3,590 MW Against a $325 Price Cap (Google–Constellation, 6 October 2026): Why AI Is Now Buying Its Own Power — and What PJM's 'Bring Your Own' Rule Changes

Google has contracted 3,590 megawatts of power from Constellation Energy in PJM, the largest US grid, combining a 20-year deal for 890 MW of new nuclear output from upgraded reactors with a 15-year deal for 2,700 MW more. The deal arrives as PJM's capacity price sits at its $325-per-megawatt-day cap and the grid operator is asking regulators to let it curtail new data centres that don't bring their own supply. The headline is a power contract. What changed is who pays for the power: from now on, more of the cost of AI's electricity sits with the companies building it.

What was actually signed

The structure matters more than the total, so take it apart. According to Bloomberg's report, the deal has two legs:

Constellation says the agreement supports more than $4.3 billion of investment in its fleet. Its shares rose more than 13% to $304.04 by mid-morning on 6 October, on a day when the Nasdaq 100 closed at a record 31,076.

The quotes on the record are measured. Amanda Peterson Corio, Google's global head of energy and power, said the company is "committed to meeting our growth responsibly." Constellation chief executive Joe Dominguez called it "a model for how technology companies and the energy industry can work together." Neither statement is a forecast, and neither should be read as one.

For context: this is the same company that restarted the Three Mile Island unit for Microsoft under a 20-year agreement announced in 2024. What is different now is scale, and the reason the timing lines up with a rule change at the grid operator.

Why "uprate" is the most important word in the deal

An uprate raises the licensed output of a reactor that already exists. The site, the operator, the grid connection and most of the safety case are already in place; the gain comes from equipment and efficiency upgrades. That is why the new nuclear piece is spread across 11 units: it is many increments, not one plant.

This is how the contract solves the buyer's real problem, which is time. New large reactors in the US have been measured in decades and cost overruns. Uprates are measured in years. A hyperscaler that needs firm, around-the-clock power before the end of the decade will value 890 MW starting in 2028 more highly than a larger plant on an uncertain timeline. And because the output is new — megawatts that wouldn't exist without the investment — it counts as adding supply to the grid rather than taking supply other customers already rely on. That distinction is the hinge for the PJM rule below.

The grid underneath: an auction that hit its cap and still came up short

PJM runs the grid across 13 states and Washington DC. It buys "capacity" — a commitment that generation will be available at peak — through an annual auction. The price of that commitment is passed through to every customer's bill.

The results PJM published on 14 July show the strain. The 2028/2029 auction procured 138,318 MW at a clearing price of $325/MW-day — exactly the cap agreed with the PJM states and FERC — for a total cost of $16.4 billion. Even at that price, committed supply was 6,831 MW short of the reliability requirement. The year before, 2027/2028, cleared at its cap of $333.44 and was about 6,500 MW short. For scale, the 2024/2025 auction cleared at $28.92; Bloomberg puts the rise in capacity prices at more than 11-fold since 2024.

An auction that clears at the cap and still falls short is giving a specific signal: the cap, not supply and demand, is setting the price. In PJM's own words, "price caps may reduce volatility, they do not solve the supply-demand imbalance." When prices can't do the rationing, rules have to — which is where the 6 October deal comes from.

The rule behind the timing: bring your own capacity, or be curtailed first

In August, PJM filed its Interim Resource Adequacy Service with FERC. The mechanism is simple to state:

PJM asked for the tariff changes to take effect on 12 October 2026, and the filing is awaiting FERC's decision. PJM has also proposed that, from the 2029/2030 auction, new large loads that don't bring supply would be left out of the demand PJM buys capacity for — so existing customers would no longer pay to cover them.

For a data centre, the threat of being curtailed first is not a small one. A training cluster or inference fleet that can be switched off on the hottest days of the year is a weaker asset than one that can't. The cleanest protection is the one Google just bought: long-term contracts for supply that is demonstrably new. Read this way, the deal is less a bet on nuclear power than a purchase of a place in the queue under rules that are about to change.

The policy question underneath — who should carry the cost of grid expansion driven by a single industry — is one regulators, states and companies are contesting. This note takes no side on it. The market point is narrower: whichever way FERC rules, the direction has been toward large loads paying for their own supply, and the contracts are being written to match.

Where it reaches the indices

AI capex has a new fixed line. Hyperscaler spending has been read mostly as chips and buildings. A 15- or 20-year power contract is a different kind of commitment: it does not depreciate the way a GPU does, but it is a long-dated fixed obligation whose value depends on the data centres it feeds staying busy. The same distinction ran through Nscale's pre-IPO numbers, where contracted backlog and billed revenue measured very different things. Megawatts under contract are capacity, not revenue, until the compute they power is sold.

Power producers are now part of the AI trade. Constellation's 13% move came on a day when the Nasdaq 100 closed at a record and Nvidia's market value was approaching $6 trillion, per Bloomberg. When the companies selling electricity to data centres re-rate on contract news, the AI theme stops being only a tech-sector weight in US500 and spreads into utilities and power. That widens the theme's footprint in the index — and means a change in sentiment about AI demand now moves more sectors at once.

Rates are part of the arithmetic. A 20-year power contract is a long-duration cash flow, and long-duration cash flows are valued against long-term yields. With the 10-year Treasury above 5% and the S&P 500's earnings yield roughly level with it, the discount rate on decade-long infrastructure contracts is noticeably higher than it was when the first hyperscaler nuclear deals were signed.

What has to be true for this to matter — and what would change it

What has to be true: the data centres this power is bought for must be built and kept busy. A 20-year contract is worth signing only if demand for compute over that span is close to what the buyer is planning for. The deal says a great deal about Google's planning assumptions; it says nothing certain about demand in 2040.

What is already priced: a large part of the reaction happened on the day. Constellation's re-rating captured the contract news; the Nasdaq record captured the broader AI theme. Nothing in the deal itself is still unknown — the terms are public.

What would change the picture: - FERC's ruling on the PJM proposal. Approval would make "bring your own capacity" the operating rule for new large loads in PJM and probably bring more contracts like this one. A rejection or significant changes would take away part of the reason for signing them early. - The next capacity auction. Another clear at the cap with a shortfall would confirm that the constraint is physical, not a pricing anomaly. - Uprate delivery. First output is due in 2028. Delays would push back the supply the contract was signed to create. - The demand side. Any sign that AI compute is being used less than planned would put these long contracts in a different light — the risk shifts from "can we get power" to "do we need it."

The currency angle — and why it is thin

This is a story about US power, hyperscaler spending and equity-sector weights. The channel to the eight majors is weak. A domestic electricity contract doesn't move Federal Reserve expectations, the trade balance or the risk sentiment that drives safe-haven flows. Its only indirect route to the dollar is the wider case for US investment spending — one input among many into the growth factor, and too small a piece of it to move the currency on its own. The meter scores USD and the other seven majors on five factors — interest rates, growth, positioning, risk sentiment and commodities — and for this story, the honest reading is that none of them changed in any measurable way. More on how the meter frames those inputs is on the about page.

The useful lesson is about the mechanism, not the currency: when a market can't ration by price because a cap is binding, it rations by rule — and the companies with the balance sheet to sign decade-long contracts move first.

Educational macro context only — not investment advice.