Skip to content
,

How Much Energy Does AI Use? Data Centers, ChatGPT Prompts and GPUs Explained

How much electricity does AI really use? Data center energy demand, the energy of a single AI prompt, the cost of training large models, why GPUs draw so much power, water use, and what is being done to make AI more efficient.

5 min read

How Much Energy Does AI Use? — cover image from TensorCore.si

Short answer

AI's energy use is growing fast but is often misunderstood. The International Energy Agency (IEA) estimates data centers used about 415 TWh of electricity in 2024 (around 1.5% of global electricity) and projects this will more than double to about 945 TWh by 2030, with AI the main driver. A single chatbot text prompt is small: Google measured the median Gemini text prompt at about 0.24 watt-hours, roughly like running a TV for nine seconds. The big totals come from billions of prompts, longer reasoning and video generation, and from training ever-larger models on tens of thousands of GPUs.

The big picture: data centers

AI data centers are buildings full of GPU servers, networking, storage and cooling. According to the IEA's 2025 *Energy and AI* report:

  • Data centers used roughly 415 TWh in 2024, about 1.5% of the world's electricity.
  • Demand is projected to reach about 945 TWh by 2030, slightly more than Japan uses today.
  • AI-focused, accelerated servers are the fastest-growing part of that demand.
  • The United States and China account for most of the growth.

Data centers are still a modest share of total global electricity, but they are concentrated: in some regions they already use a large share of local supply, which strains grids and drives new power plant and transmission projects.

How much energy does one AI prompt use?

ActivityApproximate energy
Median Gemini text prompt (Google, 2025)0.24 Wh
Typical ChatGPT query (OpenAI CEO estimate, 2025)~0.34 Wh
Long reasoning answerSeveral to tens of times a short reply
AI image generationUsually more than a text reply
AI video generationMuch more, scaling with length and resolution
10 W LED bulb for one hour10 Wh

Google's measurement included the AI chips, host servers, idle capacity and data center overhead, and also reported about 0.26 ml of water per median prompt. Estimates vary a lot by model size, answer length and hardware, so treat single numbers as rough guides.

Two trends push per-prompt energy in opposite directions:

  • Efficiency gains: newer GPUs, low-precision formats like FP8 and FP4, better software and mixture-of-experts models make each token cheaper. Google reported large year-on-year reductions in energy per prompt.
  • Heavier use: reasoning models generate thousands of hidden tokens, AI agents run many steps, and video generation is far more compute-intensive than text.

How much energy does training use?

Training a frontier model runs thousands of GPUs non-stop for weeks or months. A widely cited 2021 study estimated that training GPT-3 used about 1,287 MWh, roughly what 120 average US homes use in a year. Today's largest models are trained with far more compute on clusters of tens of thousands of GPUs, so their training runs use much more, though developers rarely publish exact figures. See AI training vs inference and LLM training cost for how the compute adds up.

Over a popular model's lifetime, inference (answering users) usually uses more total energy than training, because it runs billions of times.

Why AI hardware draws so much power

HardwareTypical maximum power
RTX 5090 gaming/AI GPU575 W
NVIDIA H100 SXM700 W
NVIDIA B200~1,000 W
8-GPU H100 server~10 kW
GB200 NVL72 rack (72 GPUs)~120 kW or more
Traditional server rack~5–15 kW

Modern AI chips pack tens of billions of transistors that switch billions of times per second, and moving data between memory and compute costs energy too. That is why memory bandwidth and stacked HBM close to the chip matter for efficiency. Dense racks like the GB200 NVL72 need liquid cooling because air cannot remove that much heat.

Water use

Data centers use water mainly for cooling, either directly (evaporative cooling) or indirectly through the power plants that supply electricity. Usage depends heavily on climate and cooling design. Closed-loop liquid cooling and siting in cooler climates reduce water consumption, and many operators report water use per unit of compute.

How AI is getting more efficient

  • Better chips: each GPU generation delivers far more AI performance per watt. See Hopper vs Blackwell.
  • Lower precision: NVFP4 and FP8 do more work per joule than FP16.
  • Smarter models: small language models, distillation and MoE give similar quality with less compute.
  • Software: techniques like FlashAttention, batching and KV cache optimisation squeeze more tokens from each GPU.
  • Specialised chips: AI accelerators such as TPUs and inference ASICs can be more efficient for specific workloads.
  • On-device AI: running small models on phones and AI PCs avoids data center trips for simple tasks.
  • Clean power: tech companies are signing large deals for solar, wind, nuclear and geothermal power.

What you can do

  • Use smaller or faster models for simple tasks, and reasoning modes only when needed.
  • Avoid regenerating answers or images many times without reason.
  • Run small models locally when it makes sense; see how to run an LLM locally.

Frequently asked questions

Does one ChatGPT question use a bottle of water?

No. That claim came from early, broad estimates. Google's 2025 measurement put a median Gemini text prompt at about 0.26 ml of water, a few drops. Totals become significant at the scale of billions of prompts.

Is AI worse for energy than Bitcoin mining?

They are different kinds of demand. Estimates for Bitcoin are in a similar range to today's AI share of data center use, but AI demand is growing faster and also delivers useful services. Both depend heavily on how clean the electricity is.

Will AI cause power shortages?

In specific regions with many new data centers, grid capacity is a real constraint, and utilities are investing heavily. Globally, the IEA expects data centers to remain a relatively small share of electricity growth compared with industry, air conditioning and electric vehicles.

How much power does a home AI PC use?

A desktop with a high-end GPU can draw 500 to 800 W under full AI load, but local AI usually runs in short bursts. Smaller models on laptops and Macs use far less.

Sources and further reading

Think you understand it now? Prove it.

Solve real Tensor Core math in the TensorCore.si Challenge and earn a certificate with your name.

Take the challenge

Leave a Reply

Your email address will not be published. Required fields are marked *