Solar power disappears every night, but a data centre cannot. That engineering mismatch is turning Google, Amazon and other technology giants into some of the world's most important buyers of battery storage, financing gigawatt-scale systems to keep increasingly power-hungry AI infrastructure supplied around the clock.

A few months ago, I wrote about Meta buying future electricity from solar collectors in orbit. It was one of those stories that made the AI power problem feel almost absurd. Computing demand had grown so quickly that putting part of the power station 36,000 kilometres above Earth could pass for a reasonable meeting agenda.

The more immediate response is happening on much less exotic ground. Google, Amazon and other large technology companies are signing agreements that bring hundreds of megawatts of batteries onto regional grids, usually beside the solar and wind projects intended to support new data centres.

I do not think the useful story is that Big Tech has suddenly discovered environmental virtue. These companies are creating a large part of the new electricity demand. They are also becoming unusually influential customers for the infrastructure needed to meet it. Both things can be true without cancelling each other out.

The engineering problem underneath all of this is almost embarrassingly simple. Solar follows the Sun. A data centre follows the internet, and the internet has never respected sunset.

The mismatch happens hour by hour

For years, technology companies could say they had matched their annual electricity use with renewable generation. That matters because long-term purchasing contracts help wind and solar farms get financed. But annual accounting can hide the clock.

A company might buy as much renewable electricity as it consumes over twelve months and still draw from a fossil-heavy grid on a still night. The solar farm overproduces at midday, the data centre keeps running after dark, and a spreadsheet later balances the totals. The annual claim can be legitimate while the physical electricity supply remains very different from hour to hour.

The latest International Energy Agency outlook estimates that global data-centre electricity consumption rose to about 485 terawatt-hours in 2025 and could reach roughly 950 terawatt-hours in 2030. AI-focused facilities grow faster still. The agency says 20 to 25 gigawatts of batteries could be installed at data centres worldwide by the end of the decade, partly because AI workloads can create large, rapid swings in demand as well as a large continuous load.

That forecast is not a promise. Efficiency could improve faster than expected and projects could be delayed. I am cautious with any projection drawn as a smooth rising line. Even so, the direction is difficult to miss.

Google’s projects show how quickly the scale has changed

In Arkansas, Google is acting as an anchor investor and buyer for the Steel River Energy Center. The first two phases are planned to add 1.6 gigawatts of solar generation and 1.9 gigawatt-hours of battery storage. If all three phases are completed in 2029, the project would reach 2.5 gigawatts of solar and 2.9 gigawatt-hours of storage.

Those numbers deserve a pause. This is not a little backup cabinet attached to a server room. It is power-station infrastructure. Yet it is also important not to merge the figures into one impressive but meaningless total. The 2.5 gigawatts describe the peak direct-current rating of the solar panels. The 2.9 gigawatt-hours describe how much energy the batteries can hold. Power and stored energy are related, but they are not interchangeable.

Google’s Minnesota agreement goes in a different direction. Xcel Energy says the package for a new Pine Island data centre includes 1.4 gigawatts of wind, 200 megawatts of solar and a 300-megawatt, 30-gigawatt-hour iron-air battery. The battery is designed to discharge for about 100 hours, which moves it out of the familiar afternoon-to-evening role and into multi-day weather coverage.

These are grid resources funded through agreements associated with data-centre growth, not a private Google power island. They can support other customers too. A very large new electricity user is helping make very large new assets financeable.

Amazon is assembling the same pieces in a different pattern

Amazon says it had paired 15 solar projects with battery storage by the end of 2025, representing 2.3 gigawatts of project capacity. The wording matters here too: that is the capacity of the paired projects, not necessarily 2.3 gigawatts of battery discharge alone.

In Nevada, Amazon is working with NV Energy on a package that includes 600 megawatts of solar, 600 megawatts of battery storage and 100 megawatts of geothermal power for future data-centre operations around Reno. Amazon says it will cover the costs associated with serving its facilities and the new infrastructure. That is a company statement, not the final word on how every cost and benefit will flow through a regional electricity system, but the structure is significant.

The geothermal component is the quiet clue. Batteries move electricity through time; they do not create it. Solar can charge them cheaply during the day, but several cloudy days in a row still require another source. A mix of solar, storage and steady geothermal generation is more convincing than asking batteries to carry every night and every weather event alone.

Strip away the adjectives and the portfolios are revealing. Google and Amazon are combining variable renewables, several kinds of storage, firm generation and flexible demand. No one responsible for a real data centre is betting everything on one elegant technology.

