Back
This is the fourth and final article in our series “Boon or bane: What will data centers do to the grid?”
Tom Wilson is aware that the explosive growth of data centers could make electricity costlier and dirtier. As a principal technical executive at the Electric Power Research Institute, the premier U.S. utility research organization, he’s studied the risks himself.
But he also thinks conversations about the problem tend to miss a key point. Data centers could also make the grid cleaner and cheaper by embracing a simple concept: flexibility.
“Data centers are not just load — they can also be grid assets,” he said. Turning that proposition into reality is the goal of his most recent project, DCFlex, a collaborative effort to get data centers to “support the electric grid, enable better asset utilization, and support the clean energy transition.”
DCFlex is short for “data center flexibility,” a term that encompasses all the ways that these sprawling campuses and buildings full of servers, cooling equipment, power-control systems, backup generators, and batteries can reduce or shift their power use.
Since its October launch, DCFlex has grown from 15 to 37 funding participants. On the data-center side are tech “hyperscalers” like Google, Meta, and Microsoft; major data center developers like Compass and QTS; and AI computing and power equipment suppliers like Nvidia and Schneider Electric.
On the grid side are utilities such as Duke Energy, Pacific Gas & Electric, Portland General Electric, and Southern Company; power plant owners like Constellation Energy, NRG Energy, and Vistra; and five of the continent’s seven grid operators, which manage energy markets serving electricity to two-thirds of the U.S. population.
The range of participants reflects the broad interest in solving the pressing challenge of powering the data centers being proposed around the country without driving up grid costs and emissions.
That won’t be easy. From Virginia’s “Data Center Alley” to emerging hot spots in Arizona, Georgia, Indiana, Ohio, and beyond, utilities are being inundated with demands for round-the-clock power from data center projects that can add the equivalent of a small city’s electricity consumption within a few years. Meanwhile, it usually takes four or five years to connect new power plants to the grid.
Flexibility could make a big difference, however, said Tom Wilson, who’s worked on climate and energy issues for more than four decades, including advising projects at the Massachusetts Institute of Technology and Stanford University and serving at the White House Office of Science and Technology Policy during the Biden administration.
That’s because the impacts of massive new utility customers like data centers are tied not just to how much power they need but specifically to when they need it.
Utilities live and die by the few hours per year when demand for electricity peaks — usually during the hottest and coldest days. By refraining from using grid power during those peak hours, new data centers could significantly reduce their impact on utility costs and carbon emissions.
If data centers and other big electricity customers committed to curtailing their power use during peak hours, it could unlock tens of gigawatts of “spare” capacity on U.S. grids, according to a recent analysis from Duke University’s Nicholas Institute for Energy, Environment & Sustainability.
Realizing that spare capacity will be challenging, though. For starters, every new large power customer would have to agree not to use grid power during key hours of the year, which is far from a realistic expectation today. What’s more, utilities would need some serious proof that those big customers can actually follow through with promises to not use power during those peak times before letting them connect, because broken promises in this case could lead to overloaded grids or forced blackouts.
And flexibility can’t solve all the power problems that massive data center expansion could cause.
In a December report, consultancy Grid Strategies found that key data center markets are driving an unprecedented fivefold increase in the amount of new power demand that U.S. utilities and grid operators forecast over the next half decade or so. While that analysis “really focused on the peak demand forecast,” the sheer amount of power needed over the course of a year is “potentially just as big of a story,” said John Wilson, Grid Strategies’ vice president.
Still, for utilities struggling to plan and build the generation and grid infrastructure needed to support data centers, flexibility is worth exploring. That’s because data centers could make their operations flexible a lot faster than utilities can expand power grids and build power plants.
It typically takes seven to 10 years to build high-voltage transmission lines and four to five years to build a gas-fired power plant — “even in Texas,” Tom Wilson pointed out. The concept of relying on big customers to avoid using power instead of building all that infrastructure is just starting to take hold in utility planning, but it could play a major role in managing the surge in power demand.
