The Customer Becomes the Grid

The line already moved

For most of the AI buildout, hyperscalers bought power. They signed PPAs, paid utilities, and let someone else own the iron. That changed this year.

Alphabet closed its $4.75 billion acquisition of Intersect in March, the first time a hyperscaler bought the developer instead of the electrons. Intersect expects roughly 10.8 GW online or in development by 2028. In June, Microsoft signed a 20-year agreement with Chevron for Project Kilby, a 2.67 GW gas plant planned beside a Microsoft data center in West Texas, with a final investment decision expected by year-end. Regulators are clearing the path too. FERC’s December 2025 order directed PJM to write clear rules for data centers co-located with generation, and the commission has kept pushing on large-load interconnection since.

The easy read is “hyperscalers are becoming utilities.” That’s directionally right and analytically early. The more useful question is where, on what timeline, and under whose contract.

Read the number before you read the trend

Start with scale. The U.S. runs on roughly 1,300 GW of generating capacity. Kilby, one of the largest co-located projects in the country, is about 0.2% of that. Add every hyperscaler-owned or -controlled gigawatt announced so far and you’re still in low single digits as a share of the national fleet.

Then look at what the plant is designed to do. Chevron says Kilby will run independently of the regional transmission system, a choice made to keep local ratepayers from carrying the cost. That’s self-supply. It adds generation to the country without adding a megawatt the grid can dispatch.

So nationally, through 2030, hyperscalers aren’t the grid’s primary source and aren’t trying to be. Regionally is a different story. In a West Texas county or a single PJM zone, one company can become the largest generator and the largest load at the same time. At that point it sets the local price, shapes transmission planning, and sits across the table from the utility as a peer. “Primary” arrives region by region long before it shows up in a national chart.

Forced competence becomes a product

Here’s the part I think gets underweighted. Hyperscalers aren’t building power because they want to. They’re building it because the grid can’t deliver on their timeline. Necessity is teaching them development, interconnection, turbine procurement, fuel contracting and plant operations, one gigawatt-scale project at a time.

We’ve seen this movie. Amazon built world-class infrastructure to run its own retail business, got very good at it, and realized the capability was worth more sold than kept. That was AWS. Power has the same shape: a cost center, run out of necessity, that slowly becomes a competence the board can see.

Land is cheap. The scarce things are an interconnection position, a substation, permitted transmission, and a fuel supply, and hyperscalers are accumulating all four at every campus. Adding generation behind an interconnection you already hold costs far less than starting from a bare field. That’s where the supplemental-profit math starts to work.

The board conversation writes itself once you notice who earns the return today. Chevron has said Kilby targets mid-teen returns. That margin currently flows to the partner. A board that watches its own team learn to site, permit and run these plants will eventually ask why.

There’s an honest counterweight. Generation returns sit well below cloud margins, so a pure return-on-capital case is hard to make. The case that wins is probably strategic: a hedge on the company’s own power cost, and a permitting story that changes the room. A company that can tell a state “we bring more generation than we consume” has a very different consent conversation than one blamed for rate increases. On the consent rung, that may be worth more than the margin.

The surplus has a buyer

The bear case for this thesis is that AI demand cools and the plants sit idle. It’s worth taking seriously, and it doesn’t break the argument. It accelerates it.

Electricity demand outside data centers isn’t going away. Electrification, reshored manufacturing and vehicle charging keep pulling load onto grids that were already short of firm capacity. A hyperscaler holding gigawatts of long-lived generation in a demand slowdown doesn’t write it off. It sells into the market, and becomes an independent power producer by default rather than by strategy.

Either path ends with a hyperscaler selling power to the grid. The only question is whether the contracts were written for it.

Where “primary” becomes literal

Outside the U.S., the timeline compresses. In countries where national generation is measured in hundreds of megawatts or a few gigawatts, a single AI campus can exceed existing supply. Whoever builds generation for that campus becomes the anchor of the national grid, and the host government will want the surplus. In those markets the data-center operator could become the primary generation source within this decade, well ahead of anything in the U.S.

There’s a precedent for exactly this, and it’s instructive. Ghana’s Akosombo Dam was built in the 1960s on the strength of an aluminum smelter. Valco, owned by Kaiser and Reynolds, guaranteed to buy half the dam’s output for thirty years, and that guarantee is what made the project financeable. The smelter was the data center of its day: a huge, constant load that justified capacity well beyond itself. The dam went on to power the country.

Then the terms aged. The favorable power price that launched the project became a political problem. Ghana’s government renegotiated the rates in the 1980s, and by the early 2000s the two sides were back at the table, with the government arguing Valco was still paying about 1.1 cents per kWh while the cost of producing power had moved well past that. When drought dropped the reservoir, the smelter’s supply was cut and its pot lines went dark.

Akosombo worked. That’s the point. The contract that justifies the plant isn’t the contract that survives it. The anchor customer always gets renegotiated, and the only choice is whether that happens on a schedule both sides wrote in advance or in a crisis, when the grid is short and the politics have turned.

Structure it now, region by region

That’s why I think the parties to these deals should be designing their agreements today for the scenario where the customer becomes the grid. You don’t remove a bottleneck, you relocate it. The power constraint is moving off the utility’s balance sheet and onto the hyperscaler’s, and from there it relocates into the contract. A handful of terms decide who carries it:

• Export and surplus rights. Can the plant sell power to the grid, at what volume, and at what price? A formula indexed to cost or market holds up better than a fixed rate that looks generous in year one and indefensible in year fifteen.
• Shortage priority. When the grid is short, who gets curtailed first, the data center or the region? Define the ladder before the heat wave, not during it.
• Grid services and cost allocation. Standby service, backup and transmission upgrades cost money. Who pays should be explicit, so “we insulated the ratepayer” stays true once the plant starts exporting.
• Scheduled reopeners. Build renegotiation triggers into the deal, such as a price review every five years or a review when the host region’s demand passes a set threshold. Valco’s rates got reopened anyway, just under pressure.
• End of life and handback. If the data center leaves or shrinks, who owns the plant, who runs it, and on what terms? Build-operate-transfer structures exist for exactly this question.
• Local benefit. A share of export revenue or a firm allocation to the host community turns the plant from “their power” into “our power,” which matters most on the day there isn’t enough to go around.

The right mix differs by region. In a market like ERCOT, the issue is market access and price exposure. In a vertically integrated state, it’s the regulator and the utility franchise. In an emerging market with a sovereign utility, it’s offtaker credit, currency, and the political durability of the deal. One template won’t travel.

The question underneath

Today these plants are built to serve one customer. Over a thirty-year asset life, that customer’s needs will change, the region’s needs will grow, and the politics will move. The companies that structure for becoming a grid supplier will earn a second business. The ones that structure only for self-supply will likely find themselves renegotiating someday, on someone else’s timeline.

If your data center’s power plant outlives your data center, who was it really built for?