The next line on your AI bill isn't a model price — it's a power bill

Wholesale electricity prices in the US have already risen 2-6% nationwide because of AI data center demand, according to Federal Reserve Bank of Dallas research published in August 2026 — and the same research projects electricity generation costs could climb 20-30% higher by 2028 than in a world without the AI buildout. None of that shows up on a per-token pricing page yet. It's heading there.

A cost that doesn't show up on the invoice — yet

The Dallas Fed's research found that existing AI data centers have already pushed average wholesale electricity prices 2% to 6% higher nationwide, with larger surges in the regions where facilities are concentrated. In the PJM grid region, home to roughly 67 million residents, electricity bills are running about 15% higher in 2026 than in a pre-AI-datacenter baseline. Energy costs make up roughly half of a typical retail electricity price, and while wholesale increases take time to fully reach household and commercial rates, the Dallas Fed's own framing is that the trajectory is a steady climb, not a plateau.

Where this is headed

The same research projects electricity generation costs could rise 20% to 30% by 2028 compared with a scenario where the AI data center buildout never happened. Goldman Sachs, working the same problem from a different angle, separately projects the AI infrastructure buildout adding 6% to electricity costs between 2026 and 2027, with another 3% by 2028. Different models, same direction: the input cost underneath every GPU and every inference call is getting structurally more expensive, and that trend has years left to run, not months.

States are already pulling the emergency brake

This isn't just a modeling exercise. Texas Governor Greg Abbott ordered state regulators to halt new data center connections to the power grid pending a full capacity audit. New York passed the Responsible Data Center Development Act, imposing a one-year moratorium on large data center permits and requiring utilities to create separate service classifications so infrastructure costs get assigned to data centers rather than spread across general ratepayers. When a state government literally pauses new grid connections for data centers, that's not a distant policy debate — it's a capacity constraint with an immediate effect on how much new GPU capacity can come online, and where.

Why this is a budgeting problem, not just a policy story

Every inference call runs on power that's getting more expensive at the generation level, and every grid capacity halt tightens the supply of new data center capacity available to lease. Token-price wars between model providers are real and have driven sticker prices down sharply over the past two years — but those prices are set by competition between vendors, which can discount margin. Electricity generation cost and grid capacity are physical constraints competition can't discount away. That pressure doesn't disappear; it moves upstream into infrastructure cost, and eventually into the price of the token itself.

What to watch for

• Rate limit or capacity waitlist changes from your provider — these often move before public pricing does, especially in regions facing grid constraints.
• Whether a provider's status or incident communications start referencing power or capacity constraints, rather than purely software or demand-driven issues.
• Reserved-capacity versus on-demand terms, if you self-host or use dedicated capacity — power cost pass-through tends to land there first, ahead of standard API pricing.

The bottom line

Nobody's model API bill has a line item for electricity generation cost — but the physics of where that power comes from doesn't disappear just because it's abstracted behind a per-token price. Wholesale electricity is already up, state grids are already constrained, and the projections point one direction through 2028. Budgets built entirely around today's competitive token pricing are missing the input cost that competition can't discount away.