Energy-Storage.news Premium speaks with Scott Blalock, general manager, integrated applications engineering, at BESS integrator Wärtsilä Energy Storage.
The rapid growth of AI data centres is creating an unprecedented challenge for energy infrastructure, forcing developers to in turn, rethink how they approach battery energy storage systems (BESS).
“Most of the developers understand to some extent their power demand, meaning how many engines, how many turbines do they need,” explains Blalock. “What they are struggling with is how much [energy] storage they need.”
Unlike traditional data centres with relatively stable power consumption, AI-specific facilities experience dramatic fluctuations.
Try Premium for just $1
- Full premium access for the first month at only $1
- Converts to an annual rate after 30 days unless cancelled
- Cancel anytime during the trial period
Premium Benefits
- Expert industry analysis and interviews
- Digital access to PV Tech Power journal
- Exclusive event discounts
Or get the full Premium subscription right away
Or continue reading this article for free
“Basically the GPUs ramp up very quickly, they go at nearly full power, and then they all ramp down really quickly,” Blalock says. “It’s that fluctuation that’s the hard part, because the engines or the generation can’t change that fast. Even if you’re grid-tied, the grid does not like that fast of a change.”
He continues, “That’s where storage comes into play, to smooth that out and basically make it invisible to the grid.” BESS help to protect generation assets from rapid changes that could otherwise damage them.
For traditional grid-tied BESS projects, sizing is fairly straightforward. Developers know their interconnect agreements, their power purchase agreements (PPAs), and can specify exactly how many MWs they need and for what duration.
AI data centres are different. Blalock explains, “With an AI data centre developer, they don’t know that target, and they’re coming to us saying, ‘Tell me how much battery I need.'”
Wärtsilä’s approach starts with the load change. “If it’s going from 20% to 80%, whatever that load change is, start with one-to-one power,” he says. “So, if it’s a 600MW load change that you have to absorb, start with 600MW of batteries. Then for the duration, start from 2-hour.”
However, that is only the beginning of the process. “We have to model the whole system. We have to know what the generation assets are. If it’s grid-tied, we need to know about that grid connection. We need to know as much as we can about the load profile itself.”
The moving target
Complicating matters further is the fact that AI load profiles are closely guarded secrets.
“It’s something that chip providers like NVIDIA, or whoever’s building those chips, hold very close to their chest—what that load profile looks like,” Blalock notes. “It can also change from one generation of chip to the next.”
Even training methodologies could alter power consumption patterns. “If that load profile changes, you may have built the BESS that could support the original load profile, but then something about it changes and it becomes a little more aggressive, and the power system struggles with it because it’s out of spec from what it was built to do.”
Given interconnection delays and the unique power requirements of AI facilities, many developers are opting for islanded or off-grid configurations with on-site generation.
Blalock bluntly states, “It’s basically impossible to have that without a BESS. You just can’t do it, really, at least not with an AI data center, just because the generators themselves just can’t handle that type of a load profile by themselves.”
Connecting volatile AI loads directly to the grid without storage, also is not viable, because as Blalock notes, “you kind of destroy the grid.”
While Wärtsilä has experience with islanded power systems in places like island nations, AI data centres are a significant leap in scale.
“Those [island systems] are generally in the 100MW or less range. Some of the data centres are talking around a gigawatt. So, an order of magnitude more complex,” he highlights.
He’s direct about the challenge, “Contained, islanded grids of that size simply don’t exist. Nobody’s ever done it before.”
Coordinating generation assets, managing maintenance schedules, and ensuring uninterrupted power to the data centre while maintaining battery state of charge requires sophisticated control systems operating across multiple timescales, from milliseconds to hours, adding to the complexity.
Pairing gas and storage
In April, Wärtsilä announced it would supply an off-grid power solution for a new data centre in Texas. The 790MW power plant will operate with 42 Wärtsilä 50SG engines running on natural gas.
When asked about the installation of large gas engine installations for data centres without mention of batteries, Blalock suggests it’s a matter of timing rather than intent.
“The thing that takes the longest—first you’ve got to get the land, then you have to start planning the structures. Then, for the power, the longest lead times are the transformers and the generators, those gas engines,” he explains.
Batteries have much shorter lead times, around 18 months compared to years for generation equipment. “Once they get the engines squared away, then they start looking at the next thing on the list, which is batteries.”
However, he predicts for facilities that skip storage, “I think they’re going to regret it. I don’t think it’s going to work very well, or at all, if they don’t have a BESS.”
