AI faces a power problem. It takes an incredible amount of power to run it. Demand in the United States is growing faster than the electric grid can keep up, giving companies that generate and supply that demand enormous leverage.
On June 2, the Electric Reliability Council of Texas voted to overhaul the way it welcomes large-scale power users onto the power grid, while addressing backlogs such as data centers, crypto mines, and industrial sites that reach the same megawatt capacity.
The same week, lawmakers in Albany, New York, were rushing to pass a one-year moratorium on new large data centers that could make the state the first in the nation to completely halt construction.
Companies training frontier models continue to run into walls built by copper, concrete, and regulatory patience. The beneficiary of all this demand is the humble entity on the other side of the wire. That means power companies, grid operators, and power producers who decide who gets electricity, when they get it, and at what price.
Electricity has become the scarcest asset for AI
For most of the past decade, every conversation about AI revolved around software, and the most important constraint people were concerned about was the availability of advanced GPUs.
Now we turn to the industrial economy, where the limited inputs are land, generating capacity, water, high-voltage transformers, and local switchboards.
Goldman Sachs predicts that U.S. data center power demand will increase from 31 GW in 2025 to 41 GW in 2026 and 66 GW in 2027, with data centers’ share of U.S. summer peak demand rising from 4.1% to 8.5% over the same period.
However, the bank noted that due to delays and cancellations, only about 50-60% of its planned capacity over the next one to two years is likely to arrive on time. Even at a discount, the grid is being asked to absorb in two years what would normally take 10 years to add.
The International Energy Agency predicts that data center electricity usage will nearly double by 2030, and demand from AI-centric facilities will triple. The report focuses on bottlenecks, from strengthening supply chains for gas turbines and transformers, to grid connections that take years to break into on-site generation, which remains largely on paper.
Utilities now have incredible influence. Regardless of which company wins the race, the utility company collects. All it takes is for the race to keep demanding more power. Regulated utilities can profit from approved capital investments, so each wave of grid upgrades becomes a wave of rate-based revenue.
Independent power producers sell to tougher markets, but at higher prices. Grid operators have finite connection capacity and are the gatekeepers in determining which projects are viable.
Texas shows how gatekeeping can turn into rules. Under Senate Bill 6, ERCOT currently employs a “self-pay” model in which large customers pay for interconnection costs and shut down service in the event of an emergency. There is a non-refundable fee of $50,000 per megawatt and large deposits to eliminate speculative claims.
The burden cannot be overstated, as nearly 200 large-scale users lined up in the first few months of 2026 alone, seeking a total of 438 gigawatts, more than five times what the entire state currently consumes.
New York state’s proposed moratorium approaches a similar issue from a political perspective, weighing the growth of AI data centers against household bills, water usage, and power grid reliability. Electricity becomes a rationed source, and the party doing the rationing has the most power at the table.
Bitcoin miners were the first to witness this battle, and now everyone will pay up.
The Bitcoin market is used to this bottleneck because miners were the first to face it. The mining industry built its business on cheap, interruptible electricity, using flexible loads that switch off when the grid is stressed and absorb the surplus when prices spike.
That’s why Texas created a new demand-response program around it, and why miners spend years chasing wasted watts to windy plateaus and hydro-power spillways, where energy is often stranded and cheap. Some analysts go further, arguing that the grid should welcome that flexibility as a service given how quickly miners can cut back.
That’s almost the opposite of what AI wants and needs. Hyperscalers want stable, always-on power and long-term certainty, backed by jobs and national competitiveness arguments with real political weight. When BlackRock warned in January that AI data centers could consume up to 24% of U.S. electricity by 2030, the company effectively declared the truce on cheap power over.
The CryptoSlate analysis, which compared the energy footprint of streaming, AI, and cryptocurrencies overall, reached a similar verdict, with miners currently facing a severe squeeze as AI companies drive up the price of corporate supplies.
Utilities now arbitrate the dispute and profit from it no matter how it is resolved.
As utilities build out generation and transmission to meet the demands of AI hyperscalers, ratepayers could end up paying some of the costs unless regulators lock in those costs or force them to pay more to cover their own share.
Federal projections are already trending that way, with EIA projecting U.S. electricity usage to set new records in 2026 and 2027. Home prices are already up 5% in 2026, with the steepest increases concentrated along the East Coast.
AI promises abstraction, where intelligence is rendered as weightless, infinitely copyable software. That expansion has made electricity a scarce commodity that determines who scales, who sets prices, and who collects the checks, regardless of which companies control most of the market. While power companies will keep a tight grip on the meters, businesses will continue to follow the headlines.
