Ep. 026 - PJM's $12B Modeling Mistake Is Hitting Ratepayers Again (Datacenter, Energy)

Summary

• SemiAnalysis’s core call is that PJM’s capacity-auction modeling—not data center demand alone—is what’s driving the region’s electricity-price spike, with $12B of the $63B spent across the last four auctions estimated as avoidable. All $12B came from the first two auctions ($7B for 2025–26, $5B for 2026–27); price caps, pushed primarily by Governor Shapiro, limited the effect of the modeling adjustment in the later ones. Boswell’s framing: “that’s due to market design or not really even market design—more like sort of bureaucratic auction design.”
• The mistake is two-sided: PJM constrains supply while inflating demand. Auctions now run on one-to-two years’ notice instead of roughly three, so no new plant can enter and the supply curve “goes vertical”—then PJM overbuys because it models the fleet “as if it’s summer all year round,” ignoring that denser winter air makes gas turbines materially more efficient, and rates winterization on backward-looking failure history despite post-Storm-Elliott federal orders and local mandates. “Past performance is not a predictor of future performance.”
• The multiplier comes from PJM’s unusual pay-as-clear design covering the entire existing fleet, not just new capacity. Every plant—even fully paid-off ones—gets the marginal clearing price: bid in at $5/MW-day and you still collect $300 if that’s what the last unit needs, so a tiny scarcity move at the curve’s edge can reprice bills across a 66-million-person grid.
• The near-term event is an emergency auction this September or October in which PJM’s board overrode its own members’ two-thirds approval. PJM models a 6.8 GW shortfall for 2028–29; Boswell’s estimate, after accounting for cold-weather performance and winterization, is that only about 3 GW would be needed. The board’s design has PJM signing contracts out to 2042–43 at a $555/MW-day cap—potentially as much as $21B over 15 years—expecting to pass the liability to new demand, including data centers. If they don’t show up, “PJM is stuck holding the bag, and PJM can’t hold the bag because it doesn’t have its own money”—existing ratepayers do.
• Boswell agrees with the FERC chair’s “too big to function” characterization as a governance and regulatory-capture problem, not a problem with grid scale—and estimates something like 40% of America’s first-tier electricity-price debate stems from PJM alone. Over 1,000 members in five voting groups, footprint-wide deliverability studies, “copper plating” overbuild, and rules written by incumbent generators earning windfalls off constrained supply are central to the problem.
• The tradeable silver lining: the current high-price auctions only cover their respective years and run out partway through 2029, while data centers could structurally lower rates, not raise them. Electricity is an extreme-fixed-cost commodity where load growth spreads costs; data centers represent “by far the largest differential between cost of electricity in versus value out that has ever existed” and would happily overpay to get built. Boswell points to batteries and behind-the-meter power but is “gently skeptical” of data center demand flexibility.

Deep dive

1. PJM’s bureaucratic auction design, not data center demand alone, is the price story—and $12B was avoidable

• Boswell’s scene-setting: PJM (nominally Pennsylvania–New Jersey–Maryland, actually 13 states plus DC, 66 million people) is America’s largest grid by electricity delivered and “data center alley”—and almost every article claiming data centers raise electricity rates is specific to PJM. “I always check this whenever I see that data centers are leading to increased electricity rates. I look at what the source is, and it’s almost always specific to PJM.” His diagnosis: “that’s due to market design or not really even market design—more like sort of bureaucratic auction design.”
• The headline number: of $63B spent across the last four capacity auctions, SemiAnalysis estimates $12B was avoidable—$7B in the 2025–26 auction and $5B in 2026–27. Price caps, pushed primarily by Governor Shapiro along with other political pressure, limited the effect of modeling differently in the later two; Boswell says the caps were “probably the right thing to do given the other constraints.”
• The structural quirk: unlike the usual forward auctions for new capacity needing long-term certainty, PJM auctions the entire system—fully paid-off plants included. “Any price increase has a massive multiplier effect… a small price increase on the Y-axis that’s multiplied across the entire X-axis.”

2. The modeling mistake: endless summer and a stick with no carrot

• Supply is artificially constrained: PJM delayed auctions and now runs them on one-to-two years’ notice when roughly three years is already tight—“just not long enough for a new power plant to get built”—alongside a slow interconnection queue. The supply curve shifts left and goes vertical, so any demand move produces scarcity pricing.
• Demand is simultaneously inflated by two rating errors. First, PJM models plants “as if it’s summer all year round”—but winter air is denser, so gas turbines run cooler and move more mass. “As much as endless summer would be a lot of fun, it’s simply not the case… winter is coming.”
• Second, winterization. After the 2014 polar vortex and Winter Storm Elliott in 2022—when components froze or pipelines cracked, including components that failed within their operational margins—federal orders, local mandates, and fines pushed plants to weatherize. Yet PJM rates reliability purely on past failures extrapolated forward: “as any good financial professional knows, past performance is not a predictor of future performance.” Boswell’s fix: reward winterized plants with higher ratings—introduce a carrot alongside the stick—and stop overbuying capacity the system already effectively has.

