This week has been all about the debate unfurling over how to redefine reliability for the AI era, and PJM has been at the center. Over the weekend, a months-old regulatory capacity accreditation ruling collided with a fresh report from the region’s Independent Market Monitor that dismissed data center flexibility as “regulatory fiction,” sparking a debate over whether “flexible load” can be a verified grid asset — or simply a regulatory loophole.
The two separate-but-related conversations converged in the energy corners of social media, where proponents of demand response found themselves defending the reliability of virtual power plants against critiques specifically aimed at hyperscaler loads.
Much of the debate conflates two arguments made by the IMM at different points this year. Back in April, Monitoring Analytics pushed back against PJM’s proposal to increase the accreditation value of demand response resources for 2027-2028, arguing the change was based on “a combination of speculation and incorrect ELCC modeling of demand response.”
FERC disagreed and approved the proposal anyway, and in August PJM published updated ratings crediting demand response at 92% — a 23% bump from prior years, and a higher rate than every other resource except nuclear. Importantly, that doesn’t mean PJM expects demand response resources to dispatch the same number of hours as a nuclear plant. Rather, it’s an indicator that PJM believes that during key hours when the grid is stressed, demand response is likely to be available around 92% of the time.
Then, this month, in its regular State of the Market report, the IMM sharply criticized the rapid addition of large data center loads that agreed to be flexible, arguing the only way to ensure system reliability is for those loads to bring their own matching generation. Despite the hype, it said, AI data centers can’t be relied on to turn off when the grid is stressed, and PJM doesn’t have the ability or authority to force them to throttle. Allowing such massive loads to skip the line if they promise flexibility — as proposed by some stakeholders as part of PJM’s Critical Issue Fast Path process — would cost ratepayers billions of dollars.
The two debates are distinct, and the first was essentially closed before its revival on social media this week. PJM believes demand response works and intends to leverage it, as evidenced by the updated rating for the upcoming 2027-2028 capacity auction.
Still unresolved, though, is whether individual data center loads are as reliable as aggregated grid resources. A vote earlier this week failed to pass a single one of the 12 separate proposals in the Critical Fast Path process — including a proposal that would require data centers to bring their own matching capacity, as the IMM indicated, and another that would have required data centers to ramp down first during emergency periods.
The 24/7 trade off
PJM’s demand response programs currently have significant constraints, and are only able to participate in the market during certain hours, like hot summer afternoons or winter mornings. Demand response is cut off at 9pm in the winter, even as PJM’s risk modeling shifts. The effect is that even in the event of a capacity shortfall at 9:30pm, demand response has no obligation to show up.
That will change in the 2027-2028 delivery year, when demand response participation will expand to 24/7/365, allowing virtual power plants and other demand response resources to respond any time, not just during peak hours.
It’s a change that PJM agreed to back in May, following advocacy from the demand response industry. In its filing, PJM explained the move sets a requirement for all demand response resources “to be available for all hours of an operating day and imposes a performance obligation.”
It’s that expansion that led to the increase in demand response’s ELCC rating, explained Dana Guernsey, co-founder and CEO of Voltus. Prior ratings were assuming zero availability for the hours in which VPPs weren’t even considered for participation, despite the data showing their load was available, she said.
Voltus and others are already frequently ramping loads, including large loads of 50 MW facilities or greater, in response to PJM price signals in both the day-ahead and real-time markets, Guernsey added. They’re not just emergency dispatches — which Guernsey describes as an “on/off switch” compared to the “dial” approach of ancillary services — that Voltus is dispatching every day. Their portfolio includes loads that are well-suited to ramping capabilities, like battery storage, building control systems, and cryptocurrency mining. Aggregations too can ramp up and down, with smaller assets coming in and out as Voltus’ platform orchestrates them, Guernsey added.
That’s Guernsey’s main addition to the flexibility debate: Demand response has already reached scale and is being delivered today, in PJM and elsewhere around the country. Critically, she added, participation isn’t voluntary once committed to the market. There are significant penalties for failing to meet performance obligations; it’s a significant operational liability.
Which brings the conversation back to the criticism outlined in the IMM’s report on flexible data centers. The issue, Guernsey said, isn’t whether large loads can flex; Voltus’ daily dispatches and PJM’s upgraded capacity rating show that it can. Instead, it’s whether it makes economic sense for AI data centers in particular to do so, whether they are obligated to do so by regulation — and if so, how exactly.
“In my experience, data centers would rather install storage or distributed generation than actually stop computing or risk too much increased latency,” Guernsey explained. Demand response portfolios can technically enroll AI data centers, she added, and Voltus has built out the capabilities to ramp them in the markets, but decisions for those particular loads are based more on economic incentives. “We’ve shown it’s possible, but often data centers will decide the compute is more valuable, and turn to other things we can help them with,” she added, pointing flexing load instead with onsite generation and storage, and bilateral agreements to contract flexible megawatts.
The option to bring your own
Efforts to obligate data centers to participate in order to get online early have faced pretty significant pushback. Earlier this year PJM proposed its “non-capacity backed load” initiative, which would have required new customers over 50 MW — namely data centers — to agree to be first in line to have their power cut off during shortages in order to get connected to the grid. Under the NCBL proposal, data centers would have been called upon even before traditional demand response resources.
That plan met significant opposition from stakeholders including utilities and data center companies, and PJM ultimately dropped the initiative in October.
But to Guernsey, the IMM proposal to require large loads to build physical power plants, and effectively island themselves with onsite power, is an economic mistake. Doing so means the grid loses out on the efficiencies of shared resources, she explained.
That’s the point Voltus is trying to hammer home in its comments to FERC regarding broader interconnection reforms: “I totally agree, more generation will need to be built, but if you island everything, and all of the capital goes towards everybody building their own microgrid…it’s always been the case that if you can provide benefits to the broader grid, that’s better,” she added.
Guernsey pushes back against the IMM’s insistence that reliability must be physically co-located behind the data center’s meter to count. “Is it 100% as valuable to have it one zip code over? Maybe not in all situations,” she conceded. “But it’s definitely not zero.”
Voltus is pitching an alternative path it’s calling “bring your own capacity.” In that model, rather than a data center promising to self-curtail or building a gas plant, the data center instead pays an aggregator (like Voltus) to bundle verified reductions from the surrounding grid.
That’s a model that provides grid resources today, regardless of whether it turns out that new AI data centers are willing to be flexible or not, Guernsey added. (The jury is still out on that, in her eyes.)
Ultimately, she said, the friction, failed votes, and sparring on the internet are all an indication that the system is undergoing an evolution forced by the AI boom — one that may ultimately push the market toward demand response.
“I’m actually really grateful for all the data centers, because it’s been this huge forcing function,” she said. “It’s not just about data centers, but data centers are helping us see…all of the things that aren’t working so great if you apply a lot of load.”