A megawatt tells you the size of the tap, not the tank

Battery announcements are particularly easy to misunderstand because power and energy use nearly identical units.

A megawatt measures the rate at which electricity can be delivered. A megawatt-hour measures an amount of stored energy. Think of one as the width of a tap and the other as the size of the tank behind it. A 600-megawatt battery could produce full output for one hour if it stores 600 megawatt-hours, or four hours if it stores 2,400 megawatt-hours.

This changes how the Arkansas number should be read. In a deliberately simple example, 2.9 gigawatt-hours could support a constant 500-megawatt load for 5.8 hours. Real operation would be shorter after conversion losses, reserve margins and limits intended to protect battery life. The project also serves the wider grid rather than dedicating every stored unit to one building.

Five or six hours can carry solar into the evening peak and give operators time to respond to an outage. It is not the same as supplying a large data centre through a long winter storm. “Gigawatt-scale” shows seriousness; duration shows what the equipment can do.

There are already batteries inside the servers

The grid-scale projects are only one layer. Google reported in 2025 that the lithium-ion backup units built into its racks had reached 100 million individual cells.

Those cells bridge short disturbances, hold servers steady while power sources switch and allow a clean shutdown if an outage continues. Their timescale is seconds or minutes, not an entire night.

Facility-level uninterruptible power systems sit above them. Grid batteries then operate across longer windows, charging when electricity is plentiful and discharging when the network is strained. Long-duration technologies such as iron-air or Google’s planned 23-megawatt, 200-megawatt-hour carbon-dioxide battery in Ireland try to stretch storage further.

The data centre can sometimes move instead of the electricity

Storage is not the only way to reconcile a fixed supply with a flexible workload. Google says it has now integrated one gigawatt of demand-response capacity into US utility agreements. Some machine-learning jobs can be delayed, reduced or shifted to another time when the grid is under pressure.

Not every workload can wait. A search request or online payment needs an answer now. Training runs and background jobs may have more room. Moving computation by an hour can be easier than storing enough electricity to pretend the hour never happened.

This is a point I keep returning to. When I wrote about Denmark producing 60 percent of its electricity from wind, the impressive percentage was only half the story. The supporting system mattered just as much: interconnectors, flexible generation, imports and prices that encourage demand to move away from scarce hours.

Data centres need the same humility. Batteries help. So do transmission, nuclear power, geothermal energy, gas in some current projects, and software that knows when a job can wait. A resilient system is usually less pure and more practical than the diagram in a company presentation.

Why technology companies have become unusually important buyers

A battery developer needs more than cells and land. It needs confidence that someone will pay for the service over many years. A long-term contract with Google or Amazon can provide predictable revenue, lower financing risk and turn a proposed project into one that lenders will support.

A technology company can therefore reshape the electricity system without owning every battery. It can be an anchor customer, sign an offtake agreement or accept a utility tariff that pays for new infrastructure.

The less comfortable side is that these buyers are influential because their demand is enormous. Local residents are entitled to ask who pays for substations and transmission, who receives the reliability benefit and what happens if promised jobs disappoint.

Promises that existing customers will not bear the cost should be tested by regulators, not merely repeated. Equally, it would be odd to complain about data-centre power demand and dismiss the storage it helps finance. Both belong in the assessment.

Batteries are becoming essential without becoming sufficient

I spend a lot of time writing about distant machines surviving on almost no power. Voyager 1 loses roughly four watts of available power each year, and engineers debate which instrument can stay warm. AI infrastructure forces the same word, power, into a scale so different that it almost stops meaning the same thing.

That contrast is part of why I am wary of tidy verdicts. The booster version says batteries and renewables will make AI clean. The cynical version says every storage announcement is public relations wrapped around unavoidable demand growth. I am not convinced by either.

What I see is an industry colliding with the physical world. Chips need substations. Training runs need transmission lines. Solar generation needs somewhere to go at noon and something to return after sunset. Large technology companies are becoming important battery buyers because the reliability of their core business now depends on solving those ordinary engineering problems.

The contracts can accelerate storage that benefits the wider grid. They can also accompany a build-out that raises total electricity demand, keeps some fossil generation operating and creates legitimate local concerns. Progress here will be measured less by one headline capacity number than by hourly carbon emissions, reliability, who bears the cost and whether promised projects actually get built.

Solar power disappears every night. Batteries can move part of the day across that boundary. They cannot abolish weather, create energy from nothing or replace the rest of an electricity system. That is not a failure. It is a realistic job description, and it is already large enough to change who builds the world’s power infrastructure.