The rapid growth of artificial intelligence is creating a new electricity challenge for the United States. AI data centres need huge amounts of power to train and run models, pushing utilities, technology companies and policymakers to find new ways to generate and deliver electricity.
The US Energy Information Administration (EIA) expects electricity consumption to hit a record 4,270 billion kWh in 2026, up from 4,195 billion kWh in 2025. It expects demand to rise further to 4,349 billion kWh in 2027. AI-focused data centres, along with the wider electrification of homes and transport, are among the key drivers.
The pressure is already visible in Texas. Governor Greg Abbott on September 21 ordered a halt to new state-issued data-centre permits until regulators assess their impact on the power grid, water supply and other resources. More than 470 GW of data-centre projects and other large energy users were waiting to connect to the Texas grid, reported Reuters.
The US is therefore facing a new infrastructure race. It needs not just more AI data centres, but also enough electricity and grid capacity to power them.
1. Solar and batteries are expanding rapidly
Solar power and battery storage are emerging as some of the fastest ways to add new electricity capacity. Carnegie Endowment for International Peace, citing EIA figures, said developers planned to add more than 90 GW of electricity capacity in 2026, including about 43.4 GW of utility-scale solar and 24 GW of battery storage.
The EIA expects renewable sources to account for about 25% of US electricity generation in 2026, rising to 27% in 2027.
But adding solar plants alone will not solve the problem. Some of the country’s strongest solar and wind resources are far from major data-centre hubs. The US therefore needs new transmission lines, substations and grid connections to move that electricity to where it is needed.
2. Natural gas is becoming a near-term solution
Natural gas is gaining importance because gas-fired plants can provide electricity around the clock and respond relatively quickly when demand changes.
The EIA expects natural gas to remain the largest single source of US electricity generation, accounting for about 40% of generation in both 2026 and 2027.
Carnegie cited an RMI review of utility plans that identified about 94 GW of planned natural-gas capacity through 2035. Texas alone has plans for roughly 10 GW of new gas capacity, backed by more than $5 billion in state loans.
Technology companies are also exploring gas generation directly at data-centre sites. Microsoft, Amazon and other hyperscalers have sought or received permits for such projects in several states.
The trade-off is that gas can help meet AI’s immediate power needs, but greater reliance on fossil fuels could make it harder for the US to reduce emissions from its power sector.
3. Nuclear power is returning to the AI energy debate
Nuclear energy is attracting renewed interest from technology companies because nuclear plants can generate large amounts of electricity continuously.
Several hyperscalers have announced nuclear-related plans or agreements. But nuclear cannot solve the immediate power crunch because new plants take years to develop and build.
Carnegie pointed to Georgia’s Vogtle Units 3 and 4, the latest nuclear reactors completed in the US, which finished in 2024 after years of delays and at a cost of more than $30 billion. Small modular reactors could eventually provide another option, but they still face challenges involving manufacturing, supply chains and skilled workers.
That makes nuclear primarily a longer-term answer to America’s growing electricity needs.
4. Data centres are starting to generate their own power
The AI boom is also changing how large technology companies think about electricity. Instead of relying entirely on the local grid, some data-centre developers are looking at generating power at or near their facilities. This could reduce dependence on grid connections and allow some projects to become operational sooner.
Natural gas is one option because it can provide continuous electricity. BloombergNEF’s July 2026 US data-centre outlook identified 124 GW of announced on-site gas-generation capacity associated with data centres, although only 56% had a specific target commissioning date.
BNEF estimated that if grid connections do not accelerate, about 48 GW of on-site gas generation could be needed by 2035.
Hence, the biggest data centres could increasingly function not just as electricity consumers, but as major energy projects with their own generation, storage and long-term power contracts.
5. US is racing to expand the grid
More power generation will not be enough if the electricity cannot reach data centres.
BloombergNEF raised its forecast for installed US data-centre capacity in 2030 to 118 GW, 52% higher than its December 2025 forecast. It expects capacity to reach 194 GW by 2035.
But the grid is already struggling to accommodate new projects. Texas, for example, had more than 470 GW of data-centre projects and other large energy users waiting to connect.
Carnegie cited a November 2025 survey in which 78% of respondents said they were concerned that new data centres would increase household electricity bills. In March 2026, major hyperscalers and AI companies signed the White House’s Ratepayer Protection Pledge, under which companies agreed to build, bring or buy new generation and pay for the power-delivery infrastructure needed for their data centres. Carnegie noted that the pledge did not include enforcement mechanisms.
AI’s power problem is becoming an infrastructure race
America’s AI expansion is creating demand for electricity at a scale that the existing power system was not designed to handle.
There is no single solution. Solar and batteries can add capacity quickly, natural gas can provide near-term reliable power, nuclear energy offers a longer-term source of large-scale electricity, and on-site generation can help data centres bypass some grid constraints.
But all of these solutions ultimately depend on the same thing: a bigger and more capable electricity network.
The EIA expects US electricity consumption to set new records in both 2026 and 2027.
For the AI industry, the challenge is no longer just building more chips and data centres. It is building enough power infrastructure to keep them running.