Can sustainable AI practices help reduce AI energy consumption?

So why can’t AI switch to using more electricity from wind, solar and water? The UK generated enough clean energy last year to power every household. The problem comes when you try to use green electrons to meet utility demand. Mousavi says that in states like California and Texas, where renewable energy capacity has grown rapidly, “the grid is not optimised to deliver renewable energy where it’s needed. When there’s not enough transmission capacity for renewable generation, particularly during peak times like midday for solar, they have no choice but to resort to curtailment – ​​simply disconnecting renewable energy capacity from the grid.”

As an example, he cites the extensive investment in solar power plants in California’s Central Valley. “The grid was not designed to transport the majority of the electricity demand in the San Francisco Bay Area from locations in the Central Valley where we’ve seen growth in solar power capacity, such as the suburbs of Davis and Merced, and it hasn’t been upgraded to do that yet,” he explains. “As a result, while the demand for power in the Bay Area is high, there isn’t the grid infrastructure to deliver all the renewable energy it needs. So what happens is they resort to taking installed and readily available renewable energy sources, such as solar, offline during peak demand to keep the electrical grid stable, and they have to use traditional power plants, commonly called ‘peakers,’ in the local consumption area during periods of high demand.”

Infrastructure investment will help, but it will take years or even decades to build out energy systems. What can be done about AI’s growing electricity demand? One option is to ration computing power, as AWS reportedly was forced to do in Ireland, or to encourage computing power to be moved and expanded to areas with high renewables, thereby reducing the additional pressure on the electricity grid to match renewable demand and supply. Another option is to continue burning fossil fuels. Despite all the problems, oil and gas plants are plentiful, reliable, and evenly distributed around major population centers.

But the more these sources are used, the less need there is to launch new renewable energy projects. The United States has a long list of clean energy projects awaiting regulatory approval. But as AI expands its reach from businesses to homes and personal devices, any “additional” elements of new renewable energy generation may be delayed.

To overcome the lack of access to clean energy, data centers have adopted workarounds such as power purchase agreements (PPAs) and renewable energy certificates (RECs) to offset or write off their carbon footprint. “PPAs and RECs are two drivers of generation growth, and while they have been a boon to the renewable energy industry, they are certainly not a panacea,” Mousavi explains.

“They do little to improve transmission or enable workable accountability mechanisms through proper attribution of emissions responsibility or e-responsibility to end users of electricity, and emissions trading through mechanisms such as RECs makes it virtually impossible for anyone to know the true impact of their emissions.

“As discussed in the paper ‘What is Scope 2 for?’, you will never know the exact impact of non-Scope 3 Category 3 (S3C3) emissions, and market-based Scope 2 (MBS2) emissions do not allow for that,” Mousavi continues. “We need to look at improving the balance between production and demand for renewable energy. We have rapidly increased clean generation over the years, but we have failed to commensurately improve the infrastructure to transport this increased capacity to where the demand is.”