In particular, the computing power highly concentrated in cloud "hyperscalers" such as Amazon, Microsoft and Google is pivotal to crunching the colossal datasets used for training large language models and providing the distribution channel that model providers need to reach users.
Both AI model training and distribution require massive amounts of power, however, pushing tech companies to get increasingly involved in energy infrastructure and prompting worries in Europe about stability of the electricity grid.
Goldman Sachs has estimated* that the demand for power from data centers is set to grow 160 percent by 2030, boosting their worldwide share of power consumption from 1-2 percent to 3-4 percent.
And AI is expected to be the driving force of this growth: The research notes that for household AI chatbot ChatGPT to process a query requires almost 10 times as much electricity as a standard Google search.
Calculating the AI sector’s potential energy cost and carbon footprint, especially concerning the training and deployment of large language models, is extremely difficult and raises some fundamental questions about the sustainability of this exploding industry.
— Uncertainty factors —
The major players are even exploring nuclear power as a possible solution to provide carbon-free power to meet demand.
Microsoft, for example, is restoring the Three Mile Island nuclear reactors in the US in an exclusive deal to power its data centers. Earlier this month, Google announced that it has ordered several small modular nuclear reactors precisely to feed the data centers that support its AI technologies.
Big Tech investing in building its own infrastructure is nothing new, of course. Giants such as Google and Meta Platforms have become dominant in building submarine Internet cables as traditional players couldn’t keep up with their skyrocketing demand for high data capacity and low latency.
But tech companies face significant uncertainty as they enter these long-term investments in power plants. For all the hype around AI-powered growth, the sector is still in its infancy, as it only exploded two years ago. How much of this is a bubble that will eventually burst is anyone’s guess.
“The uncertainty stems from the fact that we don't know exactly how much this industry is going to grow,” Conall Heussaff, a research analyst at Bruegel, an economic think tank in Brussels, told MLex. “Some people say we'll use them for everything, and they will displace everybody's jobs. Other people say they'll be limited as writing and coding tools.”
Another factor casting doubt on future energy demand concerns the energy efficiency of the hardware on which AI will run in the near future. The graphics processing unit chips that have become an essential part of AI infrastructure were initially meant for gaming and were not designed to be particularly energy-efficient, given that they were not expected to become so widespread.
In a similar vein, the International Energy Agency published a report in July stressing that future energy consumption for data centers and AI is “highly uncertain” due to key uncertainties over demand, technological advances, and improved efficiency (see here).
— Managing the demand —
For Bruegel’s Heussaff, another determining factor regarding AI’s energy consumption relates to the technical characteristics of this power demand, which might become more manageable if data centers have the technical capacity to handle it flexibly.
That does not seem to be the case at the moment, which explains the Big Tech companies’ interest in nuclear power, a much more stable energy source than renewables.
By contrast, a flexible demand would entail training a large language model during off-peak hours, avoiding straining the energy grid. It could also involve moving the load to data centers located in different countries based on where there is abundant wind or solar power at that moment in time.
Ireland, the European tech hub where this core digital infrastructure is projected to account for up to one-third of the total electricity demand before the end of the decade, is a critical testbed for how the growing demand from data centers will be managed in the EU.
The current situation does not have a rosy outlook. Since last November, Ireland’s state-owned electricity system operator, EirGrid, has not authorized new data centers due to concerns related to power shortages and supplies, forcing Microsoft and Amazon Web Services to put their plans on hold.
“Given the intermittency of our renewables, flexibility of data center demand to ‘match’ renewable generation will become increasingly important in the future and can assist in meeting our decarbonization objectives,” the Irish government wrote in a statement published in July.
Especially in countries that have taken an anti-nuclear stance, at least part of that flat demand would likely have to be satisfied by gas, which is more polluting and is known for pushing wholesale electricity prices up on average due to the price-setting role it plays in Europe’s energy markets.
— Regulatory efforts —
How strongly policymakers can regulate these energy uses remains an open question, made even more challenging by the overwhelming lack of reliable and up-to-date data, which consistently blurs the picture.
The European Commission is setting up a database based on energy consumption data from the EU’s largest data center operators. The idea is to develop a report by May next year that could inform a legislative proposal mandating minimum performance standards for data centers (see here).
Moreover, the EU’s AI Act envisages the development of codes of conduct and technical standards for the energy-efficiency design, training, and use of AI models. It also mandates that model providers disclose their models' known or estimated energy consumption.
To this end, the EU's executive arm is currently working on commissioning a study to develop a measurement framework for the energy efficiency of general-purpose AI models.
Meanwhile, academics have pointed** to challenges in translating the AI Act’s reporting obligations to reality, suggesting new publishing requirements such as a sustainability risk assessment and renewable energy targets for data centers.
In the short to medium term, EU policymakers are chasing the growing problem, with the Irish experience suggesting that leaving data centers’ demand unchecked might imperil energy grid stability. For their part, tech companies investing in costly energy infrastructure must take the risk that it will be worth it if — when — the AI bubble bursts.
* Goldman Sachs, "AI is poised to drive 160% increase in data center power demand," online analysis article, May 14, 2024.
** Nicolas Alder, Kai Ebert, Ralf Herbrich and Philipp Hacker, "AI, Climate, and Transparency: Operationalizing and Improving the AI Act," research paper, Aug. 28, 2024.
Please email editors@mlex.com to contact the editorial staff regarding this story, or to submit the names of lawyers and advisers.