AI is not an easy field; model training requires enormous computational power and energy consumption. According to OpenAI According to publicly available data, a single training run of the ChatGPT model requires approximately 3,640 PF-days of computational power. ^[1] This is equivalent to possessing a computer capable of performing 10 quadrillion computations per second, operating continuously for nearly 10 years.

Model training is by no means a one-and-done endeavor; OpenAI releases a new version almost every year, with the number of parameters doubling each time. NVIDIA’s extraordinary performance alone offers a glimpse into just how intense this arms race in hardware has become. Data center Business revenue grew 409% year over year in fiscal year 2024!

The crux of the issue lies in the training process. Artificial intelligence models must learn from massive datasets, which requires powerful computers to run complex computations for extended periods. This continuous operation translates into enormous electricity consumption. Musk has voiced his concerns, emphasizing that AI could become a “power-hungry behemoth.”

Running AI models is extraordinarily costly—literally a “power hog.” According to a report in The New Yorker, citing research from foreign institutions, ChatGPT processes roughly 200 million requests per day, consuming more than 500,000 kilowatt-hours of electricity in the process. By comparison, the average U.S. household uses only about 29 kilowatt-hours per day, meaning ChatGPT’s daily energy consumption is equivalent to that of 17,000 American households.

　　According to estimates by the International Energy Agency (IEA), global data centers, cryptocurrency mining, and artificial intelligence consumed approximately 460 terawatt-hours of electricity in 2022, accounting for nearly 2% of total global electricity demand. By 2026, depending on the pace of deployment, improvements in efficiency, and trends in AI and cryptocurrency, this figure could double. This would represent an increase of 160 terawatt-hours—bringing total demand to 590 terawatt-hours—in 2026 compared with 2022, roughly equivalent to adding the electricity consumption of at least one Sweden or at most one Germany.

What is the way out of the energy predicament?

On the other hand, Bezos has proposed a solution: strategic data-center siting. Amazon invested $650 million to build a data center next to a nuclear power plant. This 1,200-acre facility is directly powered by the adjacent 2.5-gigawatt Susquehanna Steam Electric Station, which is the sixth-largest nuclear power plant in the United States.

Why relocate data centers next to nuclear power plants?

Nuclear power plants can provide data centers with a stable and reliable power supply, reducing their reliance on the grid and enhancing power supply reliability. Moreover, the relatively low fuel costs of nuclear energy can help data centers lower their operating expenses.

Data center energy efficiency can also be improved through the utilization of waste heat from nuclear power plants. The waste heat generated by nuclear power plants can be used for heating and cooling in data centers, thereby reducing their overall energy consumption.

By locating data centers near clean-energy sources such as nuclear power plants, the environmental impact can be minimized. This approach acknowledges the necessity of AI development while seeking ways to reduce its ecological footprint.

OpenAI CEO Sam Altman envisions a future in which AI and energy are deeply interconnected: if the deployment of artificial intelligence expands on the scale he anticipates, it will require “vast, vast” amounts of energy.

Oklo is one of the nuclear startups backed by Altman; he led Oklo’s seed-round financing in 2015 and subsequently served as Chairman of the Board. Oklo is working to commercialize nuclear fission—the same reaction that powers all existing nuclear power plants—but with significantly smaller reactors.

He has also invested $375 million in Helion, one of the emerging sectors composed of start-ups dedicated to demonstrating and commercializing nuclear fusion—the process by which the Sun generates energy without producing long-lived nuclear waste—yet which has never been replicated or scaled up on Earth. ^[6] In 2023, Microsoft also agreed to purchase electricity from Helion starting in 2028.

Turning to the domestic scene, miHoYo, in partnership with NIO Capital, led a nearly RMB 400 million investment in Energy Singularity, a company dedicated to developing controlled nuclear fusion technology to achieve “energy freedom” for humanity. ^[8] Previously, miHoYo also invested in several cutting-edge technology fields, including artificial intelligence and brain-computer interfaces.

Could Nuclear Energy Be AI’s “True Destiny”?

This issue extends beyond environmental impacts. The ever-growing energy demands of artificial intelligence could overburden existing power grids, leading to potential blackouts and service disruptions. Addressing this risk requires exploring alternative energy sources and developing more energy-efficient methods for training AI models.

Because nuclear energy can generate large amounts of electricity while emitting very few greenhouse gases, it is regarded as a viable alternative to fossil fuels. This perspective has rekindled interest in nuclear power and spurred a series of initiatives, including the launch of new nuclear power plant projects, the research and development of advanced nuclear technologies, and the exploration and development of new uranium deposits to meet anticipated demand.

The World Nuclear Association (WNA) projects that by 2030, 30 new nuclear power projects will be under construction worldwide. France has already planned to build a nuclear power plant specifically to supply electricity to data centers, which will offer new avenues for the development of artificial intelligence.