The pitch for data center flexibility is a relatively simple one: If artificial intelligence load can flex, AI growth can accelerate without overwhelming the grid. In the last year, startups and hyperscalers alike have begun pitching their vision of massive AI loads that ramp up and down, act like giant batteries, or disconnect from the grid entirely.
So when PJM’s independent market monitor poured cold water on the excitement — publishing a report last week that frames large-load flexibility as an expensive and unproven reliability gamble — the reaction was harsh and swift. Critics accused the IMM, Monitoring Analytics, of misunderstanding or even opposing demand response, and of missing the point on data center flexibility proposals.
But analysts told Latitude Media that the IMM, tasked with ensuring PJM’s wholesale electricity markets are fair and efficient, wasn’t necessarily attacking flexibility in principle. Instead, the report exposes a gap between the way flexibility exists today, at least in PJM, and the way proponents envision it working in the near future.
Paradoxically, that gap may move the conversation forward. By exposing where PJM’s frameworks fall short, the IMM report could serve to push industry and regulators toward more concrete, enforceable solutions on what has become “a hot button issue,” explained Joe De Lia, VP of power markets at Orennia.
The IMM’s central warning is that the cost of getting flexibility wrong is extremely high. If new load promises flexibility but then fails to curtail during grid emergencies, PJM will need to buy additional backup capacity to cover the gap, driving up market prices and, ultimately, consumer bills. In PJM, flexibility means voluntary, emergency-only demand response. PJM can ask loads to turn down when needed but can’t physically shut off a customer’s power.
“PJM is still trying to get their handle on how these loads will operate…and there just isn’t necessarily that operational data to back up that these loads are curtailable yet,” De Lia said. “There has to be a way for [data centers] to prove to an ISO…that they’re actually able to curtail.”
A PJM problem
For those working to build and implement flexibility, the news that market design is key to scaled deployment comes as no surprise. In PJM, treating a large load as a dispatchable resource, able to ramp consumption up and down, isn’t possible, explained energy market economics expert Arushi Sharma Frank, who is a senior advisor to flexibility startup Emerald AI.
Frank recently filed a proposal with ERCOT to establish a new category of resources called Controllable Load Resources: large loads controlled like power plants with strict, realtime commands and onsite flexibility tools. ERCOT is well positioned to leverage data centers as dispatchable resources, Frank said, because the market can send real time, enforceable dispatch instructions to specific loads at specific locations. ERCOT also already has mechanisms for real-time telemetry, performance monitoring, and penalties if loads don’t comply.
PJM just isn’t ready for that type of setup. Current demand response programs in that market are made up of groups of smaller customers lumped together, and PJM doesn’t have a way to control or send realtime signals to a large load at its exact location on the grid. PJM can ask groups to cut back power use during grid emergencies, but can’t enforce precise load reductions at individual data center nodes, Frank explained.
“From where the monitor sits, it is not credible to assume that the same asset will quietly act as a deeply reliable, frequently interruptible resource whenever PJM is short on capacity, Frank wrote on her Substack this week. The report, she added, “is a hard indictment of the status quo, and I think that’s appropriate.”
Changing market structure to enable dispatchability won’t happen overnight in PJM, she acknowledged, but the report diagnoses real problems with current demand response products in the market, which is an important step.
“From my vantage point, dispatchability at the load’s site controller is an essential element of getting this right,” Frank wrote. “If you recognize that the data center can arrive with its own flexibility toolkit…and expose that to PJM as a dispatchable resource, then it stops being just a giant new load. It becomes part of the reliability toolkit, as long as the RTO can see it, dispatch it, and hold it accountable.”
Getting it right
The flip side to the report’s explanation that getting flexibility wrong will be costly is that, according to the same modeling, getting it right would be immensely beneficial.
As part of its more than 900-page analysis, Monitoring Analytics, which was established as the independent external monitor for PJM in 2008, uses the scenario of adding 20 gigawatts of data center load to the grid to assess future costs in the capacity market.
That modeling shows that the cost of integrating new data center load swings massively depending on whether a small fraction declines to participate in demand response.
If 100% of that new data center load agrees to act as demand response, the extra cost for existing consumers would still be around $396 million per year, the report found. However, if just 10% of that new load opts out of flexibility, those extra costs jump up to $5.48 billion, assuming the resulting shortfall pushes PJM’s capacity market into scarcity pricing — a more than 1,000% increase. It’s essentially a measure of how expensive inflexibility is: even a small wedge of inflexible load drives cost into the billions.
As the IMM put it in its report, there is not now and not likely to be in the near future sufficient capacity supply in PJM to meet large data center load. “The solution is not to create reliability issues and wealth transfer issues by clearing the capacity market at the maximum price and at a quantity less than the reliability requirement by allowing the ongoing interconnection of large data center loads without adequate generation to serve them.”
The only form of “flexibility” that the IMM endorses as ready to deploy, is AI load that brings matching new generation with it, matched in location and timing to the data center profile. This “bring your own generation” approach is being endorsed by a handful of governors and the Data Center Coalition as part of PJM’s Critical Issue Fast Path stakeholder process, expected to wrap up this week.
The “gridlock” over flexible large loads is partly the result of the measured pace utilities and ISOs take to these types of technological changes, explained Julia Hoos, who covers the eastern U.S. at Aurora Energy Research.
“The IMM is rightfully raising that we haven’t solved how to serve data centers in isolation,” Hoos said. It’s a critical conversation, she added: “Nobody thinks about PJM until the lights go out or the costs skyrocket, and that’s what’s been happening this year…We tend to forget that there is no way to supply data centers with power without impacting everyone else that relies on the grid.”
Many of the next-generation designs, including for co-located generation, islandable microgrids, and workloads designed for rapid curtailment, are largely conceptual. “This fully islanded large load just doesn’t exist today,” Hoos added. “That’s where the primary tension is, that a lot of these plans for new large load are developing projects that are untested, and they’re asking the grid operators to trust them that things will work out as planned.”
