The data centers are eating the grid

In May 2027, an electric utility called NV Energy will stop selling power to roughly 49,000 people who live around Lake Tahoe. The reason is not a storm or a shortage of fuel. It is a queue. NV Energy says it needs the transmission capacity for data centers being built east of Reno, and the homes are at the back of the line. Liberty Utilities, the small company that serves Tahoe, buys about three-quarters of its electricity from NV Energy. That supply is the part being cut. "It's like we don't exist," one resident, Danielle Hughes, told Fortune. She is, in a narrow accounting sense, correct: the household load lost the bid.

This is the story the AI boom does not put in its renderings. The compute gets the headlines; the power gets the invoice. And the invoice is now large enough to reorder who gets electricity first. The useful way to read the last twelve months is not as a technology story at all. It is a grid-economics story, and like all grid-economics stories it comes down to three numbers: how much new demand is arriving, how long it takes to connect, and who pays the difference.

The demand is real, and it arrives fast

Start with the load, because everything else follows from it. A 2024 report from Lawrence Berkeley National Laboratory, funded by the Department of Energy, found that US data centers consumed about 4.4% of national electricity in 2023 and could reach somewhere between 6.7% and 12% by 2028. In absolute terms that is a jump from 176 terawatt-hours in 2023 to as much as 580 by 2028. S&P Global's 451 Research puts US data-center grid demand at 75.8 gigawatts in 2026, 108 by 2028, and 134.4 by 2030 — roughly a tripling from where it stood last year.

The number that matters about this demand is not its size but its speed. A steel mill announces itself years in advance and ramps slowly. A hyperscale campus wants 500 megawatts and it wants them in eighteen months, because the chips inside depreciate faster than the building does. Grids were not built to absorb load on that schedule. They were built to absorb it on the schedule of a coal plant.

The queue is the bottleneck, and it is four years long

Here is the constraint that defines the whole problem. The US interconnection queue — the line you join to plug new generation or load into the grid — has swelled to roughly 2,600 gigawatts of backlogged projects. The median wait from application to commercial operation is now close to five years, up from under two years in 2008. In the places data centers actually want to build — Northern Virginia, Phoenix, Dallas — the wait runs four to seven years. For some large loads the estimates stretch past a decade.

Four years is not a delay. Four years is a different business plan. If you are building AI infrastructure on a depreciation clock, a four-year wait for grid power is the same as no power at all. So the industry has done the rational thing: it has stopped waiting in the line.

Four years is not a delay. Four years is a different business plan.

Batteries are the way around the line

This is where the storage boom comes from, and it is worth being precise about why. The batteries going into data centers are not there to be green. They are there to be fast. A battery system, paired with on-site gas or an existing grid connection, lets a campus firm its own power, ride through the millisecond-scale load swings that AI training produces, and — critically — energize before a full grid interconnection clears. Behind-the-meter, you skip the line.

The scale is the part that should make you sit up. According to GGII, the research house, global shipments of lithium-ion batteries specifically for AI data centers were about 12 gigawatt-hours in 2025, the validation year. They project 61 gigawatt-hours by 2027 — a 114% jump — and 272 by 2030, pushing total demand past 300 gigawatt-hours. UBS Securities pegs the AI data-center power-supply market at $24 billion by 2028. For context on how fast storage now moves: China installed 18 gigawatts, about 65 gigawatt-hours, of grid battery storage in December 2025 alone — more in one month than the United States installed in the whole of the prior year.

Why batteries and not just more gas turbines? Because turbines are also stuck in a queue — the one for gas turbines themselves, whose order books stretch years out — and because batteries respond in milliseconds, which is exactly what a training cluster's pulsing load demands. The reasons are operational, not idealistic. The cleanest framing I have seen, from data-center developers, is that batteries buy time. They bridge the gap between when you need power and when the grid can give it to you.

The cost line is what decides whether this scales

Now the number I actually care about: cost per kilowatt-hour, because that is what determines whether this is a strategy or a stopgap. Installed grid-scale battery storage runs about $73 per kilowatt-hour in China, roughly $177 in Europe, and around $219 in the United States. Outside China, an all-in four-hour utility-scale system averages near $125 — about $75 for the cells, $50 for installation and connection. Behind-the-meter data-center systems sized for short daily discharge land higher, $250 to $300, because you are paying for reliability and speed, not duration.

The spread between $73 and $219 is the whole global story in a single line. Storage is cheap enough, in China, to deploy at continental scale without anyone writing a press release about it. In the US it is expensive enough that the economics only work when the alternative — waiting four years in a queue while your chips depreciate — is worse. That is a fragile basis for a 300-gigawatt-hour boom. It holds as long as the queue stays broken. The day interconnection reform actually works, a good share of the behind-the-meter case evaporates.

Who pays

Which brings us to the part the brochures skip. New transmission, new substations, new generation — somebody funds all of it, and the default mechanism is the rate base, which means everyone's bill. The US residential rate hit 17.45 cents per kilowatt-hour in January 2026, up 9.5% year over year, well ahead of inflation. In Virginia, Dominion proposed its first base-rate increase since 1992, about $8.51 a month, largely to serve data-center load. The Federal Reserve Bank of Dallas has estimated wholesale prices could rise as much as 50% as data-center demand doubles.

There are attempts to make the load pay its own way. More than 30 states have proposed or adopted special large-load tariffs that push infrastructure and risk costs onto the big customers. Some developers are funding their own supply directly: Google is paying for 1,900 megawatts of wind, solar and storage with Xcel in Minnesota; Meta is funding gas plants, transmission and battery systems with Entergy in Louisiana. In March 2026, Microsoft, Meta, OpenAI and Amazon signed a "Ratepayer Protection Pledge" promising to cover their own power. Consumer advocates called it unenforceable. They are not wrong: a pledge is not a tariff, and a tariff is not a meter reading.

A pledge is not a tariff, and a tariff is not a meter reading.

The global lens

It is tempting to read this as an American problem, and the loudest version of it is American. But the underlying arithmetic is universal, and it lands hardest where grids are weakest. The IEA noted that global investment in data centers overtook investment in crude oil supply for the first time in 2025. That capital will look for the cheapest, fastest power it can find. In places with thin grids and chronic load-shedding — much of West Africa, parts of South and Southeast Asia — a hyperscaler that can self-supply with batteries and gas does not strain the grid so much as bypass it entirely, building a private island of reliable power next to households that have never had it. Whether that island ever connects to the people around it is a policy choice, not a technical one.

That is the quiet stake in the storage boom. Batteries are genuinely the right tool: cheap enough where it counts, fast enough to matter, abundant enough to scale. The 300-gigawatt-hour figure is not hype. But a tool is not a settlement. The same battery that lets a data center skip the queue can also let it skip the grid, and the obligations that come with the grid — serving everyone, in order, at a regulated price.

So watch the number that actually decides this, which is not gigawatt-hours shipped or cents per kilowatt-hour installed. It is the residential bill. If the data centers firm their own power and the line at the homes still gets longer and the rates still climb, then the storage boom did exactly what it was built to do for the people who paid for it, and nothing at all for the 49,000 at Lake Tahoe. The technology works. The question was never the technology.