How Much Energy Does AI Actually Use?

1. Where AI energy use actually happens

2. What the numbers look like in the real world

3. Why data centers are reshaping power markets

note

Data centers are changing electricity demand

A modern AI data center can require 50 to 100 megawatts or more. At the top end, that is comparable to the electricity demand of a small city.

This is why AI companies are signing deals for:

• Utility-scale solar and wind
• Battery storage
• Grid upgrades and transmission
• Nuclear power, including existing plants and restarts

Why utilities care

New data centers create:

• Large, predictable loads
• Demand for 24/7 reliability
• Pressure to expand generation and cooling infrastructure

That can speed up investment in clean power. It can also stress local grids if growth is too fast.

diagram

illustration

4. How engineers cut AI energy use

# Simple illustration of quantization savings
# 32-bit float uses 4 bytes per number
# 8-bit integer uses 1 byte per number

parameters = 7_000_000_000
bytes_fp32 = parameters * 4
bytes_int8 = parameters * 1

print(f'FP32 storage: {bytes_fp32 / 1e9:.1f} GB')
print(f'INT8 storage: {bytes_int8 / 1e9:.1f} GB')
print(f'Size reduction: {bytes_fp32 / bytes_int8:.0f}x')

5. The hard tradeoff: ambition versus footprint

Transcript

Welcome to Slate. Today we're looking at How Much Energy Does AI Actually Use?. We'll cover The true energy cost of training and running AI models, How data centers are driving nuclear and renewable investment, Efficiency techniques: distillation, sparsity, edge computing, and The tension between AI ambition and environmental cost. Let's get into it.

A model does not use electricity in one dramatic burst. It uses power in two very different phases. First comes training. That is the expensive part. Thousands of graphics processing units, or G-P-U-s, run for days or weeks. Then comes inference. That is the steady drip of energy every time a user asks a question, generates an image, or gets a recommendation. The diagram shows the split. Training is like building a highway. Inference is the traffic that uses it every day. The scale can be large. Researchers at the University of Massachusetts Amherst estimated in 2019 that training a big natural-language model could emit about 284 metric tons of carbon dioxide equivalent, more than five cars over their full lifetimes. A 2023 study from the University of California, Berkeley found that training power can vary wildly depending on model size, hardware, and data center efficiency. That range matters. A model trained on modern chips in a well-run facility can use far less energy than an older setup doing the same work. So when people say, “AI uses a lot of energy,” the real question is: which stage, which model, which hardware, and which data center?

The visual on the canvas helps here because the numbers are easier to feel than to imagine. A single large training run can indeed rival the annual electricity use of dozens or even around 100 homes, depending on the model and the grid mix. But that comparison is not a law of nature. It is a snapshot. For example, in 2023, researchers and industry analysts estimated that training frontier models could require hundreds of megawatt-hours to several gigawatt-hours of electricity. One widely cited estimate from the International Energy Agency says data centers, cryptocurrencies, and AI together used around 460 terawatt-hours in 2022, and data center demand could more than double by 2026. That does not mean all of that is AI. It means AI is entering an already power-hungry system. Here is the important distinction. A model trained once and used by a few thousand people is not the same as a model embedded in search, office software, and phones. The first has a big up-front cost. The second can accumulate a much larger lifetime footprint. Think of it like buying a sports car versus driving a taxi all day. The car is the same. The mileage changes everything.

Here the story moves from chips to infrastructure. AI does not just consume electricity. It changes where utilities, governments, and companies build new power plants and transmission lines. A hyperscale data center can draw tens to hundreds of megawatts. That is enough to matter to a regional grid. In the United States, utilities and tech companies are increasingly signing long-term contracts for wind, solar, batteries, and nuclear power. Microsoft, Amazon, Google, and Meta have all announced large clean-energy purchases. Nuclear is back in the conversation because it can provide steady power around the clock, unlike solar and wind alone. In 2024, Microsoft signed a deal tied to restarting the Three Mile Island Unit 1 site in Pennsylvania. That is a striking example of how AI demand is influencing old energy assets. The tension is simple. AI firms want reliable power now. Grid planners want low-carbon power that can scale quickly. Those goals do not always line up. The map of a data center is also a map of the energy system around it.

The good news is that AI energy use is not fixed. Engineers have real tools to reduce it. The first is distillation. A large model teaches a smaller one. You keep much of the behavior, but the student model needs fewer parameters and less compute. The second is sparsity. Instead of activating every part of a model for every token, only some parts turn on. That is like lighting only the rooms you are using. Quantization is another major technique. It stores numbers with fewer bits, such as 8-bit or 4-bit precision, which reduces memory traffic and often speeds inference. On modern hardware, that can cut power use and cost at the same time. Edge computing helps too. If a small model runs on a phone or laptop, the request may never need a round trip to a distant data center. These methods are not free. Smaller models can be less accurate. Sparse systems can be harder to train and debug. Quantization can hurt quality if pushed too far. The engineering job is to find the sweet spot where the energy saved is worth the performance tradeoff.

AI is not going away, so the real question is how to grow it responsibly. The cleanest path is not to stop computing. It is to use less compute for the same value. That means better models, better chips, better scheduling, and cleaner grids. There is a policy angle too. If a company trains a massive model on a coal-heavy grid, the carbon cost is much higher than the same training run on a low-carbon grid. That is why location matters. So does time of day. Some data centers now shift flexible workloads to periods when renewable power is abundant. The final lesson is balance. AI can help optimize power grids, discover materials for batteries, and reduce waste in logistics. But those benefits do not cancel the emissions from the systems themselves. The question is not whether AI has a footprint. It does. The question is whether we measure it honestly and shrink it aggressively. That is what sustainable computing looks like in practice.