Flexible data centers may also be able to secure space on capacity-constrained grids more quickly than inflexible competitors, Tom Wilson said. A U.S. Department of Energy report released last year included interviews with dozens of utilities, and one key takeaway was that “electricity providers often can accommodate the energy and capacity requests of a data center for (say) 350 days but need to find a win-win solution for the remaining 15 days.”
“If you have two projects in the queue, and one says they can be flexible and the other says they can’t be flexible, and they’re about the same size, then the one that can be flexible is more likely to be successful,” Tom Wilson said.
That’s not lost on Brian Janous, cofounder of Cloverleaf Infrastructure, which develops what the company describes as “clean-powered, ready-to-build sites for the largest electric loads,” mainly data centers.
“You need to understand, when a utility says, ‘I can’t get you power,’ what they mean is, ‘There are certain hours of the day I can’t get you power,’” he said. The data center industry “lacks visibility into this, which is kind of shocking,” given that data center flexibility is nothing new.
In fact, back in 2016, when he worked as energy strategy director at Microsoft, Janous helped structure a deal for a data center in Cheyenne, Wyoming, to use fossil-gas-fired backup generators to reduce peak grid stress for utility Black Hills Energy. That promise, combined with Microsoft’s agreement to purchase nearly 240 megawatts of wind power, got that deal over the line.
Janous thinks many utilities are eager for similar propositions today. One unnamed utility executive told him recently that the backlog for connecting large data centers to its grid is now at least five years. “I asked, ‘What if the data center could be dispatchable?” And he said, ‘Oh, we could connect them tomorrow. But nobody’s asking me that.’”
Getting utilities and data centers together to ask those kinds of questions is what DCFlex is all about. Project partners are now developing five to 10 demonstration projects, Tom Wilson said, none of which have been announced. But he described the scope of work as ranging from the development stage to “existing sites that are ready to roll.”
As for how these projects will help the grid, he laid out two broad methods: They’ll use on-site power generation or storage to replace what they’d otherwise pull from the grid, or they’ll use less electricity during peak hours.
Janous thinks on-site generation is the simpler approach. To some extent, it’s already happening today but typically with dirty diesel generators. Janous, Tom Wilson, and other experts say these diesel generators are not a viable option for hyperscalers, however. They’re simply too dirty and too expensive to rely on, except during grid outages or other dire situations.
Biodiesel and renewable diesel could work for some smaller data centers, Tom Wilson said. But it’s not yet clear whether air-quality rules would permit generators burning those fuels to run during nonemergencies. Nor are the economics viable for larger-scale data centers, he said.
Fossil-gas-fired backup generators like those Microsoft used in Cheyenne are another option — albeit one that still pollutes the local air and warms the planet. Still, a growing number of data center developers are looking to use them as a workaround to grid constraints. “We’re in the process of developing sites in many parts of the country,” Janous said. “Every one of them has access to natural gas.”
It would be a problem for the planet — and for meeting the climate goals major tech companies have committed to — if data centers planned to use fossil gas for a majority of their power. But if relied on sparingly and strategically, this choice might be less harmful than the alternatives: If a data center burns fossil gas just to power itself during grid peaks, that might reduce pressure on utilities to keep old coal-fired power plants open or to build much larger gas-fired plants that would lock in emissions for decades.
Other gas-fueled options for on-site power might be less damaging to the climate — although this remains a hotly debated topic. Microgrid developer Enchanted Rock plans to install gas generators at a Microsoft data center in San Jose, California, which will burn regular fossil gas but will offset that usage by purchasing an equivalent amount of “renewable natural gas” — methane captured from rotting food waste. It claims this will make the project emissions-free.
And utility American Electric Power has signed an agreement to buy 1 gigawatt of fuel cells from Bloom Energy, which it plans to install at data centers. The firm’s fuel cells still emit carbon dioxide, but they avoid the harmful nitrous oxides caused by burning gas.
