AI energy demand
This is the amount of electricity consumed by training and daily use of artificial intelligence, mainly through machines operating in data centers.
In plain terms
When you type a question into a chat AI, the answer is not generated on your phone but in a remote data center where thousands of high-performance computers work. These consume electricity, and their cooling also costs electricity, often water. AI energy demand is the name for this total consumption.
It is like tap water: you just turn on the tap; you do not see the pumps, the water tower, or the treatment plant. A single glass of water costs almost nothing, but the consumption of an entire city requires serious infrastructure. With AI, a single question is a tiny item, but billions of questions a day are power-plant scale.
The topic made headlines because data centers built for AI are growing faster than the electrical grid can expand. Thus, today, not only the number of chips but also available power determines where and how quickly AI can develop.
How it works
- Training: the model is trained for weeks or months on thousands of specialized chips (GPUs), and on tens of thousands for the largest models. This is a one-off but huge energy draw per model, and it is repeated for every new model version or experiment.
- Usage (inference): every question, image generation, or agent task starts a new calculation. Individually small, but due to the large number of users, this accounts for the majority of consumption over time.
- Cooling and power delivery: chips convert almost all consumed electricity into heat, which must be dissipated. Cooling, power conversion, and backup systems consume additional energy.
- Network connectivity: a large AI data center can consume as much electricity as a medium-sized city, requiring dedicated high-voltage lines, substations, and often its own power plant agreement.
- Efficiency improvement: newer chips and models solve the same task with less energy, but cheaper usage leads to more consumption, so total consumption has continued to grow so far.
An everyday example
You ask an AI to summarize the gist of a long contract and then to create an image for it. According to the providers' own figures, an average text response uses roughly as much electricity as ten seconds of watching TV; processing a long document uses more, and an image or a short video far more. None of this shows up on your electricity bill, because the consumption occurs in the data center.
Who is affected
A common misconception
Limits and risks
- The exact figures are uncertain: companies share little data, and estimates vary by multiples depending on the method used.
- Not all AI tasks are equal: video generation and long-running agents consume many times more resources than a short text response.
- A 'green' or 'carbon-neutral' data center often means only an accounting solution, not that the machines run on renewable energy every hour.
- Alongside electricity, the environmental impact of cooling water usage and chip manufacturing matters; consumption data alone says nothing about this.
Similar, but not the same
- Data center energy consumption – Data centers serve the entire internet (video, storage, email), while AI is only one part, albeit the fastest-growing.
- AI carbon footprint – The carbon footprint measures the amount of greenhouse gases emitted, which depends on which power plant supplies the electricity; the same consumption can be clean or polluting.
- computing capacity – Computing power shows how many operations machines can perform, while energy demand shows how much electricity is needed for this.
Related concepts
AI infrastructure, Environmental impact of AI, Sustainable AI
Training energy demand, Inference energy consumption, Data center cooling, AI water consumption
Data center, GPU, computing capacity, AI carbon footprint, Green AI
A brief history
The issue was brought to professional awareness in 2019 by a study by Emma Strubell and colleagues, which quantified the energy and carbon dioxide costs of training large language models. After the launch of ChatGPT in 2022, the topic moved from research debates to energy policy and daily news, and in 2025 the International Energy Agency (IEA) dedicated a special report to it.
Where things stand
By the end of 2026, power supply has become a major bottleneck for AI development: large companies are signing long-term power plant contracts, and in many countries, grid connection is the primary obstacle. According to the IEA's 2025 estimate, data center electricity consumption could more than double from approximately 415 TWh in 2024 to roughly 945 TWh by 2030. This is a forecast, and measurement and transparency remain inadequate.
In the news
- PJM grid moves to tackle AI power costs to avoid capacity shortfallsNov 4, 2026 · Ipar
- 81% of energy executives say AI power demand is underestimatedOct 22, 2026 · Ipar
- Emera to buy Canadian Utilities for $14.3 billion in deal tied to AI power demandOct 6, 2026 · Ipar
- India's data-centre expansion tests energy, land and water, and renewable capacitySep 28, 2026 · Ipar
- World Energy Council: AI power demand and grid constraints are reshaping the global energy systemSep 21, 2026 · Ipar
- Regional grid strains: Northwest, Iowa and California wildfire riskSep 14, 2026 · Ipar
- Zach Dell bets on battery fleet and new generation models to ease AI power crunchSep 14, 2026 · Ipar
- US to produce more gas for AI power demand as data centres weigh on-site generationAug 31, 2026 · Ipar
Also known as
AI energy demand · AI energy demand · AI energy consumption · AI electricity consumption · AI's energy appetite · AI energy demand · AI energy consumption · AI power demand
Sources
- IEA: Energy and AI (2025)
- Strubell, Ganesh, McCallum: Energy and Policy Considerations for Deep Learning in NLP (2019)
- Wikipedia: Environmental impact of artificial intelligence
This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.