Introducing The State of AI, a novel partnership between the Financial Times and MIT Technology Review. For the next six weeks, each Monday, journalists from both organizations will explore a facet of the generative AI transformation impacting global influence.
This week’s discussion features Casey Crownhart, senior energy reporter at MIT Technology Review, alongside FT columnist Pilita Clark, who examine how China’s swift renewable energy expansion might allow it to surpass the US in AI advancements.
Casey Crownhart comments:
In the era of AI, the primary hurdle to advancement isn’t financial but rather energy. This is particularly concerning in the US, where large data centers are poised to activate, yet it seems unlikely that the nation will develop the consistent power supply or infrastructure to accommodate them all.
This situation wasn’t always present. For approximately ten years before 2020, data centers managed to balance rising demand with increased efficiency. Currently, however, electricity consumption is rising in the US, with billions of interactions with popular AI systems daily—and efficiency improvements are not keeping up. The insufficient introduction of new power capacity is beginning to reveal its stresses: Rising electricity costs are hitting residents in areas where data centers are increasingly burdening the grid.
If we aspire for AI to fulfill its significant promises without inflating electricity prices for the general public, the US must glean insights from global examples of energy abundance. Take China as an illustration.
In 2024, China added 429 GW of new power generation capacity, over six times the total capacity increase in the US during the same period.
While China still generates a considerable portion of its electricity from coal, its share of the energy mix is steadily declining. The nation is instead concentrating on deploying solar, wind, nuclear, and gas at unprecedented rates.
Conversely, the US is attempting to revive its struggling coal sector. Coal-powered facilities are both polluting and significantly costly to operate. Aging US plants are also less dependable than previously, producing electricity only 42% of the time, compared to a 61% capacity factor in 2014.
It’s a troubling scenario. Without change, the US risks evolving into a consumer rather than an innovator in both energy and AI technologies. Presently, China gains more from renewable exports than the US does from oil and gas exports.
Establishing and approving new renewable energy facilities would undoubtedly assist, given that they are now the most affordable and quickest to deploy. However, wind and solar are politically unpalatable to the current administration. Natural gas presents a sensible option, although there are worries regarding delays tied to essential equipment.
A swift solution could involve data centers being more adaptable. If they consented to refrain from drawing power from the grid during periods of strain, new AI infrastructure could potentially activate without new energy infrastructure being required.
A study by Duke University revealed that if data centers would scale back their consumption by 0.25% of the time (approximately 22 hours per year), the grid could accommodate power for roughly 76 GW of new demand. This is akin to adding about 5% to the total grid capacity without necessitating new construction.
However, flexibility alone would not suffice to genuinely accommodate the surge in electricity demand from AI. What are your thoughts on this, Pilita? What could alleviate the US from these energy constraints? Are there additional considerations we should examine regarding AI and its energy consumption?
Pilita Clark replies:
I concur. Data centers that can minimize their energy usage during periods of grid stress should become standard practice, not an exception. Additionally, we require more agreements similar to those offering lower-cost electricity to data centers that permit power utilities to access their backup generators. Both approaches reduce the necessity for constructing additional power plants, a sensible strategy regardless of how much energy AI ultimately consumes.
This is a crucial issue for nations worldwide, as we still lack precise knowledge about AI’s future energy consumption.
Projections for data center energy requirements in as little as five years differ significantly, with estimates ranging from less than double the current rates to quadruple.
This discrepancy is partly due to the scarcity of public data regarding AI system energy demands. Additionally, we cannot predict how much more efficient these systems will become. Last year, US semiconductor firm Nvidia reported that its specialized chips had become 45,000 times more energy efficient over the past eight years.
Furthermore, we have miscalculated technology energy requirements previously. During the pinnacle of the dot-com boom in 1999, there was an inaccurate assertion that the internet would consume half of the US’s electricity within a decade—leading to a need for increased coal power.
Nevertheless, certain nations are already experiencing the impact. In Ireland, data centers use such a significant amount of energy that new connections have been limited around Dublin to prevent overloading the grid.
Some regulatory bodies are contemplating new regulations mandating tech firms to generate adequate power to meet their consumption. I hope these initiatives expand. I also hope AI itself will enhance energy abundance and, crucially, hasten the global energy transition necessary to address climate change. OpenAI’s Sam Altman stated in 2023 that “once we acquire a truly powerful super intelligence, tackling climate change will not be particularly challenging.”
The current indications are not encouraging, especially in the US, where renewable initiatives are being terminated. Yet, the US could ultimately be an exception in a world where ever-lower renewable costs constituted over 90% of new global power capacity added last year.
Europe intends to primarily power one of its largest data centers with renewables and battery storage. However, the country leading the green energy charge is undoubtedly China.
The 20th century was characterized by nations abundant in fossil fuels, whose dominion the US presently seeks to extend. In contrast, China may emerge as the world’s first green electrostate. Should it achieve this in a manner that enables it to triumph in an AI competition that the US has predominantly controlled, it would mark a significant chapter in economic, technological, and geopolitical history.
Casey Crownhart responds:
I share your doubts regarding tech leaders’ assertions that AI will be revolutionary in the effort to combat climate change. While it is true that AI is advancing swiftly, we cannot afford to wait for technologies that make grand promises without substantial evidence to support them.
Regarding the grid, for instance, experts suggest there’s potential for AI to assist in planning and even operations, yet these initiatives remain in the experimental stage.
In the meantime, many parts of the globe are making tangible progress transitioning to newer, greener energy sources. The impact of this transition on the AI surge remains uncertain. What is evident is that AI is transforming our grid and our world, and we must remain realistic about the implications.
Further reading
MIT Technology Review journalists quantified the energy requirements of an AI query.
Several reasons exist to maintain a hopeful outlook on AI’s energy consumption.
The FT’s visual data team examines the relentless competition for AI capacity.
And global FT journalists inquire if data centers can genuinely be environmentally friendly.
