The power bill is the pro forma

$28.92. That is what PJM Interconnection's capacity auction cleared at, per megawatt-day, for the 2024-2025 delivery year. Three auctions later, the 2027-2028 auction cleared at $333.44, the temporary price cap, and 809.6 megawatts of capacity did not clear at all because suppliers wanted more than that and PJM said no.

Eleven and a half times, in three delivery years. Nobody's zoning fight moved that fast.

What is a capacity auction, and why does a developer care?

PJM runs the mid-Atlantic and Midwest grid covering roughly 65 million people, and its capacity auction is how the grid pays generators to be available when demand peaks, years in advance. It is boring market plumbing until data centers show up in the load forecast, and then it becomes the fastest-moving number in the entire site-selection stack. The 2026-2027 auction, held July 9-15, 2025, cleared at $329.17, a 22% jump from the prior year's $269.92, and PJM's own forecast attributed the underlying load growth (about 5,500 MW) mainly to data centers. Total auction cost: $16.1 billion, up 9.5% year over year. Without a Pennsylvania-brokered price cap, PJM says the clearing price would have hit roughly $389.

Run the full sequence and the shape is unmistakable.

The 2027-2028 auction, held in December 2025, put the forecast peak load about 5,250 MW above the prior year's forecast, and PJM attributed 5,100 MW of that, nearly all of it, to data centers. The total auction cost came to $16.4 billion. Whatever a site memo says about "power availability" as a yes-or-no gate, this table is the honest version: power is a price, and the price is data centers bidding against everyone else who also needs electricity in 2028.

Is this a PJM problem, or is it everywhere?

Everywhere, just at different speeds. The EIA reported industrial electricity revenue per kilowatt-hour up 5.5% year over year through April 2026, and the all-sector average hit 13.88 cents/kWh, up 6.0%. In January, the EIA forecast the strongest four-year growth in US electricity demand since 2000, driven largely by data centers, with demand up 1% in 2026 and accelerating to 3% in 2027. A grid that grew load at roughly flat for two decades is now absorbing years of unplanned growth inside a single budget cycle, and industrial ratepayers, the steel mill next door to your data center site, are in the same rate case you are.

Corporate power-purchase agreements moved even faster than the regulated numbers. LevelTen's Q1 2026 index put average North American wind PPA prices at $79.40 per megawatt-hour, up 24% year over year, and solar at $64.49, up 13%, both record highs. LevelTen's own language for the buyer pool: demand from data-center operators and hyperscalers "continues at a fever pitch," while everyone else pulls back. Some hyperscalers have stopped waiting for the market and started buying it. Google has acquired renewable-energy developers outright rather than compete for contracted output, a decision that only makes sense if you assume the PPA market stays this tight for years.

That is not a hedge against rising power costs. That is an admission the power line is now the biggest variable in the model.

What happens before you ever get a rate?

You wait. Roughly 2,300 gigawatts of generation and storage capacity sit in US interconnection queues today, more capacity than the country has installed. Queue times that ran two to three years fifteen years ago now run four to seven years in the markets absorbing most of 2026's announced load. Texas shows the scale best: ERCOT's large-load interconnection queue stood at roughly 410 gigawatts in April 2026, a state whose entire grid peaks around 85 gigawatts. The queue is not a formality anymore. It is a rationing mechanism, and it prices in years before a single kilowatt-hour changes hands.

States are writing new rules because the old ones assumed load this concentrated would never show up. Twenty-four states had approved at least one large-load tariff as of June 2026, with four more pending, and every one of those tariffs exists to answer a single question: does the data center pay for the grid it requires, or does everyone else's electric bill absorb the difference. Michigan answered that question inside a single contract. The Saline Township Stargate campus's power approval carries an 80% minimum billing demand across a 19-year term, meaning the site pays for four-fifths of its contracted load whether it draws the power or not. That single clause is not boilerplate. It is the power bill functioning as a floor under the entire pro forma, negotiated line by line, for two decades.

An 80% minimum billing demand, signed for 19 years, is not a utility footnote. It is the pro forma, wearing a different cover page.

Does the market you pick change the math?

It changes which number moves, not whether one does. PJM prices capacity years ahead because it is a capacity market; ERCOT in Texas runs energy-only, no capacity auction at all, which is why its stress shows up as a 410-gigawatt interconnection queue instead of a headline auction price. MISO, the market covering Michigan and much of the Midwest, handles large new load through utility-specific special contracts reviewed case by case, which is exactly the forum that produced Saline's 80% minimum-billing floor. Three different market designs, three different pressure points, and the same underlying fact in each one: someone has to decide who pays for the grid a gigawatt-scale campus requires, and in every design the answer increasingly comes back to the developer. A site memo that treats "which ISO" as a footnote is skipping the paragraph that tells you which cost shows up first, the auction bill, the queue delay, or the contract floor.

The conditions attached to these approvals are getting more specific, not less. Michigan's Public Service Commission approved the Saline contracts 3-0 with a 19-year minimum duration, customer-funded energy storage, quarterly reporting requirements, and an order that the utility file a generally applicable large-load tariff within 90 days. Regulators are no longer approving a number. They are approving a structure, and that structure outlives whatever the campus's own economics do over two decades.

So why do site memos still treat power as a checkbox?

Because the checkbox era was recent enough to still feel normal. Through most of the last decade, a data-center site either had power or it didn't, and the number that mattered was megawatts available, not dollars per megawatt-day, PPA escalators, or minimum-billing floors. That framework worked when capacity auctions cleared under $50 and interconnection queues ran two years. It stopped working somewhere around the 2025-2026 PJM auction, and most underwriting templates have not been rebuilt since.

The fix is not complicated, just uncomfortable, because it means treating the power contract with the same line-by-line scrutiny a lease gets. Model the capacity-price trajectory in the ISO you are siting in, not a flat assumption carried over from the last deal. Ask whether your minimum billing demand looks like Saline's, and if it does, run your model at 80% utilization even in a downside scenario, because the utility already did. Price the interconnection queue as a schedule risk with its own probability distribution, the way we've argued land-use risk deserves, because a 2026 queue position is worth real money and a 2030 one might not clear before your tenant walks.

Here is my call, and it is checkable. Within three years, published all-in power cost, capacity charges plus energy plus the minimum-billing floor, per contracted megawatt, becomes a standard disclosure line in hyperscale earnings materials and financing decks, the same way price per acre used to be standard in a land deal. I expect at least one major PJM-territory hyperscale project to publicly renegotiate or shrink its power contract before 2028 over minimum-billing exposure, the way Michigan's Saline terms already hint is coming. If capacity prices flatten from here and nobody renegotiates a floor, I was wrong about how fast this repriced, and I will say so.

This analysis is a source-cited research summary drawn from public records, not legal advice. It can contain errors and should be verified independently before any investment decision.