The Grid the Hyperscalers Are Trying to Outrun

JULY ISSUE OF THE CURRENT

The data-center power story keeps getting told as an exit from the grid. The hyperscalers building their own generation in northern Nevada, West Texas, and Memphis are not walking away, though. They are trying to get online before the grid can get to them.

What a utility sells is interconnection. Firm capacity when your own generation trips, the ability to draw more as you grow without building all of it yourself, and a balancing area that smooths a load no single plant can follow. That is the product. Onsite generation energizes a campus in a hurry, but it does not give you what interconnection gives you, and at hyperscale the distance between the two is where the pressure is building.

The Hurry

The reason they build is time. Bloom Energy’s 2026 Data Center Power Report found that time-to-power now runs roughly one and a half to two years longer than the industry expects, on top of interconnection queues that already stretch five to seven years. When waiting in that queue costs multiple years of revenue at hyperscale, standing up your own power is the fast path, so that is the path they take.

In December, Google paid $4.75 billion to acquire Intersect Power, a generation developer with 2.2 gigawatts of operating solar and a much larger pipeline behind it. In effect, a software company bought its own power plants. Oracle has gone further, powering West Texas campuses with behind-the-meter gas and no utility interconnection for primary power, much of it contracted through operators like VoltaGrid rather than owned outright. And in March, the Ratepayer Protection Pledge, signed at the White House by Amazon, Google, Meta, Microsoft, OpenAI, Oracle, and xAI, made the intent plain. Build it, bring it, or buy it yourself.

Told as a permanent exit, that looks like the shared system coming apart. It is closer to something the industry has seen many times. Load that cannot wait for the queue finds a bridge, and bridges are temporary.

Interconnection Is the Product

Bloom’s survey found that roughly a third of hyperscalers expect to run entirely on onsite generation by 2030. As a headline, that is a lot of load leaving the system. As a plan, it runs straight into what the grid is for.

A hyperscale campus needs a level of availability that onsite generation alone struggles to hit. When a turbine trips or a fuel supply hiccups, an islanded site either rides it out on its own reserves or it goes dark, and going dark is the one thing these facilities are built never to do. A balancing area handles that. It matches generation to a load no single plant can follow, it absorbs growth without forcing you to overbuild years ahead of need, and it holds you up through a bad day. Reproducing that behind one fence, at a price that beats simply interconnecting, is far harder than it looks.

So the off-grid campus tends to want a grid tie anyway, for backup if nothing else. And the campus running on its own power today, betting it can scale that way, usually discovers that the cheapest way to grow is to stop building generation and start drawing it. Pretending these loads will never interconnect is a pipe dream. Most of them will, on their own timeline, and when they do they get in the queue like everyone else.

The grid’s product is interconnection. Self-generation is what a customer builds while it waits for the thing it actually wants.

Deferred, Not Gone

This is where it lands on planners. A load that is dark today but means to interconnect tomorrow never enters the queue, so it never enters the forecast. It is not gone, only deferred, and deferred load is invisible in exactly the way that causes trouble. You size upgrades and sequence outages against the demand you can see, and then a gigawatt that had been quietly self-supplying files for interconnection and expects to be served on a hyperscale schedule.

The problem gets sharper when a self-supplying campus leans on the grid as standby without a firm interconnection behind it. It draws reliability from a system it is not helping to keep ready, and the cost of keeping that system ready still lands on someone.

We covered in June how much trouble the large-load model is already in, with recently interconnected data centers whose operating characteristics still aren’t fully understood. A customer that comes and goes on its own schedule adds a layer the forecast was never built to carry, because the forecast assumes customers interconnect and stay.

No Clean Option

None of this is a failure of the utilities. It is the corner they have been backed into. A utility carries an obligation to serve, and it has to keep the lights on for everyone while a few enormous customers arrive faster than the system can absorb them.

TVA shows the bind plainly. In February, its board approved an additional 150 megawatts of firm power for xAI’s Memphis facility and, at the same meeting, voted to keep the Kingston and Cumberland coal plants running indefinitely, both of which had been scheduled to retire. TVA’s own planning documents had described Kingston as in high cost and challenged condition, and a Cumberland unit had failed during Winter Storm Fern. The board kept them running because the load required it and because someone has to answer for reliability when the demand shows up. It reads like capitulation only if you ignore the position TVA is in. The board was serving all sides at once, weighing a new customer’s timeline against retirement plans, ratepayer cost, air quality, and the reliability record of its own equipment, with no clean option on the table.

And the bill is shared. The fixed cost of the transmission lines, substations, and interconnection facilities that keep the whole thing reliable gets recovered through rates. When large customers self-supply, or come and go, that cost shifts onto whoever is steadily connected. Utilities requested more than $29 billion in rate increases in the first half of 2025, double the same period a year earlier, affecting an estimated 40 million customers. Some of that pays for infrastructure to serve genuine new load. Some of it reflects a system sized for demand that increasingly wants to arrive on its own terms.

No One Owns the Outcome

The framework underneath all of this assumes load interconnects, stays, gets planned for, and falls under a regulator who owns the result. Lake Tahoe is what happens when those assumptions slip. Last month, NV Energy told Liberty Utilities it would stop supplying wholesale power to the 49,000 residents of the Lake Tahoe region in May 2027. Liberty’s own regulatory filings pointed to data-center demand in northern Nevada as one of the reasons. To be fair, NV Energy disputes this, saying the arrangement was always meant to be temporary. And it says service will continue until Liberty secures its own transmission, and that data centers didn’t drive the decision.

The structural problem holds either way. Liberty is a California investor-owned utility, but its grid sits inside NV Energy’s Nevada balancing authority, interconnects with NV Energy at 38 points, and runs entirely on Nevada transmission lines. Building a direct interconnection to California’s grid would cost hundreds of millions and require new transmission over the Sierra. The CPUC sets Liberty’s rates but can’t compel NV Energy. FERC governs the wholesale market but can’t dictate Nevada’s resource planning. No single regulator owns the outcome.

“It’s like we don’t exist,” a Tahoe resident told Fortune. The line was less about data centers than about what happens to the obligation to serve when demand shifts faster than the rules meant to manage it.

Thanks for reading. Subscribe to The Current and follow Simple Thread on LinkedIn for fresh analysis, project highlights, and event updates, delivered straight from the front lines of utility software.