AI Energy and Water Consumption

Training Chat GPT-3 Uses How Much Energy?!!! ⚡️🤢

It’s Tuesday conscious designers! ⚡️ Today we’re exploring the energy and water usage of AI. 💦 Read to the end because I’m also sharing solutions to get us started. ⭐️

“The development of the next generation of A.I. tools cannot come at the expense of the health of our planet,” — Massachusetts Senator Edward Markey

How Does AI Consume Energy?

To understand this we need a quick reminder on how AI works. 🤖 If you’re a complete beginner, please start with this article, Your Comprehensive Guide to AI. AI relies on huge amounts of data and computationally intensive calculations to extract patterns, which results in the answer to your GPT-4 question. 🪄 It’s not magic, it’s brute forcing a shit ton of data.

Servers and Data centers

Large AI models like GPT-4 and Claude 3.5, are trained and deployed on large clusters of servers with multiple graphic processing units (GPUs). This is why AI training and usage results in massive energy consumption. 💡 Data centers’ electricity consumption in 2026 is projected to reach 1,000 terawatts, roughly Japan’s total consumption.¹

Learn more about the environmental impact of digital products here. 💻📱

How is Energy Measured?

Before we go into energy and water usage, I want to give you a clear understanding of what this means. Energy is measured in two key ways: 1. how much energy something uses (joules) and how quickly it uses that energy (watts or kilowatts).

1. Joules: Amount of energy

Think of joules as a tiny drop of energy.

Example:

• Lifting a small apple about one meter (3 feet) uses about 1 joule of energy.

2. Watts, Kilowatts, Megawatts: Speed of energy usage

Watts measure how quickly energy is being used, like water flowing from a tap.

• 1 watt = 1 joule per second

• 1 kilowatt = 1,000 watts

• Megawatt (MW) = 1,000 kilowatts = 1,000,000 watts

Examples from small to large. 📈

• A light bulb uses about 60 Watts.

• A microwave uses about 1 Kilowatt.

• A typical coal power plant might produce 600-700 MW.

• New York City's peak power demand can reach over 11,000 MW on hot summer days.

• The world's largest solar farm, Bhadla Solar Park in India, has a capacity of 2,245 MW.

3. Kilowatt-hours: total energy used over time.

We often see kilowatt-hours (kWh) on electricity bills.

1 kilowatt-hour = using 1 kilowatt for 1 hour = 3,600,000 joules

Example: If you run a 1,000-watt (1 kilowatt) microwave for 1 hour, you've used 1 kilowatt-hour of energy.

Make sense? 😊

Here are some real examples of energy used over time:

• A smartphone battery holds about 40,000 joules or 0.011 kWh.

• An American home uses about 30 kWh per day.

• Charging a Tesla Model 3 for 8 hours uses about 61.6 kWh.

• A large Bitcoin mining operation might consume 200-300 MWh per day.

• A large data center can consume 400-500 MWh in a day.

• A 100 MW wind farm operating at 35% capacity could produce about 840 MWh per day.

Great! With that out of the way, let’s talk energy. 👇🏽

How Much Energy Does AI Consume?

While we have some estimates, such as, “Training a large language model like GPT-3 is estimated to use just under 1,300 megawatt hours (MWh) of electricity,” the truth is machine learning models can vary drastically; figures are contingent. This means we only have a ballpark idea of how much energy and water is actually being used. It’s going to depend on the product. Companies are simply not sharing the data.

When contacted by The Verge, Judy Priest, CTO for cloud operations and innovations at Microsoft said in an e-mail that the company is currently “investing in developing methodologies to quantify the energy use and carbon impact of AI while working on ways to make large systems more efficient, in both training and application.” OpenAI and Meta did not respond to requests for comment.

We don’t know enough on how much energy is being consumed because there are no clear reporting standards. This will change next year with the EU AI Act that requires reporting on consumption, resource use, and other impacts throughout their systems’ lifecycle.² Thankfully, the US is trying to pass a similar bill.³

The Water Crisis Accelerated by Artificial Intelligence

One of our most finite resources, water, is also being depleted. AI also requires millions of gallons of water to cool data center equipment and for offsite electricity generation.

1. Onsite water consumption is necessary to cool servers. Heat is dissipated into the outside environment. Data centers use cooling towers and/or outside air, which requires an enormous amount of clean, fresh water. Less than 2.5% of the world’s water is fresh water.

2. Offsite water consumption is perhaps lesser known. Generating electricity also consumes water through cooling at thermal power and nuclear plants and expedited water evaporation caused by hydropower plants.

3. There is even a third culprit for water usage if we expand to AI supply chains. For example, to produce a microchip takes approximately 2,200 gallons of Ultra-Pure Water (UPW).⁴

An example of a data centre’s operational water usage: on-site scope-1 water for server cooling (via cooling towers in the example) and off-site scope-2 water usage for electricity generation. The icons for AI models are only for illustration purposes. (OECD)

Looking at Microsoft’s global data centers, AI consumes 1.8 – 12 litres of water for each kWh of energy usage, with Ireland and the state of Washington being the most and least water-efficient locations, respectively.⁵ Forbes also cites, “a single ChatGPT conversation uses about 50 centilitres of water, equivalent to one plastic bottle.”⁶ How crazy is that?!😳

AI Training vs Deployment

There is a distinction here to be made in training versus deployment. Training is more energy intensive and thus, consumes MORE than regular data center activity. For example, GPT-3 is estimated to use just under 1,300 megawatt hours (MWh) of electricity; about as much power as consumed annually by 130 US homes. 🤯 To put that in context, streaming an hour of Netflix requires around 0.8 kWh (0.0008 MWh) of electricity. That means you’d have to watch 1.6 million hours to consume the same amount of power it takes to train GPT-3.⁷

Do you feel sick thinking about that? I do. 🤢 There are serious environmental concerns here as everyone gets caught up in the hype of AI and mindless consumption.

What’s the Problem?