We’re back with The State of AI, a fresh collaboration between the Financial Times and MIT Technology Review. Every Monday, writers from both outlets discuss a facet of the generative AI revolution and its impact on global influence.
This week, Casey Crownhart, MIT Technology Review’s senior energy reporter, teams up with FT columnist Pilita Clark to explore how China’s swift development of renewable energy could enable it to outpace AI advancements.
Casey Crownhart states:
In the era of AI, the primary obstacle to advancement is not finances but energy supply. This is particularly concerning in the US, where large data centers are poised to come online, yet the nation appears unprepared to develop the consistent power resources or infrastructure necessary to support them.
This wasn’t always the case. Prior to 2020, for roughly a decade, data centers were capable of meeting heightened demand through efficiency enhancements. Currently, however, electricity demand in the US is on the rise, driven by billions of requests to popular AI models daily—and efficiency improvements are failing to keep up. With insufficient new power capacity being added, the pressure is becoming evident: Electricity rates are skyrocketing for residents in areas where data centers increasingly burden the grid.
If we intend for AI to fulfill its substantial promises without escalating electricity costs for the general public, the US must adopt some energy abundance lessons from abroad. Look no further than China.
In 2024, China added 429 GW of new power generation capabilities, exceeding the net gain in the US by more than six times during the same period.
China continues to rely heavily on coal for electricity generation, yet this constitutes a diminishing fraction of its energy mix. The focus now lies on ramping up solar, wind, nuclear, and gas at unprecedented speeds.
In contrast, the US is concentrating on resurrecting its struggling coal sector. Coal-powered facilities are not only polluting but also notably costly to operate. Aging plants in the US have become less dependable, generating electricity just 42% of the time, a decline from a 61% capacity factor in 2014.
The situation is far from ideal. Unless the US implements significant changes, we face the prospect of being mere consumers rather than pioneers in both the energy and AI technology sectors. Currently, China gains more revenue from its renewable energy exports than the US does from oil and gas exports.
Constructing and permitting new renewable energy plants would unquestionably assist, as they are presently the most cost-effective and rapid to deploy. However, wind and solar face political unpopularity in the current administration. Natural gas emerges as a potential option, although there are concerns regarding delays in acquiring essential equipment.
A rapid solution would be for data centers to adopt greater flexibility. If they could commit to not drawing power from the grid during peak stress times, new AI infrastructure might be introduced without necessitating additional energy infrastructure.
A study from Duke University revealed that if data centers would agree to reduce their power consumption by merely 0.25% of the time (approximately 22 hours per year), the grid could supply energy for about 76 GW of new demand. That equates to adding roughly 5% of the total grid’s capacity without requiring any new constructions.
However, flexibility alone would not suffice to adequately accommodate the surge in AI’s electricity demands. What’s your perspective, Pilita? What measures could extricate the US from these energy limitations? Are there alternative approaches we should consider regarding AI and its energy consumption?
Pilita Clark replies:
I concur. Data centers capable of reducing their electricity usage during times of grid stress should become standard, not an exception. Similarly, more arrangements like those providing low-cost electricity to data centers that permit power utilities to utilize their backup generators should be encouraged. Both strategies reduce the need for new power plants, which is logical regardless of how much electricity AI ultimately consumes.
This is a crucial consideration for nations worldwide, as we still lack precise knowledge of AI’s future power consumption.
Predictions for data center requirements in merely five years’ time vary significantly, ranging from needing less than double today’s consumption to four times that amount.
This uncertainty stems partly from the insufficient public data concerning the energy needs of AI systems. Additionally, we are uncertain about how much more efficient these systems will evolve. Last year, US chipmaker Nvidia claimed its specialized chips had become 45,000 times more energy efficient over the past eight years.
Furthermore, we have previously miscalculated technology’s energy requirements. During the dot-com boom peak in 1999, it was incorrectly predicted that the internet would consume half of the US’s electricity within a decade—necessitating a substantial increase in coal power.
Nonetheless, certain countries are already feeling the urgency. In Ireland, data centers consume so much electricity that new connections around Dublin have been restricted to prevent grid overload.
Some regulators are contemplating new regulations mandating tech firms to generate sufficient power to meet their demand. I hope these initiatives expand. I also wish AI itself would contribute to increased energy abundance and, importantly, expedite the global energy transition essential for combating climate change. OpenAI’s Sam Altman remarked in 2023 that “once we obtain a truly powerful superintelligence, tackling climate change will not be particularly challenging.”
The current indications, especially in the US, are not encouraging, with renewable projects being cut. However, the US could ultimately be an anomaly in a world where increasingly affordable renewables accounted for over 90% of new global power capacity added last year.
Europe aims to power one of its largest data centers mainly with renewables and batteries. However, the nation spearheading the green energy movement is evidently China.
The 20th century was ruled by countries rich in fossil fuels, whose influence the US now seeks to extend. In contrast, China may emerge as the world’s first green electrostate. If it manages to accomplish this while succeeding in the AI race that the US has thus far dominated, it will represent a remarkable phase in economic, technological, and geopolitical history.
Casey Crownhart responds:
I echo your skepticism regarding tech leaders’ assertions that AI will be a transformational aid in the battle against climate change. To be fair, AI is evolving rapidly. Yet, we cannot afford to wait for technologies that rely on grand promises without supporting evidence.
For instance, when it comes to the grid, experts suggest there exists potential for AI to assist with planning and even operation, but these initiatives remain experimental.
Meanwhile, considerable portions of the globe are making significant strides in transitioning to sustainable energy sources. The ramifications of this shift on the AI surge are yet to be determined. What is evident is that AI is transforming our grid and our world, and we must be realistic about the outcomes.
