AI Needs More Electricity: Can Batteries and New Energy Technology Keep Up?

01 / The story

Demand: what has happened versus what is forecast

The International Energy Agency estimates global data centers used about 415 TWh in 2024 and roughly 485 TWh in 2025, with a central projection of about 950 TWh in 2030—around 3% of global electricity demand. AI-focused consumption grows faster than the total, but these figures cover all data centers, not AI alone. [1, 2]

For the United States, a DOE-supported Lawrence Berkeley National Laboratory report put data centers at about 176 TWh in 2023—4.4% of U.S. electricity—with a 2028 scenario range of 325–580 TWh, roughly 6.7–12%. EPRI's February 2026 scenarios go further: 9–17% of U.S. electricity by 2030, with AI workloads estimated at 15–25% of data-center electricity today. That AI share is an estimate with uncertain attribution, not a meter reading. [3, 4, 5]

One unit warning that prevents most confusion: gigawatts (GW) measure power or capacity; terawatt-hours (TWh) measure energy over time. A 1 GW load running continuously uses 8.76 TWh per year. Saying a company 'added 24 GW of batteries' tells you nothing about how long those batteries can run anything. [2]

02 / The story

The grid is a delivery problem, not just a supply problem

Berkeley Lab's Queued Up: 2026 Edition reports more than 2,060 GW of generation and storage projects seeking transmission-grid connections at the end of 2025. Most applications never become operating projects—and this is a generation queue, not a count of data centers waiting for service. [6]

A signed energy contract does not provide a substation, a transformer, a transmission path or permission to connect. Interconnection studies identify network impacts, upgrades and who pays for them. Local deliverability can constrain a project even when national electricity production is sufficient. [2, 6]

The IEA points to workarounds: onsite generation, storage, demand response and non-firm grid connections that bring faster access in exchange for occasional curtailment. Flexible access is not the same as guaranteed round-the-clock supply. [2]

03 / The story

Batteries: scaling fast, with real limits

The deployment numbers are genuinely impressive. The U.S. Energy Information Administration reported a record 15 GW of utility-scale battery additions in 2025, with 24 GW planned for 2026—developer intent reported in February 2026, not verified completed capacity. IRENA's cost series shows global battery-storage installed costs falling from $2,571 per kWh in 2010 to $192 per kWh in 2024, a decline of about 93%. [7, 8]

But batteries move electricity between times; they do not generate it. Charging requires electricity, and every cycle loses some of it. Duration matters: a 100 MW / 400 MWh battery can supply a constant 100 MW load for about four hours under idealized conditions—not indefinitely. Milliseconds of power-quality smoothing, several hours of load shifting and multi-day renewable shortfalls are different jobs requiring different sizing. [2]

The IEA's April 2026 report suggests 20–25 GW of battery storage could be installed in data centers globally by 2030. That is a forecast, not current deployment, and a different scope from the EIA's U.S. grid-wide figures. [2]

04 / The story

Renewables add energy; they do not deliver it around the clock

IRENA's July 2026 report found that more than 90% of utility-scale renewable projects commissioned in 2025 produced electricity more cheaply than the cheapest new fossil-fuel plant in their market, with global solar at about $44 per MWh and onshore wind around $33 per MWh. [9]

Those are generation costs—not the cost of firm, 24/7 delivered power including transmission, storage, backup and balancing. Annual renewable purchases that match annual consumption are not proof that every hour runs on carbon-free electricity; Google's own nuclear announcement explicitly frames its goal as carbon-free energy every hour, which is a different and harder target. [9, 11]

05 / The story

The nuclear deals: restarts versus first-of-a-kind reactors

Tech companies have signed striking nuclear agreements, but the details matter. Microsoft's September 2024 deal with Constellation is a 20-year power purchase agreement supporting the restart of Three Mile Island Unit 1—renamed the Crane Clean Energy Center, about 835 MW, originally targeting 2028. This is a restart of an existing reactor, not a small modular reactor, and not the accident-damaged Unit 2. [10]

Google's October 2024 agreement with Kairos Power covers up to 500 MW from multiple advanced reactors, with first deployment targeted for 2030. Amazon's October 2024 agreements with Energy Northwest, X-energy and Dominion support potential small modular reactor development. Target capacities and dates are development plans, not electricity already supplied. [11, 12]

More recently, on September 9, 2026, Google announced Finnish infrastructure and energy partnerships, including its first nuclear plant life-extension agreement and heat reuse at Hamina—company-announced plans, not independently audited outcomes. Reuters' April 2026 reporting documents continuing hyperscaler financing interest in next-generation nuclear, not a guarantee that proposed fleets arrive on schedule. [13, 15]

06 / The story

So can technology keep up?

There is no verified single yes or no. Short-duration storage is scaling quickly. Renewables are the cheapest new generation in most markets. Grid connections and firm supply remain the binding constraints in many places, and advanced nuclear is a longer-horizon proposition in the cited deals. [6, 7, 9, 10, 11, 12]

Efficiency is the wild card. The IEA reports rapidly improving energy per AI task—but also increasingly demanding use cases. Better efficiency does not guarantee falling total demand, and cooling, at 7% to over 30% of facility electricity depending on the facility, is only one component of the total. [2, 14]

The honest summary: AI's electricity problem is solvable in pieces and unsolvable as a slogan. Watch completed projects versus announced ones, interconnection queue progress, and hourly carbon-free matching claims. Those numbers will tell you more than any single deal headline.