Why AI Data Centers Use So Much Electricity and Water

AI data centers consume large amounts of electricity and water primarily because training and running AI models requires massive parallel computation on thousands of power-hungry processors, and cooling those processors often relies on water-based systems. A single large AI training run can use several gigawatt-hours of electricity, and a medium-sized data center can use millions of gallons of water per day for cooling.

Background: The Rise of AI Computing

Traditional data centers already used significant power, but AI workloads are different. AI models, especially large language models and image generators, are built by feeding billions of examples through neural networks with billions of parameters. This process, called training, involves repeated matrix multiplications and other mathematical operations. To do this in a reasonable time, data centers use specialized chips like GPUs (graphics processing units) and TPUs (tensor processing units) that can perform many calculations at once. These chips are much more power-hungry than standard server CPUs. For example, a typical server CPU might draw 200-400 watts, while a high-end GPU can draw 300-700 watts under load, and a single AI server can contain 8 or more such GPUs.

Once a model is trained, it is deployed for “inference” — responding to user queries. Inference also requires substantial compute, though less than training. The sheer number of requests (millions per day for popular services) adds up to a constant high power draw. This is why AI data centers are often built with dedicated power substations and can consume as much electricity as a small town.

How AI Data Centers Use Electricity

Electricity in an AI data center goes to three main places:

• Compute hardware: The processors, memory, and storage that run the AI workloads. This accounts for roughly 70-80% of the total electricity use.
• Cooling systems: To remove the heat generated by the compute hardware. This can be 20-30% of total electricity.
• Power distribution and lighting: Transformers, uninterruptible power supplies (UPS), and building systems. These losses are small but not negligible.

The power draw is not constant. Training a large model can take weeks, during which the hardware runs at near maximum capacity. Inference loads fluctuate with user demand, but AI services often have high baseline usage. Data centers also use power for redundancy — backup systems that ensure uptime, though these only run during outages.

Why Cooling Is So Important

All that electricity turns into heat. A single GPU can produce as much heat as a small space heater. If not removed, this heat would damage the chips and reduce their lifespan. The cooling system must keep the chips below their operating temperature (usually around 70-85°C). The most common cooling methods are:

• Air cooling: Fans push air through the server racks, and the warm air is expelled to the outside or passed through chillers. This is simple but less efficient for high-density AI racks.
• Liquid cooling: Coolant (often water) is pumped through plates attached to the chips or through tubes in the server. Liquid can carry away heat much more efficiently than air. This is becoming standard for AI data centers because of the high heat density.
• Evaporative cooling: Water is evaporated to absorb heat, similar to how a swamp cooler works. This uses a lot of water but less electricity than air conditioning.

In many data centers, a combination is used: liquid cooling for the hottest components and air cooling for the rest. The choice depends on climate, water availability, and cost.

Water Usage: Where Does It Go?

Water is used in two main ways in AI data centers:

1. Cooling towers: In a typical water-cooled system, water is circulated through the data center to pick up heat, then sent to a cooling tower. In the tower, some of the water evaporates, which removes heat. The remaining water is recirculated. Because evaporation leaves minerals behind, a portion of the water is periodically drained and replaced (called “blowdown”). This is the largest water use.
2. Humidification: In dry climates, data centers may add moisture to the air to prevent static electricity. This water is also evaporated and must be replenished.

Additionally, water is used indirectly to generate electricity. Most power plants (coal, natural gas, nuclear) use water for cooling. So the electricity consumed by a data center also represents water consumed at the power plant. This is called “indirect water use.” Some studies estimate that the indirect water use can be several times the direct use, depending on the local energy mix.

Comparing AI Data Centers to Other Users

To put these numbers in perspective:

• A typical single-family home uses about 10,000-15,000 gallons of water per month. A large AI data center can use that much water in a few hours.
• A large data center might use 1-5 million gallons of water per day, which is comparable to a small city of 10,000-50,000 people.
• In terms of electricity, a 100-megawatt data center uses about as much power as 75,000 homes.

However, data centers are often located in regions with abundant water or cooler climates to reduce cooling needs. Some companies are building data centers in places like the Nordics or using seawater for cooling to reduce fresh water consumption.

Common Misconceptions and Realities

Misconception: “AI data centers use more water than all agriculture.” This is false. Agriculture is by far the largest water consumer globally, accounting for about 70% of freshwater withdrawals. Data centers are a small fraction of total water use, though they can be significant locally.

Reality: Water usage is not always reported. Many companies do not disclose their water consumption. This makes it hard to get accurate numbers. Some companies have started to report water usage in their sustainability reports, but it is not yet standard.

Reality: Water use can be reduced. New cooling technologies, such as closed-loop liquid cooling, use water more efficiently. Some data centers use recycled water or air cooling exclusively. The industry is working to reduce water intensity, but the rapid growth of AI is increasing total demand.

Frequently Asked Questions

Why not use air cooling only?

Air cooling is simpler and uses no water, but it is less effective at removing heat from high-density AI servers. As chips get more powerful, the amount of heat per square inch increases. Air cooling would require very high airflow, which uses a lot of electricity and may not be enough to keep chips cool. Liquid cooling is more efficient and can handle higher heat loads, which is why it is becoming the standard for AI.

Can AI data centers use seawater for cooling?

Yes, some data centers are built near the coast and use seawater for cooling. However, seawater is corrosive, so it requires special heat exchangers to isolate it from the internal cooling loop. This adds complexity and cost. It is not yet common but is a viable option for coastal locations.

How much water does a single AI model training use?

It varies greatly. Training a large model like GPT-3 is estimated to use several million liters of water for cooling, but this is an estimate and depends on the data center’s location and cooling efficiency. Smaller models use much less. The exact number is not publicly known for most models.

Are there any alternatives to water cooling?

Yes, there are alternatives like using refrigerants in a closed loop (similar to a refrigerator), or using immersion cooling where servers are submerged in a non-conductive liquid. These methods reduce or eliminate water use but may have higher initial costs. Some data centers also use free cooling by drawing in cold outside air when the climate allows.

Does the electricity used by AI data centers come from renewable sources?

Many large tech companies have committed to using 100% renewable energy for their data centers. However, this is often achieved through purchasing renewable energy credits, not necessarily by using solar or wind power at the exact moment of use. The actual grid mix varies by location. Some data centers are built next to hydroelectric dams or wind farms to use clean power directly.

Will AI data centers’ water use cause water shortages?

In most regions, data centers are a small fraction of total water use, so they are unlikely to cause shortages on their own. However, in water-stressed areas, a large data center can strain local water supplies. This is why some communities have raised concerns about new data center construction. Companies are increasingly being asked to use water more efficiently or use non-potable water.

Related

• AI-Generated Database Queries: Read-Only Access
• AI Hyperscalers’ Storytelling Problem Explained
• LangGraph Agents with Shared Persistent Python Memory