Data centres globally consumed 220-330 TWh in 2021 (California uses around 278 TWh/year). How much more will they consume if AI takes off, given AI queries consume an order of magnitude more energy, and there are over 5bn internet users worldwide? The first step is to make a decent evidence-based prediction, but the U.S. and the EU are only expected to enforce reporting requirements next year, explains Meredith Fowlie at UC Berkeley’s Energy Institute at Haas. Next, a keen eye must be kept on “power usage effectiveness” (PUE), a measure of efficiency that reveals that global data centre energy use increased by only 6% between 2006-2018 while computing output and storage capacity increased by a factor of 6 and 25, respectively. Very good news! The bad news is that PUE has flatlined in recent years. Tensions between decarbonisation and the growth in data centres have resulted in some U.S. states considering postponing the phase-out of coal and gas plants to cope. So, on top of the global concern over the regulation of AI because of its impact on economies and society we must add its impact on energy consumption and our decarbonisation goals, says Fowlie.
Generative AI models like ChatGPT are delivering artificial intelligence to the masses. Anyone with an internet connection can now use sophisticated deep learning tools to complete complex tasks in seconds – from writing term papers to creating weird blog post art.
These models are trained on petabytes (a word I just learned) of data to generate new content based on patterns and information they have learned to recognise and mimic. All of this learning and inference requires a lot more computing power than the quaint Google searches of yesteryear. Alphabet’s John Hennessy estimates that my AI-supported art project required 10 times the computing costs of a standard internet image search.
What will happen to global energy use/GHG emissions as AI computing demand growth is compounded across 5.3 billion internet users? The range of opinions on this topic is pretty astounding. Some see climate destruction in the rise of AI. Others see a smoother clean energy transition. I can’t pretend to know where all this is going. But I can see complications ahead.
Data centres as energy hogs?
Nearly all the world’s internet traffic travels through data centres. These are large, increasingly “hyper-scale” buildings that house computing machines and related equipment. If we want to understand how increased computing demand will impact electricity demand, data centres seem like an important place to start.
The IEA estimates that data centres consumed 220-330 Terawatt hours in 2021. To put this into perspective, the entire state of California uses around 278 TWhs per year.
I’ve been curious to understand what rapidly expanding AI applications might imply for data centre electricity demand. If you ask the internet, the estimates are all over the place: AI is on track to consume our entire energy supply; US data centre demand will more than double by 2030; Data centres will draw up to 21% of the world’s electricity supply by 2030.
“PUE” efficiency innovations help, but for how long?
These projections are kind of alarming given all the other challenges we have on our decarbonisation plate. However, it is important to keep in mind that analysts have, in the past, overestimated data centre electricity use. In the 2000s, there were similar demand increases forecast. Big improvements in “power usage effectiveness” (PUE) meant that global data centre energy use increased by only 6% between 2006-2018 while computing output and storage capacity increased by a factor of 6 and 25, respectively.
The chart above tracks the remarkable improvements in PUE prior to 2019 using survey data from the Uptime Institute. But it also shows that these efficiency gains have flatlined in recent years. If efficiency gains stagnate while computing demand explodes, we’ll have an energy hog problem on our hands.
Data centres = climate angels?
There is a more optimistic counter-narrative out there arguing that data centres are not an energy problem, but rather a critical component of green and innovative climate solutions.
Barry Fischer, Google’s “Data Storyteller”, has several podcasts devoted to this topic (see here for example). And if you ask ChatGPT about the climate impacts of data centres and generative AI, it will quickly generate a very long list of innovations that will “enhance the development, deployment, and optimisation of clean energy technologies”.
As I understand it, the angelic vision is that data centres will procure their own additional 24×7 renewable energy so as to meet their electricity needs. Relatively flexible data centre loads will provide value to the grid. And the innovative AI applications that data centres support will help optimise/accelerate the clean energy transition. I see the potential here. But I think we’ll need more than potential and private sector promises to keep data centre energy consumption on the right track.
Across the country, we’re already starting to see tensions between our decarbonisation goals and a pressing need to accommodate growth in data centre electricity demand. Some recent examples include:
- New data centres in Virginia are driving significant and unexpected electricity demand increases. To keep up, Virginia utilities are asking to extend the lifespan of gas- and coal-burning plants.
- Wisconsin is also looking into delaying coal plant retirements to accommodate a new Microsoft data centre.
- In Maryland, a new data centre applied to have 504 MW of backup diesel generators exempted from environmental regulations on the grounds of “public convenience and necessity.”
A few anecdotes do not a crisis make. But delayed coal plant retirements and angry ratepayers do underscore some challenges ahead. These challenges are greatly exacerbated by a dearth of energy data from data centres. A lack of reliable information makes it really hard to anticipate and plan for data centre energy demand, let alone monitor and manage their environmental impacts.
Directing responsible data centre growth requires data
Mis-information about data centre energy use is feeding exaggerated claims about where data centre energy use is trending and what we should do about it. This is a problem we can solve. Why aren’t we systematically collecting good data on data centre energy use?
Back in 2018, the EIA did “assess the feasibility of collecting data and publishing estimates for data centre buildings”. Participation rates in a pilot survey were very low. A frustrated EIA concluded that collecting reliable data on data centre energy use would require “cooperation from the industry” and that effective data collection is “likely not feasible with current methods”.
Recently, Europe managed to establish detailed reporting requirements for the “energy performance and sustainability of data centres” which will take effect next year. Here in the US, Senator Whitehouse is developing draft legislation that would similarly require data centres in the U.S. to report operational data and environmental performance information. Following!
The widespread use of generative AI is raising all sorts of promise-or-peril-type questions around political polarisation, intellectual property, higher education, the clean energy transition. Policymakers are scrambling to understand how we can use regulation and policy tools to steer the responsible growth of AI. When it comes to managing data centre energy use and GHG emissions, it seems we already have the policy tools in hand. What we need now are better data and the political will to use them effectively.
Meredith Fowlie is an Associate Professor in the Department of Agricultural and Resource Economics at UC Berkeley. She is also a research associate at UC Berkeley’s Energy Institute at Haas and the National Bureau of Economic Research.
This article was first published on the Energy Institute Blog, the blog of the Energy Institute at Haas at the University of California, Berkeley, and is republished here with permission.
Keep up with Energy Institute blogs, research, and events on Twitter @energyathaas