'Volatile' generative AI has changed the power demand game: Hitachi Energy CEO

The volatility of demand for power from generative artificial intelligence is posing entirely new problems for electricity networks, the CEO of Hitachi Energy has warned.

“We are finding two big differences in terms of energy demand from data centres used for generative AI," he said.

The first difference highlighted is in terms of the sheer size of the connection demand.

“What we might call a 'normal' data centre will start up with say 30, 40, 50 or maybe even 100MW, but with generative AI it is jumping to 500MW, 700MW, 1GW, or even higher.

“We are connecting now to high voltage clients, rather than medium voltage lines. It’s a completely different animal,” Schierenbeck stated.

But in his view, the more fundamental difference is in the volatility of the demand profile.

Prior to the advent of generative AI, demand from data centre clients would involve forms of optimisation for a steady power supply, but this has changed.

“You may be transferring load from one data centre to another because they want the optimal working point with the best contractual terms for that kind of energy consumption, But at the end of the day it was still a steady load, and it was balanced," said Schierenbeck.

“You can no longer do that with generative AI, because the behaviour is completely different. You train an algorithm and give it the data to learn and digest, and it will start to use as much computing power as it can get. You'll see a spike power consumption as they draw as much energy as fast as they can.”

To make matters more challenging, this rapid spike in demand can drop off equally dramatically when the learning process finishes.

"This is less of an issue for us when we provide something like the transformer, but it definitely is a problem for the grid operator,” Schierenbeck added.

Contrast with heavy industry

Schierenbeck offers a contrast with more traditional industrial loads, such as an electrical furnace or a paper mill.

“They would expect to be penalised if they caused demand to spike in just 15 minutes then disappear 15 minutes later, and did so without even giving notice or scheduling,” he said.

“But this is what the generative AI data centres are doing. Behaviour is quite different than what we have seen before, and this has come as a surprise to everybody.

"We are adding volatility on the production with renewables, but also on the consumption end, and we have to deal with that volatility."

Solutions to this are likely to involve more investment in mitigation infrastructures, such as battery energy storage systems, he reckons.

How big is the boom?

Broader unpredictability in terms of power demand has been one of the dominant themes in the global debate about AI.

Other big energy infrastructure companies followed a similar pattern.

Analysts at Citi agreed that the DeepSeek AI model could impact the growth projections for energy infrastructure if it proved to be as computationally efficient as claimed.

These energy infrastucture stocks found price support, however, as Chinese claims and the broader theme of energy demand was subjected to a more nuanced analysis.

“If the computational efficiency claimed by DeepSeek is true, then much less electricity would be required to achieve the same level of processing,” agreed consultancy firm Rystad Energy.

“But if data centres become more efficient it is also likely that they process more data. It is therefore too early to conclude if our forecast needs to be revised down proportionally to the efficiency gains in computational power.”

Growth spurt

Power demand from data centres started to grow exponentially over the past few years on the back of the technological revolution, Rystad points out.

Amazon, for example, has signed more than 500 PPAs in 27 country, securing 77TWh of electricity annually.

In 2024, Microsoft signed the largest corporate PPA to date with Canada-headquartered Brookfield Renewable Partners for a total capacity of 10.5GW of renewable energy in the US and Europe.

Power consumption from data centres in the US totaled close to 140 terawatt-hours (TWh) in 2023.

"While this represents only 3.5% of total consumption, it has grown from less than 50 TWh over the past decade, representing a compound annual growth rate (CAGR) of 13%, Rystad said.

In research conducted prior to the DeepSeek claims, Rystad forecast that global electricity demand from data centers will more than double by the end of the decade, driven by the tech companies' expanding their processing capacity.

"When we did this research, on average, we weree seeing a utilisation of around 40%," said Carlos Torres Diaz, head of power with Rystad.

"What I've been wondering about DeepSeek is whether it will reduce demand, or will we increase processing? In my view, more efficiency in the data process doesn't necessarily translate into lower utilisation of the data centres."

Displacing coal?

The consultancy firm also pointed out that any slack from reduction in demand from Big Tech could be gobbled up as affordable energy, perhaps even accelerating the energy transition.

“If electricity consumption for data processing is reduced, then it is likely that tech companies will also reduce the number of PPAs signed to secure energy for their facilities. This would in turn make more renewable energy available for other sectors, helping displace faster the use of fossil fuels,” it stated.

This assessment could be relevant to China's continuing heavy reliance on coal power generation.

"They are building a lot of renewable energy capacity, but given that their demand is growing quite fast, then they are struggling to displace completely the use of fossil fuels. If they suddenly manage to slow down their demand growth because they are requiring less for data centers, then that might help them displace fossil fuels faster," Diaz said.

More efficiency in computation can hardly be a bad thing, according to Hitachi Energy's Schierenbeck.

"If you look back in history, we have always had a tendency to underestimate demand and capacity in the future, especially in terms of computing power but also things like solar.

"With AI thing, I think we will have great growth curve. There may be significant energy savings to be made, but the underlying trend is not endangered.