- It's up for debate whether enough capacity can be added to the grid to keep up with data center power demand
- Some have argued a new compute paradigm is needed
- But replacing entrenched systems will be wildly easier said than done
It’s no secret that data center operators are scrambling to find power. For many, the issue is framed as a battle to buy time until new sources of nuclear power can be ramped up and added to the grid in the coming decade. But Neil Sahota, CEO of ACSILabs and lead artificial intelligence (AI) advisor to the United Nations, argues this is simply a race that the industry cannot win.
Energy consumption, Sahota said, will continue to rise faster than new power generation can keep up. And there’s only so much you can do to improve efficiency in existing systems. “You can’t continuously improve a candle and get to a lightbulb,” he told Fierce. “We try to tweak energy sources and make optimizations, but there’s a limit to that and I think we’ve probably hit it.”
According to a December 2024 report from consulting firm Grid Strategies, data center power demand in the U.S. alone could exceed 90 gigawatts (GW) – or nearly 10% of the total forecast 2029 load of 942 GW. Indeed, the report adds “improving energy efficiency may only lead to more computing demand, not reduced energy demand.”
So then, what can be done? According to Sahota, the answer is to redesign compute from the ground up to come up with systems that use less power and don’t generate heat. And ironically, he said AI – which is partially responsible for the power crunch – can help.
For the record, Sahota isn’t the only one who has come to this conclusion. Rodolfo Rosini, CEO of startup Vaire Computing, is going all-in on reversible computing for exactly this reason.
Absent some sort of systemic change, Sahota warned the outcome could be dire.
“The ultimate thing is who gets the electricity,” he said. “We always worry about a war over water, but it might be a war over electricity.”
Massive lift
Setting aside for a moment the technical complexity and time involved in coming up with and subsequently scaling a new compute paradigm, there’s one other huge hurdle standing in the way of Sahota’s vision: money. And more specifically, the hundreds of billions of dollars that have already been or will soon be spent deploying data center infrastructure.
According to Dell’Oro Group, data center capex – that is, money spent on physical infrastructure, networking, servers and accelerators – was expected to hit $400 billion for the full year 2024. If hyperscaler announcements – like Microsoft’s plan to spend $80 billion on data centers in 2025 – are any indication, that figure could grow further this year.
Sahota isn’t blind to this fact. “When push comes to shove, it’s going to be a major burden,” he acknowledged. “That’s the toughest sell of all: Infrastructure projects don’t really have an ROI. They’re capital intensive and they take years for a payback period. But the alternative is you can’t do business.”
Reading from a different book
The other major issue is that not everyone has come to the same conclusion as Sahota and Rosini.
Wesley Cummins, the CEO of Applied Digital, told Fierce that nuclear is a good long-term solution. And in the interim, his company is focused on tapping into stranded power sources.
According to Cummins, in North America in particular there’s plenty of power that’s generated but that isn’t or can’t be used. Think wind energy created via turbines built in the middle of nowhere without a good connection to a grid capable of sending that power somewhere it’s needed. Or think of an area that previously had a large consumer of power – like an aluminum smelting plant – that went out of business and subsequently left the energy it used to use sitting on the grid.
“We mostly do wind,” Cummins said, adding that it pairs nicely with natural gas to achieve the required reliability profile. And in pursuit of that stranded resource, you “go to where wind blows the most and land is the cheapest.” Thus, Applied Digital is building a massive new 400 MW data center for a hyperscaler client in North Dakota. But with newly secured capital in hand, it also has its sights set on expanding to areas like Texas and Pennsylvania with similar stranded power profiles.
Beyond the aforementioned hurdles, it will also likely be a hard sell to convince government officials (in the U.S. especially) to prioritize sustainability over technology leadership. It’s an easier lift to, say, get those governments to greenlight new nuclear infrastructure to power the systems already in place.
“The United States dominating in the AI race is probably the most important thing we’ll ever do,” Cummins concluded. “The way you win is being able to build infrastructure and putting compute in place.”