The High Cost of Generative AI Energy Consumption

Generative AI models can consume a staggering amount of energy, with some models requiring as much as 1,000 times more energy than a typical household.

This is largely due to the massive computational resources required to train and run these models. For instance, the training of a single large language model can take up to 100,000 hours of processing time.

The environmental impact of this energy consumption is significant, with some estimates suggesting that the carbon footprint of AI training could be as high as 1.4 gigatons of CO2 per year.

Expand your knowledge: How to Learn Generative Ai

Around the globe, data centers currently account for about 1 to 1.5 percent of global electricity use, according to the International Energy Agency.

AI's energy requirements are a significant concern, with researchers warning that the world's AI boom could drive this number up a lot—and fast. A peer-reviewed analysis published in Joule quantifies the demand, showing that 1.5 million AI servers running at full capacity would consume at least 85.4 terawatt-hours of electricity annually.

The energy consumption of AI servers is staggering, more than what many small countries use in a year. This is a massive explosion in global electricity consumption, especially considering data centers are already responsible for consuming about 1 percent of global electricity.

The type of hardware used and the complexity of requests also matter, with the latest servers being more efficient than older ones. The more complicated a request, the longer the servers are working to fulfill it, the more power is consumed.

In a worst-case scenario, data centers could experience a 10-fold increase in energy consumption, which is not realistic but illustrates the energy-intensive nature of AI.

By 2027, GPUs are expected to constitute about 1.7 percent of the total electric capacity in the United States, which may seem minimal but represents a considerable growth rate over the next six years.

Data centers are projected to consume about 88 terawatt-hours (TWh) annually by 2030, which is roughly 1.6 times the electricity consumption of New York City.

GPUs will make up a significant amount of energy that will need to be supplied to data centers, with projections indicating they could account for as much as 27 percent of the planned new generation capacity for 2027.

The energy demand of GPUs is on an upward trajectory, with estimates suggesting they could make up 14 percent of total commercial energy needs by 2027.

Data center energy consumption is a complex issue, with the US Energy Information Administration (EIA) finding it challenging to quantify due to inadequate sampling frames and low response rates.

By 2030, data centers are expected to consume about 14 GW of energy annually, according to a McKinsey data center demand model based on the number of servers within data centers.

See what others are reading: How Generative Ai Will Transform Knowledge Work

Technology companies are positioning AI development as a climate solution and critical to innovation. They're exploring ways to make AI more energy-efficient.

Researchers are looking into using more efficient hardware chips to reduce the energy needed to create AI tools. This approach could have a significant impact on energy consumption.

Microsoft is building a data center in Quincy, Washington, which has raised concerns about the power it will consume. The data center could potentially suck up all the energy in the area, leading to blackouts during peak times.

Microsoft claims to work with authorities and utilities to avoid impacting local services. They also build supporting infrastructure to ensure residents don't experience utility service reductions.

Server farms that train and operate AI models may compete with local residents and businesses for power. This competition could lead to power shortages and blackouts.

Worth a look: Harnessing the Power of Generative Ai