Power Is Now the Binding Constraint on AI Growth

For most of the AI boom, the scarce input was chips. In 2026, it is electricity. Power, grid connections and community consent now decide where AI capacity gets built and how fast. For executives planning AI workloads and for regions competing for investment, the grid has become a strategy question.

The demand is real, and so is the hype

The IEA reports that data center electricity use jumped 17% in 2025, while overall global demand grew 3%. It expects data center consumption to double by 2030. In the US, the EIA expects record electricity generation in 2026 and 2027, with data centers and manufacturing driving commercial and industrial growth.

The utility view is bigger still. EPRI projects data centers could reach 9% to 17% of US electricity demand by 2030, according to Utility Dive. FirstEnergy alone now reports a 26 GW long-term data center pipeline.

But a pipeline is not a power plant. Goldman Sachs expects only about half of near-term scheduled capacity to come online on time, and Grid Strategies found that utility forecasts filed with FERC ran about 40% above credible industry projections (Utility Dive). Separating real load from speculative requests is now a core job for regulators.

The queue is the new bottleneck

Texas shows the scale. The state paused new data center interconnections while ERCOT audits a 474 GW large-load queue, more than five times the grid's all-time peak. BloombergNEF estimates the pause could delay a fifth of the US pipeline (Utility Dive).

Equipment is a second queue. Standard power transformers take two to three years, and GE Vernova's gas turbine backlog tops 100 GW with deliveries into the early 2030s (Utility Dive). Hubbell says its data center sales rose more than 80% and it is raising 2027 prices early (Utility Dive).

Who pays is the fight that matters

In PJM, the largest US grid, the 2028/29 capacity auction cleared at the $325 per MW-day cap and still came up 6,831 MW short of its reliability requirement. A study for the Pennsylvania PUC warned that PJM reliability could fall far below its planning standard by 2030 if data center growth continues.

Federal regulators are focused on cost shifts. In June, FERC issued show cause orders to six grid operators covering faster studies, cost transparency, co-location and services for flexible loads, while affirming state authority over retail sales and siting. In October, FERC rejected most of PJM's large-load plan, saying parts of it intruded on state jurisdiction.

States are filling the gap. Large-load tariffs increasingly include minimum contract terms, collateral and exit fees, and Florida now requires data centers over 50 MW to cover the full cost of grid connection (Utility Dive).

Bring your own power

Many developers are not waiting for the grid. Cleanview counted more than 90 GW of behind-the-meter projects across 59 large sites (Utility Dive). OpenAI's planned 8 GW campus in central Ohio is backed by 10 GW of new generation, mostly gas.

Nuclear is back in the mix. The Department of Energy is lending up to $4.4 billion to restart Palisades, Crane and Duane Arnold, adding 2.25 GW, with Microsoft and Google as buyers for two of them (AP). The IEA says conditional offtake agreements for small modular reactors grew to 45 GW, though most of that capacity is years away.

Flexibility is the new currency

The fastest megawatts may be the ones a data center agrees not to use at peak. Google has built 1 GW of demand response into utility contracts, including with Indiana Michigan Power, TVA, Minnesota Power and DTE Energy. Portland General Electric freed 80 MW of headroom using grid software (Utility Dive).

Reliability cuts the other way too. In July, about 3,800 MW of northern Virginia data center load tripped offline at once, and new NERC standards for computational loads are on the way.

Communities are voting

On Nov. 3, voters decided 41 local data center measures in 10 states, and Ohio had 21 of them. Results were nearly unanimous for restrictions, from a 25 MW cap in Ottawa, Kansas, to a required public vote on a proposed campus in Janesville, Wisconsin.

For heartland regions, this is the real competition. Land and transmission matter, but so do clear tariffs, credible power plans and a community benefit story residents believe.

What leaders should do now

• Plan AI capacity on power timelines, not chip timelines. Ask cloud and colocation partners where their power comes from and when it is actually energized.
• Know your state's large-load rules. Tariffs, collateral and curtailment terms are being written now and will set costs for years.
• Treat flexibility as an asset. Workloads that can shift or pause may connect faster and cost less.
• Earn local consent early. Regions and developers that lead with transparency on bills, water and jobs are the ones still winning projects.

Follow the latest on power, policy and regional readiness in the ETA AI Power and the Grid hub.