The artificial intelligence industry and power industry are working together to develop the first “power-flexible AI factory” at a 96-MW facility in Manassas, Va.
Nvidia, Emerald AI, the Electric Power Research Institute, Digital Realty and PJM are working to test the flexible capabilities of the Aurora AI Factory, which was designed from the bottom up to provide services to the grid. The power-flexible design, if adopted across the country, could unlock 100 GW of capacity on the grid, based on a study from Duke University. (See US Grid Has Flexible ‘Headroom’ for Data Center Demand Growth.)
Emerald is a startup working on the data center flexibility project with Nvidia, the largest company by market capitalization in the world because of its advanced chips that have fueled the rise of AI. The project is meant to help show the system can work, which would increase speed to market for data centers and take pressure off the grid, Emerald Chief Scientist Ayse Coskun said in an interview.
“AI data centers are facing a lot of wait time,” Coskun said. “In Virginia, we hear about five- to seven-year wait times for data centers to get connected.”
Load flexibility on the part of data centers means they can plug into the grid much more quickly because they can get offline when the system is stressed.
Nvidia provides the chips and offers control services for the project. EPRI is involved with its “DCFlex Initiative,” and the data center will bid services into PJM’s wholesale markets. (See EPRI Launches DCFlex Initiative to Help Integrate Data Centers onto the Grid.)
Data centers have varying levels of flexibility, from little to none at customer-facing facilities that make up the bulk of the facilities in Northern Virginia’s Data Center Alley, to cryptomining facilities that fall off the grid as soon as prices make their production unprofitable. AI data centers can be somewhere in the middle.
“A key ingredient in our technology is to make sure we meet these quality-of-service or priority constraints of customers,” Coskun said. “Some AI workloads fall into this category of being urgent and therefore not being flexible, but there’s a lot of other AI workloads.”
Some of the computing processes can be slowed or delayed for the few hours at a time when the grid would need to count on demand response from data centers, she added.
“Overall, when you look into the performance impact for this kind of actions, it’s minuscule,” Coskun said. “And in some cases, it’s not even noticeable.”
With ample benefits from speed-to-market concerns and little impact on AI data centers’ operations, flexibility makes sense, but it is early days of the concept for the customer class.
“Emerald AI is positioning itself to be this interface layer between the data centers and the power grid,” Coskun said. “Traditionally, there wasn’t a ton of communication between the power grids and the data centers, but as we design our data centers in a smarter and more flexible way, we believe there’s going to be this communication and programs may evolve. … There’s a ton of mechanisms that are existing in power markets that are not heavily used by data centers.”
The exciting thing about the Aurora facility is that it is being developed from the ground up for flexibility, which normally is an afterthought for data centers, EPRI Emerging Technologies Executive Anuja Ratnayake said in an interview.
EPRI’s DCFlex initiative was started to help the power industry meet the fast-growing demand for electricity from their expansion. The program also is working on real-world demonstrations at data centers in North Carolina and Arizona, the latter of which also includes Emerald.
“The major challenge for the industry is powering the data centers that are coming up at the moment, and the challenge comes from the scale and the pace of the growth in the data center sector,” Ratnayake said. “For the last 20-plus years … data centers grew up for enterprise purposes and for social media purposes and then for cloud purposes. What we are seeing happening in the last about two years is there is sort of a new type of a data center, which is what Nvidia is terming the AI factories.”
Data centers used to be five or 10 MW on the large side, but now with AI’s need for computing power and the energy to run all those Nvidia chips, it is seeing requests for 500 MW or even 1 GW, which is the size of a major city, she said.
“Think about the grid that is planned around these little loads that come together in the form of a city versus a single point in the grid that represents that same load,” Ratnayake said. “That’s new, and what that means … is the grid has to do a whole host of new investments, both potentially on the generation side and on the grid side.”
It can take up to a decade or more to build new generation and wires, but the data centers want to connect in a year or two, she noted. If data centers can respond and cut the amount of energy pulled from the grid, they can get connected while the grid is being expanded.
“This is that seven- to 10-plus-year period,” Ratnayake said. “During that period, if you’re able to be flexible, we can potentially connect you faster. That’s where the flexibility piece becomes important.”
One of the questions EPRI is studying is how much flexibility data centers might continue to provide to the grid once it has been expanded.
“It will be tied closer to business models more than really the technology viability,” Ratnayake said. “The technology viability will exist forever, but it will be up to the data center operators to really embrace which business model makes the best sense.”
