The number that should be pinned to every utility commission wall in the thirteen-state PJM footprint: $29.4 billion. That's the combined total increase in capacity market revenues across PJM's last four auctions attributable to existing and forecast data center load growth, according to PJM's independent market monitor. Not the total cost of serving data centers. Just the increment — the extra billions layered onto what ratepayers were already paying before hyperscalers started treating northern Virginia like a land grab.
This is what civilizational acceleration looks like when the grid hasn't been built for it. It's spectacular and it's brutal and it demands we look at it clearly, because the next phase — PJM's proposed fix — carries its own set of risks that the grid-optimist crowd (myself included) can't afford to wave away.
I covered the capacity auction mechanics back in September in the piece on how AI data centers are repricing the entire grid. The structural picture has since clarified substantially. The market monitor's first-half 2026 report has filled in numbers that were projections back then. Let's run through what we now know and what PJM is actually proposing to do about it.
The Auction Spiral in Real Numbers
Start with the price trajectory, because it's genuinely staggering when laid out in sequence.
The 2024/2025 PJM capacity auction cleared at $28.92 per MW-day. Within one auction cycle, the 2025/2026 auction rocketed to $269.92 per MW-day — an 833% surge that pushed total market capacity costs from $2.2 billion to $14.7 billion. Then the December 2025 auction for the 2027/2028 delivery year hit the FERC-approved price cap of $333.44 per MW-day for the third consecutive year — a 1,053% increase from the 2024/2025 baseline. Total cost for that auction: $16.4 billion.
Who's paying? Everyone, but not equally. Data centers account for $6.5 billion — 40% — of that $16.4 billion total. Here's the maddening part: of that $6.5 billion data center share, $6.2 billion is attributable to data centers not yet built. Hypothetical demand from facilities that haven't broken ground is already flowing through into ratepayer bills today. Washington D.C.'s Pepco customers saw $10 per month added to their bills from the 2025/2026 delivery year alone. Projections from the Natural Resources Defense Council suggest that by 2028, average PJM households could be looking at roughly $70 per month in increases compared to pre-surge levels.
By mid-2026, the overall wholesale power bill in PJM had surged 46% to $56.7 billion through July, up from $38 billion in the same period the prior year. Data center load, through PJM's capacity market, accounted for 9% of the wholesale price of power — $10.48/MWh of a $116.53/MWh average. That figure doesn't include the data center-related increases flowing through energy prices or transmission costs separately.
The pattern is unambiguous. Load projections for data centers are being priced into the capacity market before the physical generation exists to serve them. Demand is arriving faster than supply can respond. And the gap isn't closing — PJM projects data center and large-load demand could increase by as much as 70 GW by 2038.
The Fix PJM Is Proposing — and Why It's Complicated
PJM didn't just show up to FERC with a complaint. It submitted a two-part plan in late July and early August designed to address the shortfall before it becomes a reliability emergency.
The headline mechanism is a one-time reliability backstop procurement auction — a targeted procurement of approximately 6 GW of new capacity, the amount PJM estimates is needed to restore its reliability position to breaking even. That auction — scheduled from September 30 to October 21 — will offer 15-year capacity contracts for new power plants, covering natural gas generators, nuclear facilities, clean energy projects, battery storage, and uprates to existing resources.
The battery storage element is real and interesting. Battery energy storage systems are expected to have a competitive edge in the backstop auction specifically because of their construction timelines — they can come online faster than combustion turbines or nuclear uprates. EQT Infrastructure's backing of a $2 billion small-battery push through Madison Energy Infrastructure — targeting 1 GW of capacity additions by 2028 starting in the PJM footprint — represents exactly this logic playing out in private capital markets. Four-hour batteries at industrial and commercial sites aren't a full grid solution, but they're grid-adjacent resources that can compress interconnection timelines.
The second part of PJM's plan is the political pressure point: cost assignment. PJM wants utilities to ensure that data centers pay the capacity costs their load growth creates, rather than those costs being spread to residential and commercial customers. State regulators become responsible for enforcing that distinction. This is the mechanism that, if it actually works, changes the economics of data center siting. If hyperscalers have to internalize the full marginal capacity cost of their load — rather than socializing it across 65 million ratepayers — the incentive to co-locate with existing generation or invest in demand flexibility changes overnight.
FERC's June Order Changed the Rules of the Game
The backstop auction and cost assignment framework don't exist in isolation. On June 18, 2026, FERC issued show-cause orders to all six Regional Transmission Organizations and Independent System Operators — PJM, MISO, SPP, CAISO, ISO-NE, and NYISO — directing them to justify their current rules for large-load interconnection and tariffs or propose reforms. The order set 30-day and 60-day response windows; both had elapsed by early September.
What FERC was actually asking about matters for how we read PJM's backstop proposal. The June order focused on five categories: transmission service applications and study processes, cost assignment transparency, co-location rules, services for flexible large loads, and study processes for generation serving electrically proximate loads. The FERC order didn't prescribe efficiency standards for server racks or cooling systems. It asked grid operators whether their interconnection tariffs can handle larger, faster-moving electricity requests without shifting costs to other customers or degrading reliability.
That's the regulatory layer beneath everything PJM is doing. The backstop auction is PJM's operational answer. FERC's proceeding is the structural answer. They need to work in concert — and the history of FERC proceedings suggests the structural answer arrives on a timeline measured in years, not months.
Meanwhile, the physical problem compounds quarterly. NVIDIA's Blackwell architecture pulls between 120 kW and 132 kW per rack — roughly ten times the density of legacy enterprise servers. A single next-generation AI training cluster can require up to 300 MW of continuous power. The grid was not designed for this load profile, and the interconnection queue wasn't designed to process this volume of large-load applications simultaneously.
The Demand Flexibility Bet: Necessary but Not Sufficient
There's a third element emerging in parallel with the supply-side solutions: organized demand flexibility at the data center level. Nvidia, Google, and data center software startup Emerald AI have launched the AI Energy Management Alliance, described as the only trade group "laser focused on flexible AI data centers." The pitch is that AI workloads — particularly training runs and inference batches — can be scheduled around grid stress events, turning data centers from dumb constant loads into intelligent grid participants.
I want this to be true. The math of the capacity auction problem suggests it had better become true, fast. If 70 GW of new data center demand lands on the PJM grid by 2038 as projected, and even 20% of that load can flex — respond to price signals, shift to off-peak hours, curtail during grid stress — the capacity procurement math changes meaningfully. A data center that bids into a demand response program isn't drawing on grid reserves during a heat wave; it's providing a grid service.
The honest caveat: there's a gap between announced alliance formation and operational demand flexibility at scale. Consortium announcements and megawatt-scale curtailment agreements are different things. Until there's grid-operator data on how many gigawatts are actually enrolled in demand response programs — and how they perform during actual stress events — the demand flexibility narrative is still a forecast, not a fact.
What we can say is that the economics are now compelling enough to force the issue. When capacity prices hit $333.44/MW-day and you're a hyperscaler running hundreds of megawatts, the cost of not enrolling in demand response programs is measurable on your income statement.
What Comes Next: The Decisions That Matter
The backstop auction closes October 21. What clears it — and at what price — will tell us whether the supply response is actually keeping pace with demand growth, or whether we're about to watch another
