Why does AI use so much electricity? Because large AI models run on thousands of specialised chips packed into data centres, and those chips draw a lot of power both when a model is trained and every time it answers a request, with more energy spent on cooling them. As AI use spreads, data centres are becoming a noticeable new source of electricity demand, which is why energy has become one of the biggest stories in technology.
This article explains where the energy goes, why demand is climbing despite efficiency gains, how companies and governments are responding, and what the trend means for a country like Vietnam, whose power system is already working hard to keep up with growth.
Where does the electricity go when you use AI?
There are two big phases. The first is training: feeding a model enormous amounts of data over weeks or months on thousands of graphics processors or similar AI chips running around the clock. Training a frontier model is a one-off job, but a very large one, and companies train many models and many versions.
The second phase is inference: running the trained model every time someone asks it something. A single answer uses far less energy than training, but popular services handle a vast number of requests every day, and newer models that “think” through problems step by step, or that generate images and video, use considerably more computing per request. Over time, inference can add up to more than training.
Cooling, efficiency and the question of scale
The chips are only part of the bill. Servers need memory, storage and networking, and all of that hardware turns electricity into heat. Keeping it cool requires air conditioning, chilled water or liquid cooling systems, which consume additional power and, in many designs, significant amounts of water. Data centres also keep backup systems ready so that services never go down.
This is why the industry tracks efficiency carefully. Operators measure how much of the total energy reaches the computing equipment versus overheads such as cooling. Newer facilities are much better on this measure than older server rooms, but even an efficient data centre full of AI chips needs a large, steady supply.
For an individual user, one chat message is not an energy disaster. The issue is scale: the same convenience multiplied across a whole population and built into search, office software and phones.
Why is AI energy demand climbing despite efficiency gains?
Bigger models. Each new generation of leading models has tended to use more computing for training, because scale has been a reliable way to improve capability.
More users and more features. AI is being added to search engines, email, office software and phones, turning occasional use into constant background use.
Heavier tasks. Generating images, video and long reasoning chains requires much more computing than answering a short text question.
Always-on availability. Services must respond instantly at any hour, so capacity is built for peak demand, not average demand.
The rebound effect. When computing becomes cheaper per task, people use more of it. Efficiency gains are often absorbed by new uses rather than saved.
Concentrated locations. Data centres cluster where land, fibre and power connections exist, putting heavy loads on specific parts of the grid.
Cooling in hot climates. Facilities in warm, humid regions need more energy or water for cooling, which matters for Southeast Asia.
Competition. Companies race to build capacity ahead of demand, so construction plans can outrun actual use, at least for a while.
How are companies and governments responding?
Large technology companies are signing long-term contracts for solar, wind and, increasingly, nuclear power, including interest in restarting older plants and in small modular reactors that are not yet widely deployed. Some are building data centres next to power plants or in regions with surplus clean energy. These deals are significant, but many of the newer technologies will take years to deliver electricity at scale.
At the same time, engineers are attacking the problem from the computing side. Chips are becoming more efficient per calculation, smaller models are being trained to handle common tasks, and techniques such as distillation and quantisation let a model run with less computing. Some AI tasks are moving onto phones and laptops, which spreads the load and can improve privacy.
Grid operators and governments are adjusting planning. In several countries, new data centres face waiting lists for grid connections, requirements to supply their own backup power, or rules about where they can be built. Water use has become a local political issue in some dry regions. There are also calls for more transparency, because companies do not always disclose how much energy and water their AI services use.
The honest picture is uncertain. Forecasts of AI energy demand vary widely, depending on assumptions about adoption and efficiency. What is clear is that electricity supply has become a strategic constraint on AI, not just an operating cost.
What does AI’s energy appetite mean for Vietnam?
Vietnam is both a growing user of AI and a potential location for data centres serving the region. The issues below are worth watching.
| Issue | Why it matters | What to watch |
|---|---|---|
| Grid reliability | Data centres need stable power around the clock; Vietnam has faced power shortages during hot dry periods in recent years. | Progress on transmission lines, new generation capacity and reserve margins. |
| Clean energy supply | Global tech customers increasingly want data centres powered by renewable energy. | Mechanisms that let large users buy clean electricity directly, and rules for rooftop and grid-connected solar. |
| Cooling and water | A hot, humid climate raises cooling needs; water use can compete with farms and cities. | Use of efficient cooling designs and water recycling in new facilities. |
| Data location rules | Requirements to store some data domestically encourage local data centres. | How data regulations evolve and how they affect cloud and AI services. |
| Skills and jobs | Data centres create fewer jobs than factories but need skilled electrical, cooling and network engineers. | Training programmes for data centre operations. |
| Cost for users | Businesses using AI through foreign cloud services pay prices set partly by global energy costs. | Local cloud options, pricing changes and on-device AI alternatives. |
The opportunity and the challenge are the same thing: whoever can offer reliable, clean, affordable power will attract the next wave of digital infrastructure.
Frequently asked questions
Does using a chatbot waste a lot of energy?
A single question uses a small amount of energy, though more than a simple web search, and heavier tasks like image or video generation use much more. The concern is the combined effect of billions of requests, not one person’s use.
Why do data centres use water?
Many cooling systems evaporate water to remove heat from servers, especially in hot climates. Some newer designs use closed-loop or air cooling to reduce water use, often at the cost of more electricity.
Is training or using AI more energy-intensive?
Training one large model uses a lot of energy at once, but everyday use by millions of people can add up to more over time. Both matter, and the balance depends on how widely a model is used.
Will AI make electricity more expensive?
In places where data centres add large demand quickly, they can strain grids and influence investment needs and prices. The effect depends on local supply, regulation and how new capacity is financed, so it varies by country.
Can AI help save energy too?
Yes, in some areas, such as forecasting solar and wind output, optimising building cooling and managing grids. Whether these savings outweigh AI’s own consumption is an open question.