Batteries are another option — and the one that has the greatest potential to be clean. Most data centers have some batteries on-site to help computers ride through grid disruptions until backup generators can turn on, but relying on batteries for backup power and to provide grid support is a far more complex and costly endeavor.
“There are all kinds of trade-offs in terms of reliability, in terms of emissions, in terms of cost, in terms of the physical footprint,” Tom Wilson said. “In a storm situation that brought the grid down, you’d want something that can be dependable.”
A small but growing number of hyperscalers are looking to batteries for both backup power and grid flexibility. Google’s battery-backed data center in Belgium is one example.
“We built out battery storage as a way to displace part of our diesel gensets, to provide grid services, and to provide relief during times of extreme grid stress when we needed backup,” said Amanda Peterson Corio, Google’s global head of data center energy.
In the U.S., a Department of Energy grant is supporting a battery installation at an Iron Mountain data center in Virginia that’s meant to test the potential to store clean power for backup and grid-support uses.
It’s one of a number of DOE programs launched under the Biden administration whose purpose is to explore ways that “battery energy storage systems can provide similar levels of reliability, and without a lot of the challenges that diesel gensets or other backup power sources have,” Avi Shultz, director of the DOE’s Industrial Efficiency and Decarbonization Office, said in an October interview.
The other big idea for making data centers flexible focuses not on the power they can generate and store but on the power they use, Shultz said.
“Demand response” is the utility industry term for the practice of throttling power use during times of peak grid stress, or of shifting that power use to other times when the grid can handle it better, in exchange for payments from utilities or revenues in energy markets.
Historically, data centers haven’t been interested in standard demand-response programs and markets. The value of what they do with that electricity is just too high compared with the potential rewards.
But if a data center’s participation in a demand-response program is the difference between it getting a grid connection or not, the programs become a lot more appealing.
Shultz highlighted two key data center tasks that are particularly ripe for load flexibility.
The first is “cooling loads and facility energy demand,” he said. Data centers use enormous amounts of electricity to keep their servers and computing equipment from overheating, and quite a bit of that energy is lost in the process of converting from high-voltage grid power to the low-voltage direct current that computing equipment uses.
Data center operators have invested heavily to make this cooling and power conversion more efficient in recent decades. Further advances in efficiency — and technologies that can store power for cooling for later use — could become “part of the routine best practices of data center operations,” Shultz said.
The second big target for load flexibility is “the core of the computational operation itself,” he said. Not all data centers need to run their computing equipment 24 hours a day, and “they may not be being operated in a way that’s optimized from an energy point of view. I think there’s an opportunity there to develop more innovative and flexible operational processes.”
Again, this isn’t a new idea. Cryptocurrency mining operations in Texas have been earning millions of dollars for not using electricity during grid emergencies, even as lawmakers, regulators, and neighbors of crypto mining operations have been raising alarms over that industry’s rising hunger for power. Google has also carried out some version of this through its “carbon-intelligent computing” program, which shifts certain data center operations to use cleaner power. In recent years, Google has touted how it can also shift computing load to relieve grid stress.
“Our carbon-aware computing platform started as how we can shift nonurgent compute loads to times when the grid is more clean,” Google’s Peterson Corio said. “We’ve also done that to support utility partners in times of extreme weather events.”
But last year’s DOE report on data center power use stated that aside from Google’s activities, contributors to its research “identified no examples of grid-aware flexible operation at data centers today” in the U.S. That absence of evidence “may result from the fact that electricity providers only recently started having to say no to data center interconnection requests.”
Plus, not every data center conducts tasks that can be easily postponed, Tom Wilson emphasized. “Data centers are all different. If you are Visa, in Virginia, their data center is transacting, I think, 80,000 credit approvals per second. You don’t want to say, ‘I’ll approve your credit when the sun is shining.’ That’s the kind of thing where you don’t have much flexibility. You need to deliver.”
