Archive · 5 Dec 2025
AI Reasoning Models Consume Far More Energy
Scientists found that AI models that can think and reason use a lot more electricity than simpler AI models.
This is because they create more text and need more computing power.
For example, one AI model used as much power as a lightbulb for an hour when it wasn't thinking, but when it was thinking, it used as much power as 6,164 lightbulbs for an hour.
This is a problem because it can strain power grids and make electricity more expensive.
It also makes it harder for companies to meet their climate goals.
The scientists hope that by understanding this, people will use the right AI model for the right task to save energy.
AI reasoning models use 100 times more power than non-reasoning models.
DeepSeek's R1 model used 308,186 watt hours with reasoning enabled, compared to 50 watt hours with it disabled.
The energy consumption of AI models has raised concerns about straining power grids and increasing energy costs.
The study evaluated 40 open, freely available AI models, including those from OpenAI, Google, and Microsoft.
The researchers hope the findings will encourage smarter AI usage, choosing the right model for the right task.
- Who
- AI developers and researchers, including Hugging Face and Salesforce
- What
- AI reasoning models consume significantly more energy than non-reasoning models
- When
- Study released on Thursday
- Where
- Global, with focus on data centers
- Why
- To understand the energy demands of AI reasoning models and promote smarter AI usage
- Energy Consumption
- 100 times more power for reasoning models
- Study Conducted By
- AI Energy Score project, led by Sasha Luccioni and Boris Gamazaychikov
- Models Evaluated
- 40 open, freely available AI models
- Example Model
- DeepSeek's R1 model
- Power Usage with Reasoning Off
- 50 watt hours
- Power Usage with Reasoning On
- 308,186 watt hours
- Industry Concern
- Straining power grids and raising energy costs
- Environmental Impact
- Complicates long-term climate objectives