x the energy of a Google search. A complex reasoning query (o3, which uses 10-50K thinking tokens) uses about 50x.
An agent task with 20 tool calls? That's roughly 50 Wh. Running a laptop for half an hour to handle one request.
Training energy
| Model | Estimated training energy | Equivalent | |-------|--------------------------|-----------| | GPT-3 (2020) | ~1.3 GWh | 120 US homes for a year | | GPT-4 (2023) | ~50 GWh (est.) | 4,600 US homes for a year | | Llama 3.1 405B (2024) | ~20 GWh (est.) | 1,850 US homes for a year | | DeepSeek V3 (2024) | ~3 GWh (est.) | 280 US homes for a year | | Llama 4 405B (2025) | ~25 GWh (est.) | 2,300 US homes for a year |
Sources: Model papers (where disclosed), Epoch AI compute database, NVIDIA GPU power specs, training duration estimates. US average household consumption: ~10.8 MWh/year.
GPT-4's training reportedly used enough energy to power about 4,600 homes for a year. But training is a one-time cost. Inference is the ongoing expense.
DeepSeek V3 used roughly 3 GWh thanks to its efficient training approach. 16x less energy than GPT-4 for a model that benchmarks competitively. Training efficiency improvements directly reduce energy consumption.
Inference energy at scale
| Service | Est. daily queries | Energy per query | Daily energy | Annual energy | |---------|-------------------|-----------------|-------------|--------------| | ChatGPT | ~100M | ~3 Wh | ~300 MWh | ~110 GWh | | Google AI features | ~500M | ~1.5 Wh | ~750 MWh | ~274 GWh | | All AI APIs combined | ~2B (est.) | ~2 Wh (avg) | ~4,000 MWh | ~1,460 GWh |
Sources: SimilarWeb traffic estimates, Goldman Sachs research, IEA, my calculations.
Total AI inference globally probably consumes around 1,460 GWh per year, or about 1.46 TWh. For context, the US consumed about 4,000 TWh in 2024. So AI inference is currently about 0.04% of US electricity.
Small. But the growth rate is not small.
The growth trajectory
| Year | Est. AI electricity use (US) | % of US total | Growth rate | |------|---------------------------|--------------|-------------| | 2023 | ~3 TWh | 0.08% | Baseline | | 2024 | ~8 TWh | 0.20% | +167% | | 2025 | ~20 TWh (est.) | 0.50% | +150% | | 2026 | ~45 TWh (projected) | 1.1% | +125% | | 2028 | ~120 TWh (projected) | 3.0% | ~60% CAGR |
Sources: IEA projections, Goldman Sachs research, Microsoft and Google sustainability reports, my extrapolation.
At current growth rates, AI could consume 3% of US electricity by 2028. That's roughly the equivalent of adding another New York City's worth of power demand.
These projections assume current efficiency levels. If training and inference efficiency keep improving (they have been), the actual number could be lower.
The efficiency counter-argument
| Efficiency gain | Annual improvement | Impact | |----------------|-------------------|--------| | Hardware (H100 -> H200 -> B200) | ~50% per generation | Halves energy per token every 18 months | | Software (MoE, quantization) | ~30% per year | Reduces compute per query | | Training efficiency (DeepSeek effect) | ~60% per year | Less energy per trained model | | Total efficiency gain | ~70% per year | Significant but growth may outpace |
If efficiency improves 70% per year but AI usage grows 100%+ per year, total consumption still increases. The growth rate has been outpacing the efficiency gains.
My honest take
I'm not an energy alarmist. The 3% projection for 2028, even if accurate, is manageable with new power infrastructure. Natural gas plants, nuclear, and renewables can supply it.
But the speed of growth is unusual. Most industries don't go from 0.08% to 3% of national electricity in 5 years. The infrastructure buildout needed (new power plants, transmission lines, data center cooling) typically takes 5-10 years, and the demand curve is moving faster than the supply curve.
The data centers are already being built. Microsoft committed to nuclear power. Amazon signed renewable energy deals. The industry sees this coming.
Whether society is ready for a 3% electricity increase that primarily benefits technology companies is a question I'll leave to people smarter than me. My job is just to count the watts.
If you found this interesting, you might also like:
- 5 charts that explain why GPU prices went insane in 2021
- The training cost curve is doing something weird
- AI research papers published in 2021: a mid-year count
- My 2021 AI data roundup: the 10 numbers that mattered most
- I tracked AI image generation quality over 6 months. The improvement rate is scary.
Energy consumption estimates referenced the IEA Electricity Market Report.
-- dataku