Other tasks are better suited, he said, including, critically, much of the work of training the AI models that constitute the largest single source of increasing power demand from data centers. “There is potentially flexibility there — and a fair amount of it.”
Taking advantage of that flexibility is the idea behind Verrus, a company launched last year by Sidewalk Infrastructure Partners, a firm spun out of Alphabet. Traditional data centers separate their computing capacity into individual “halls,” each of which has its own power conversion, cooling, and backup generation. Verrus, by contrast, is planning data center complexes with “a centralized battery in the center of all the data halls, with a sophisticated microgrid controller that allows it to think about all the data halls as interruptible and schedulable,” Jonathan Winer, a Sidewalk Infrastructure Partners cofounder and the former co-CEO, said in an interview last year.
Not every “hall” will be dedicated to tasks that can be interrupted, he explained. But some of them will be — and “with AI training, I can press pause on it. These are multi-day, sometimes multi-week training runs, and you can press pause and resume.”
Verrus hasn’t built one of these data centers yet, but it is targeting Arizona, California, and Massachusetts as its first markets. In a white paper last year, Sidewalk Infrastructure Partners explained why this kind of flexibility is a must for new data centers. As Winer said, “The data center industry has realized how much it needs to be a power-led model.”
It’s not clear how many data center developers are building flexibility into their grid interconnection requests to utilities. But that doesn’t mean it’s not happening.
Data center developers tend to be secretive about their hunt for sites and the nature of the discussions they have with utilities. Developers may be talking about flexibility with one utility while also shopping around for a more traditional interconnection that gives them access to the power they want at all hours of the day.
Flexibility deals are most likely to emerge in data center markets where unimpeded interconnections simply aren’t possible anymore due to grid constraints, said Aaron Bilyeu, Cloverleaf’s chief development officer. What’s changed is that “we’ve quickly run out of those opportunities,” he said.
Utilities in grid-constrained parts of the country are beginning to take on the complexities of creating “flexible interconnection” policies and tariffs — the rules and rates for customers — that could provide data center developers some clarity on what a commitment to flexible operations could be worth to them. That’s a big part of the work underway at DCFlex, Tom Wilson said.
“We have three main workstreams. The first is aimed at defining the flexibilities that are possible and creating a taxonomy — each utility and RTO [regional transmission organization] has different words for the same things,” he said. “The second piece is aimed at incentives, rate structures, and regulatory issues on the utility side, looking at how they could effectively orchestrate flexibility.”
“The third piece is how to build the planning and operational tools that incorporate flexibility,” he added. That means figuring out how to calculate the impact of flexible versus non-flexible large customers on long-term planning for power plants, grids, and other infrastructure investments, which traditional planning processes don’t do today. “You need new tools, or to evolve tools, to think about and utilize the new opportunities that flexibility can provide.”
That work will be critical to giving regulators the tools they need to challenge utilities’ claims that they must build new fossil-gas power plants and keep coal plants open to serve peak loads, the Sierra Club wrote in a September report. Big customers “can have an outsized impact in avoiding new fossil fuel investment (or enabling coal retirements) by participating in demand management programs that allow utilities to subtract some or all of the customer’s load from its peak obligation,” the report notes.
That aligns with the needs of tech companies like Amazon, Google, Meta, and Microsoft that have aggressive clean energy goals. These companies are negotiating with utilities in hot data center markets like Georgia, arguing that letting data centers build clean energy, on-site power, and load flexibility into load forecasts could obviate some of the new dirty power that utility Georgia Power wants to build.
Putting the question of clean versus dirty power aside, data center developers can’t expect utilities and regulators to allow the costs of supplying them with round-the-clock power to fall on regular customers’ shoulders, Bilyeu said. “We believe data centers should pay their own way.”
.svg)
