What if the AI race isn’t about chips at all?

Open Editor’s Digest for free

Rula Khalaf, editor of the Financial Times, picks her favorite stories in this weekly newsletter.

Jensen Huang says China will win the AI ​​race. At first glance, it’s easy to assume that the billionaire founder of Nvidia is just talking about his book. Nvidia is certainly the biggest beneficiary of any narrative that encourages the United States to intensify its investments in artificial intelligence or ease regulatory restrictions imposed on its development, thus boosting demand for Nvidia chips. But does he have a point?

Not long ago, about a fifth of NVIDIA’s data center revenue came from China. Its fortunes depend on a steady stream of orders for its chips from governments, cloud service providers and artificial intelligence research labs around the world. Fear of China’s progress in artificial intelligence fuels this demand.

However, Hwang’s warning may hold some truth. The development of artificial intelligence has begun to shift from being limited primarily by the availability of cutting-edge chips to being limited by electricity supplies.

The GPT-4 model could use up to 463,269 megawatt-hours of electricity per year, according to research by academics at the University of Rhode Island, the University of Tunis and Providence College. This is more than the annual energy consumption of more than 35,000 homes in the United States. This demand reflects the increasing share of AI workloads in data center electricity consumption. Global electricity use by data centers is expected to double by 2030 and will reach about 1,800 terawatt-hours by 2040, enough to power 150 million American homes for a year, according to Rystad Energy.

As a result, the price and availability of energy will increasingly determine the pace of AI progress. Here, China has a head start. Last year, it added a record amount of renewable energy capacity, mostly from new solar and wind installations. Solar power alone expanded by about 277 gigawatts, while wind power contributed about 80 gigawatts, bringing the total new renewable capacity to more than 356 gigawatts, far exceeding the total capacity in the United States.

This renewed boom is part of a larger plan. Beijing has linked industrial policy to its efforts to strengthen the national grid, develop massive solar projects in Inner Mongolia, expand hydropower in Sichuan, and build high-voltage transmission lines to transport cheap inland electricity to coastal centers of demand.

Local authorities are also granting preferential electricity rates to companies such as Alibaba, Tencent and ByteDance to promote local AI computing. These subsidies help compensate for the lower efficiency of domestic chips produced by Huawei, allowing China to train AI models at a lower overall cost.

Meanwhile, in the United States, wholesale electricity costs have soared, with prices today 267 percent higher than five years ago in areas near data centers. But investment in many types of renewable energy projects, including large-scale wind and solar projects, declined in the United States during the first half of the year, reflecting policy shifts and regulatory uncertainty. The White House also released details of an executive order ending subsidies for wind and solar energy.

Some claim that China’s energy advantage cannot fully compensate for its lag in chip and model manufacturing. In fact, Nvidia’s H100 and Blackwell GPUs remain ahead of Chinese alternatives like Huawei’s Ascend 910B in terms of memory bandwidth and performance.

This imbalance was crucial in a phase of hardware-dominated technological competition, when access to the advanced chips that power computers and smartphones determined who led entire industries. For example, the United States curbed Huawei’s rise by restricting its supply of advanced chips starting in 2019.

The difference today, however, is that power is now starting to expand faster than transistors: chip performance gains have slowed to single digits, while renewable energy generation in China continues to expand at rates of more than 10% each year. Lower electricity costs increase the amount of computation that can be purchased with the same budget, and expanding network capacity allows models to be trained repeatedly for longer periods.

The race to master artificial intelligence is new but part of a centuries-old story. Throughout history, every technological superpower has risen on the back of cheap energy. Cheap and abundant coal was the engine of the Industrial Revolution in Britain. In the United States, oil and hydroelectric power fueled its dominance of manufacturing and military technology throughout the twentieth century.

The battle for control of AI is often framed as a competition over chips and the controls that govern them. But the power will belong to those who can keep the AI ​​models running.

june.yoon@ft.com