While our Legislature has proposed electrical tariffs for owners of artificial intelligence in recognition of inevitable harm it will cause to non-AI ratepayers, lawmakers are simply whistling past the graveyard. Legislators and regulators must implement targeted procedures and regulations to insulate non-AI electric users from subsidizing AI participants.
Electric utility prices are structured to pay for variable costs of electric generation — fuel, operations and maintenance — plus those for transmission and distribution, plus reimbursement for the utilities’ infrastructure investment.
Through a process called integrated resource planning, the utility forecasts transmission and distribution infrastructure to meet demand. It bundles that projected cost into a “revenue requirement” — the amount needed to recover the variable costs, overhead and investment and to make a profit. Then the utility divides that revenue requirement by expected future kilowatt-hour usage from various customer classes to arrive at electric prices.
Utility rates are premised on spreading system expansion across all ratepayers. It’s an imprecise approach that essentially “socializes” infrastructure expansion costs: One customer class is subsidized by others.
Retiring and not replacing
AI consumes massive electricity due to the vast amount of processed data, the complexity of queries and the volume of requests. During training — that is, the research AI does to answer a query — the program “learns” based on large sets of examples and data. Training for AI can take a few seconds to several months.
Credit: Fred FastiggiDumping incremental kilowatt hours into a limited supply pipeline causes generation costs to rise. By 2030, data centers are projected to consume as much as 12% of U.S. electricity, according to a study by Harvard Law School’s Environmental and Energy Law Program.
New Jersey utilities procure their power via PJM Interconnection, a regional transmission organization that serves 13 states plus Washington, D.C. PJM attributes electric pricing increases to supply and demand and “unprecedented and continuing growth” of data centers, according to an Inside Climate News report.
When demand is flat, plant owners retire old, inefficient and unprofitable generation plants. At the same time, PJM is struggling with plant connection delays. The operator has an online queue for prospective plant owners to gain interstate hookup approvals. The line, though, is a tangle of long waits.
Stranded costs
Our own process here in New Jersey also contributes to higher electric prices. When data center growth is expected, utilities seek regulator permission for more infrastructure spending. Approvals lead to higher electric rates for all customers, given the “socialization” of expansion costs.
Saddling non-AI customers with the AI customers’ costs is distasteful enough. Another risk is stranded costs for unneeded infrastructure. Historically, utilities conservatively overestimate capacity needs based on promises from prospective energy users. Even if that growth doesn’t happen, the higher rates stay.
The legislature and regulators need a targeted approach beyond simply speculating on needs and setting tariffs promoted as protection for non-AI ratepayers. Indeed, AI customers must assume all business risk related to higher projected loads.
AI users should perform grid impact studies at their cost. If demand doesn’t materialize, AI customers should pay exit fees to offset the utilities’ investment If they exceed demand estimates, they should be penalized with tariff step-ups.
Cutoff threat
An alternative is for utilities to service AI users only if they commit to flexible operations that limit infrastructure demand. The tariff provision under consideration needs to be strengthened. A utility should have the option to cut AI power if the grid becomes overloaded. That possibility alone should be data centers’ incentive to install sufficient and reliable backup systems or to tolerate power interruptions.
Regulators also need to relax the policy that restricts power distribution mostly to utilities. This requirement is a remnant of outdated thinking meant to protect utility service territories as regulated monopolies for reasons which were valid in Thomas Edison’s time but less so today. A foolproof way to protect non-AI energy users is to isolate AI data centers from utility-owned grids. This is easily done with micro grids, which aren’t interconnected to utility distribution facilities.
Once a grid impact study is done — and the direct costs are clear — AI users can then choose service from a utility, build their own grid or abandon development in New Jersey.
