‘Within a decade or so, it could be the equivalent of the transport sector in terms of energy use,’ RMIT scientist Ascelin Gordon said.
Energy consumption from data centres in Australia could eclipse the transport sector within a decade, says one RMIT scientist.
Speaking to the Select Committee on Adopting Artificial Intelligence (AI) on July 17, Ascelin Gordon, a senior lecturer at RMIT University and a research fellow at the Interdisciplinary Conservation Science Research Group, presented his findings on the environmental impact of data centres and AI.
According to Mr. Gordon, data centres currently consume 1 to 3 percent of electricity globally, and consumption in Australia is about 5 percent of the country’s energy usage.
The lecturer said that the figure could reach 8-15 percent by 2030.
“There’s a lot of uncertainties in these projections, but it’s certainly going to be significant,” he said. “Within a decade or so, it could be the equivalent of the transport sector in energy use.”
Mark Stickells, CEO of Pawsey Supercomputing Research Centre (PSRC) in Western Australia, said that as the processing power of data centres increased, so does its energy consumption.
Mr. Stickells said PSRC currently operates the fourth most energy-efficient supercomputer in the world, which came online 18 months ago.
While the new system had a 30 times increase in computing capacity compared to previous iterations, it also used twice the power.
The CEO acknowledged that as the supercomputer’s efficiency rose, it had a larger environmental impact and despite the facility having a solar panel system, it still needed the power grid.
Supercomputers have a life cycle of about five years, according to Mr. Stickells.
How Much Electricity Does AI Use?
AI has gained popularity among the public with the advent of applications such as ChatGPT.
However, RMIT’s Mr. Gordon said not many people were aware of its environmental impacts.
“The impacts of the technologies we used are often under-recognised in the general public,” he said.
“I feel like the recent explosion of popularity in generative AI potentially threatens a step-change of environmental impact due to the computational infrastructure required to train and deploy the AI models we see around the world.”
Mr. Gordon explained that a single query on ChatGPT consumed around 10 to 90 times more energy than a Google search.
“If we generate an image, that’s probably 20 times more energy intensive again,” he said.
While the RMIT scientist did not have the exact figure for the energy consumption of AI video generation, he said it would be much more than creating an image.
“Generative AI producing video has just become a big thing this year, and there are many very impressive companies out there,” he said.
“I haven’t seen any good estimates of the energy intensity of generating video, but it’s probably orders of magnitude higher yet again.
“When we think there’s now millions or possibly billions of queries on these AI platforms daily, the energy use is quite staggering.”
Tackling the Environmental Impact
Mr. Gordon suggested the government look to encourage tech companies to tackle AI’s environmental impacts.
“There’s the Climate Neutral Data Center Pact, which seems to have lots of data centres in Europe coming on board and pledging to be carbon neutral by 2030 with their electricity use,” he said.
“So I guess encouraging and incentivising programs like that or approaches like that could be really positive.”
Mr. Gordon also said funding grants could also be conditions on the environmental footprint of a project, but noted it was tricky to do so as most companies operated offshore.
“These companies are big multinationals, and the data centres spread all around the world,” he said.
“When we use AI, often people probably don’t even know or think about where the actual computation is occurring as well. So it’s a really challenging area to regulate.”
Despite the legal challenge, Mr. Gordon said raising awareness among the general public and within the industry was important.