Sam Altman, Dario Amodei, and Elon Musk need sign-off from fewer than a dozen board members to commit their companies to multibillion-dollar moves. The power plants that supply their data centers need sign-off from 13 states (plus D.C.), thousands of generators, millions of customers, and a federal regulator whose ratemaking standard predates the personal computer in order to build anything new.
Everyone in tech knows about the CEOs of the foundational artificial intelligence labs. Only energy nerds know the names of the people running our grid operators. That anonymity is a feature, not a bug. Grid operators generally think in decades, not years. But right now, they’re telling the U.S. that it has years, not decades, to figure out its own new path forward.
For decades, this process sufficed for energy generators (and regulators) grown accustomed to gradual, predictable load growth. But over the past several years, the scale and speed of increasing energy demand has overwhelmed the supply -side’s ability to respond. The resulting strain on the grid has reverberated through every rung of the supply chain, delaying development timelines, increasing costs, and elevating energy from political conversations to dinner table discussions.
The loudest creaks and groans are coming from PJM Interconnection, North America’s largest grid operator. Residential bills in the PJM service area are climbing at a dizzying pace. Recent capacity auctions have ended with record prices, which PJM’s own market monitor blames on the explosive growth in data center power demand. Pennsylvania Governor Josh Shapiro has attempted to pressure PJM to lower its capacity price cap. Even Secretary of Energy Chris Wright has called on the Federal Energy Regulatory Commission to develop new procedures to help get data centers online faster.
David Mills, PJM’s CEO, published a 70-page report in May acknowledging that current market rules cannot keep pace with AI-driven load growth. And yet he also refused to recommend a path forward, leaving the decision to “state regulators and legislatures, to FERC, to consumers.”
The most essential grid infrastructure, he explained, “is not a price curve or a performance obligation — it is legitimacy.” In other words, what’s broken isn’t a parameter inside the capacity market, but rather the capacity market itself, along with the political conditions under which it operates. PJM calls this the “credibility trap”: high prices accurately signal that new investment is needed, but when those prices become politically untenable, government intervenes and investment stalls.
The fix, Mills writes, “requires structural choices, not just parameter adjustments.”
Mills is speaking to a deeper issue with the grid than its ability to respond to shifting market dynamics, which is that hyperscalers and grid operators are built to solve two different kinds of problems. Hyperscalers solve engineering problems with specifiable objectives, known constraints, verifiable outcomes. Engineering problems reward concentrated authority and unilateral decision-making.
Grid operators, on the other hand, solve coordination problems. The information they rely on to do so is dispersed across millions of stakeholders, continuously revised and often contradictory, and operators’ preferences are not so much known as they are revealed through deliberation. FERC’s standard for wholesale rates is not whether those rates are objectively “correct,” but rather whether the market settled on those rates through fair competition. The process does not just determine the answer, it essentially is the answer.
This construction is the category error driving the current AI-grid collision. The electricity grid is not an engineering problem with coordination problems attached. It is a coordination problem with engineering problems embedded in it. Treat it as the former and you lose all the information that gets generated in the process of market-based price discovery. You also lose all the buy-in that occurs when real people are faced with real trade-offs and have to make hard, binding choices.
Mills did lay out three possible structural paths in his May letter:
- Path A uses long-term contracts to limit price shocks but also locks customers into yesterday’s prices when the market falls. The price signal remains but is muted, and decision-making shifts to the utilities and state commissions.
- Path B formalizes what already happens on a de facto basis when the system is short and load must be shed: some customers’ service gets cut before others’. Currently, the order in which customers lose service is a byproduct of the physics and architecture of the grid. Going forward, curtailment priority could be an attribute that is competitively priced. This path is the most politically volatile.
- Path C takes the opposite approach — increase the real-time scarcity price signal and reduce the administrative overhead of the capacity market. While total electricity costs would not increase, price spikes (and general consternation) would. But the burden for covering those spikes would shift to load-serving entities, shielding ratepayers.
These pathways are not equivalent — unlike with an engineering problem, there are no cut-and-dried solutions here. There are only trade-offs and questions about who bears their consequences. Path C is likely the better answer, while Path A is more expedient. The gap between them is the work PJM’s constituents have to manage over the coming years. PJM may choose the wrong path, or arrive at the right one too late.
The alternative is not hypothetical. If hyperscalers aren’t willing to wait for PJM customers to decide which path they want to take (and recent history suggests they are not) they will build behind-the-meter generation, sign bespoke deals with regulated utilities, and restart dormant nuclear plants. America would be left with two grids, one for compute, one for everything else. The first will be reliable and expensive. The second will be cheaper, fragile, and stranded with the costs of the system the first walked away from. The market would lose the dispatch signal, the error-correcting price mechanism, and the legitimacy of the system that has reliably powered the Mid-Atlantic for two decades.
Economist Friedrich Hayek described the limits of humans’ planning capabilities better than anyone in his 1974 Nobel Prize lecture, using the metaphor of the craftsman shaping his handiwork versus the gardener cultivating growth. The craftsman thinks they can make a perfect tool but repeatedly runs up against the boundaries of their own knowledge, whereas the gardener learns to manage new information as it arises, tending not to the product itself but rather to the conditions that produce it.
Hyperscalers are not bad actors. They have legitimate interests and the political capital to help shape the grid’s future. But we should resist the Newtonian urge to meet unexpected, swiftly moving demand with equally swift supply. Markets and physical systems both tend toward equilibrium, but the former finds it through deliberation, not collision. Instead of trying to unilaterally craft a better grid, hyperscalers might find a better path if they work with the practitioners who already know how to garden.
Beam me up to baseload. Virtus Solis Technology
