The energy consumption of large AI models is becoming an issue that cannot be ignored.
Training Costs
The electricity consumption of large model training is staggering. Taking a GPT-4-level model as a reference:
- A single training run consumes an amount of electricity equivalent to hundreds of U.S. households' annual usage
- AI data center construction by Microsoft, Google, and Meta is pushing up the global share of data center energy consumption
What is even more alarming is the inference side. Training is a one-time event, but inference is continuous. With hundreds of millions of AI calls worldwide every day, cumulative inference energy consumption may have already surpassed that of training.
Company Responses
Microsoft has signed a large number of renewable energy purchase agreements and is even exploring small nuclear reactors to power its data centers.
Google has committed to powering 100% of its AI-related operations with carbon-free energy, but in practice there is a significant gap between "carbon neutrality" and "truly carbon-free."
DeepSeek's approach offers a different perspective — instead of brute-force training with more GPUs, use smarter methods (MoE, FP8, more efficient architectures) to reduce the computational load itself. The practice of training a top-tier model for $5.5 million is also the most environmentally friendly from an energy perspective.
The Contradiction
There is a fundamental contradiction between AI's energy problem and the pace of AI development:
- The larger and more capable the model, the higher the energy consumption
- The deeper users' reliance on AI, the more inference requests
- The pace of data center construction cannot keep up with the growth in demand for AI computing power
Some researchers predict that by 2030, AI-related electricity consumption could account for 3-5% of global total electricity use. While this figure may seem small, the absolute amount is very significant.
Efficiency optimization (quantization, distillation, MoE, etc.) is currently the most realistic solution. Relying entirely on renewable energy to cover AI energy consumption is not realistic in the short to medium term.
Sources: IEA Data Center Energy Consumption Report, CocoLoop, and various companies' sustainability reports