Analysis-Resource-poor Japan was girding for an AI energy surge. DeepSeek raised the stakes

By John Geddie and Katya Golubkova

TOKYO (Reuters) - Weeks after Japan revealed details of a landmark energy plan partly designed to keep up with an expected AI surge, the shock rise of Chinese startup DeepSeek has upended conventional wisdom over the sector's future power demands.

It was only last year that Tokyo abandoned its long-held forecasts that its future electricity needs would dwindle with its ageing population, and began accounting for higher AI-driven usage from data centres and microchip makers.

In late December, the government released a draft of its basic energy plan, a major policy document reviewed about every three years, projecting electricity generation would rise between 10-20% by 2040 and citing those factors.

While Tokyo will not likely hurry to rejig its forecasts, DeepSeek's seemingly leaner models have triggered a broad rethink of AI energy needs that the world's most resource-poor major economy would be remiss to ignore, analysts say.

"It would be risky (for Japan) not to take this seriously," said Andrew DeWit, professor at Rikkyo University's School of Economic Policy Studies in Tokyo.

DeepSeek last week launched a free AI assistant that it says uses less data at a fraction of the cost of incumbent services. By Monday, it had overtaken U.S. rival ChatGPT in downloads from Apple's App Store, triggering a global selloff in tech shares.

Power producers, whose stock prices had outperformed on expectations of ballooning electricity demands needed to scale AI technologies, also took a hit as investors weighed DeepSeek's seemingly more energy-efficient models.

But as analysts parse DeepSeek's potential impact, an alternative view has emerged: its success may lower the barrier to entry in a sector dominated by Silicon Valley giants and catalyse higher overall electricity demand from new AI entrants.

This is a potential headache for Japan, which produces just 13% of its energy needs from domestic sources, the second lowest ratio of all 38 OECD countries, besting only Luxembourg.

"If AI proves to be cheaper to develop than currently expected, that would accelerate its mass introduction rather than slow it. If anything, it would increase power demand in the country," said Yuriy Humber, CEO of K.K. Yuri Group, an energy research and consulting firm based in Tokyo.

"Japanese officials have taken their time to adjust power demand forecasts even though the AI boom was apparent two years ago. I expect they will monitor the new developments carefully," he said.

The trade ministry, which oversees the country's long-term energy planning, did not immediately respond to a request for comment.

A senior official at the ministry, speaking on condition of anonymity, said he had been briefed on DeepSeek in December and concluded that, while impressive, the technology did not advance existing systems. He said he thought markets had over-reacted.

Tepco, Japan's biggest power company by sales, said it was monitoring DeepSeek's potential impact on power demand closely but has yet to assess its full effects.

'BITTER LEARNING'

Japan's grid monitor had for years predicted future electricity demand would decline gradually due to the adoption of energy-efficient equipment and a shrinking population.

But in 2024, it revised its outlook to reflect an overall increase, largely driven by an expected 5.14 million kWh of new power demand from data centres and chip plants by 2034.

Senior government officials have also cited AI-related energy needs as a reason to restart nuclear reactors, a sensitive subject in a country that suffered one of the world's worst nuclear disasters when the Fukushima plant was crippled by a massive earthquake and tsunami in 2011.

Mika Ohbayashi, director at Renewable Energy Institute in Tokyo, said that DeepSeek's emergence was a "clear indication" that AI may become more efficient and demand less power.

She criticised Japanese officials linking AI energy needs to the promotion of nuclear power, and said the government needs to do more to develop renewable energy instead.

The jury is out over whether DeepSeek's models could trigger more or less AI energy demands in the future, and analysts say its technology needs to be rigorously stress tested before countries rush to change their plans.

But Japan has had a bitter experience of getting its energy preparations for technology wrong in the past, said Rikkyo University's DeWit, pointing to the years before its economic bubble burst in the late 1980s.

"Japan was a chip leader at the time and they figured they were going to become number one and they built out the power system. And of course, as they entered the 90s as the bubble collapsed, that power demand did not eventuate.

"They've had a bitter learning. So it behoves the policy makers to take this seriously," he said.

(Reporting by John Geddie and Katya Golubkova; Additional reporting by Tim Kelly and Yuka Obayashi; Editing by Kim Coghill)