3. How the money actually flows: pay-as-clear on megawatt-days

• The unit is capacity, not generation—“it’s like paying a farmer for having a field as opposed to paying them for the wheat that comes from the field.” PJM draws a textbook demand curve (willing to buy roughly 170 GW if cheap, with a price ceiling if dear), bidders’ offers form the supply curve, and the intersection sets price and quantity.
• The kicker is that everyone receives the clearing price: “if you bid in at $5 a megawatt-day… you will still get paid $300 a megawatt-day if that’s what the last unit needs.” That’s why a handful of marginal megawatts at the vertical end of the curve can swing total auction cost by billions.
• Ratings determine volume, hence revenue: a 1,000 MW nuclear plant might be rated at roughly 90% or 93%, versus roughly 80% for a gas plant. Boswell’s proposal is that a well-winterized, well-sited gas plant—short pipeline, good design—could be rated at 85–88%, bid in more capacity, and collect more: “short of violation, it’s just revenue. It just comes straight in the door.”

4. The emergency auction: the board overrules its own members, and ratepayers may hold the bag

• PJM models a 6.8 GW shortfall for 2028–29, partly a price-cap artifact; Boswell estimates that accounting for the cold-weather uplift and winterization would mean needing only about 3 GW—more than half the procurement, and its cost, potentially avoidable.
• The governance twist Nanos pulls out in disbelief: PJM’s more-than-1,000 members, split into five voting groups needing the equivalent of 66% to pass a measure, agreed on an auction design without this problem—and the board overrode them. Boswell’s dry answer: “Not everything is good to be run as a democracy, right?” But “if our modeling is correct and they don’t need that much power, then they didn’t need to overrule the membership.”
• The board’s design has PJM signing contracts out to 2042–43—long-term contracts being good for new capacity—but without necessarily having a counterparty, at a $555/MW-day price cap that could amount to as much as $21B over 15 years for the amount being contracted. PJM expects to pass the cost to new demand, including data centers, but that load may not appear. “There’s a very real possibility that data centers don’t turn up… and PJM can’t hold the bag because it doesn’t have its own money”—it only has ratepayers’ money. The auction is expected in September or October.
• The membership-supported alternative, which Boswell prefers, was for data centers to take the liability directly—“show up, sign on the dotted line, say I need this amount”—and bring their own power or contract for it. The board overrode that approach because it thought the membership-backed alternative would not procure enough power; Boswell’s modeling argument is that the extra procurement was unnecessary.

5. “Too big to function”: regulatory capture, and why data centers should be making power cheaper

• Boswell agrees with the FERC chair’s phrase “too big to function”—meaning governance, not geography: “being big is good if you’re an electricity grid” given scale economies, but PJM “can’t navigate its own systems of governance.” He also flags footprint-wide deliverability studies—can a Virginia plant reach Illinois?—and possible “copper plating” overbuild, hedged as something he is “not certain about, but I’ve seen some data,” drawing the parallel to Britain’s own overbuilt system.
• The capture mechanism, from his British experience with industry codes: “the rules of the system are written by the players.” Incumbents earning windfalls off constrained supply can run rings around regulators, and Boswell admits “the bit I really don’t have a good grasp on… is the deep politics of this.” PJM itself is “really just a stack of contracts”—three unbelievably complex ones—and about 10% of the system, including vertically integrated utilities, does not participate in the auction.
• The through-line: data centers should make electricity cheaper. Grids are extreme-fixed-cost commodities with huge scale economies, and data centers represent “by far the largest differential between cost of electricity in versus value out that has ever existed in an electricity system”—meaningfully committed players bringing solutions to meetings, whose alternative is “either we figure this out or I’m moving to Texas.” You could even make them overpay: “you just have to pay more than your fair share because you’re making absolute bank”—and they might accept it to get built.
• Nanos’s forced happy ending lands on a real point: the current high-price auctions only cover their respective years and run out “partway through 2029,” while, if data centers make their case, electricity load growth is “by far the best way to bring down electricity prices”—via batteries, behind-the-meter power, and flexibility, though Boswell stays “gently skeptical of demand flexibility at data centers.” His estimate is that something like 40% of the first-tier U.S. electricity-price debate stems from PJM, with arguably much of the rest downstream of it.