Firm or flexible, the load pays

A year ago, the question about AI-era load was whether the grid could deliver it. This fall, regulators, utilities, and grid operators started answering a different question: on what terms?

The thesis: a grid connection for a large load is becoming a contract, and it comes in two grades. Either way, the load pays. It pays up front, for upgrades, and for a decade or more. Under a pending Senate bill, it would pay for both the new grid and the old one. Then it chooses. Firm service is the premium grade, and it waits for wires and generation. Flexible service arrives sooner, with an obligation to step down when the system is stressed. Value moves to whoever can make load flexible, or firm it behind the fence.

The public record moved fast in October.

The bill moved to the load

Large-load tariffs are no longer the exception. The Edison Electric Institute's October 2026 summary counts 25 states with at least one approved large-load tariff and 8 more with tariffs pending. Its list of publicly announced projects alone represents more than $1.14 trillion in investment and more than 72 GW of connected load.

The terms are getting stricter. Researchers at Lawrence Berkeley National Laboratory and the Brattle Group, working from Halcyon's tracker of 264 large-load tariff filings, found that average minimum contract terms rose from five years in tariffs proposed before 2025 to 12 years in tariffs proposed after, per Utility Dive. Upfront study payments, ramp schedules, "hold harmless" provisions, and exit fees are becoming standard.

Federal regulators are writing the same idea into transmission rules. In October FERC approved Duke Energy Carolinas' large-load terms, which require new large loads to pay a committed revenue contribution equal to 75% of their scheduled megawatts for 15 years after the upgrades they need are in service. Commissioner Lindsay See framed the issue as how "to ensure large load pays its way." In PJM, the grid operator's amended reliability backstop filing would buy capacity for up to 15 years against a shortfall of roughly 6,800 MW and allocate the cost to the large loads driving it.

Congress may go further. The Senate permitting bill would require computational loads above 20 MW to pay both the embedded cost of the existing grid and the incremental cost of expanding it, Latitude Media reports. Two buyer groups that represent large tech companies have endorsed it. The vote is set for Nov. 9.

Some deals already look like that. Georgia Power says OpenAI will pay the full cost of the infrastructure to serve its 3,200 MW project, and the utility now pitches large-load revenue as savings for households.

Why it moved now: the bills are visible

EIA's October Short-Term Energy Outlook puts the average U.S. residential electricity price at 18.2 cents per kWh in 2026, up from 17.3 cents in 2025. It expects PJM wholesale prices to rise 41% this year, the largest increase of any region it tracks, and notes that residential prices also carry transmission upgrades and, in some regions, capacity auction costs. EIA attributes much of this year's wholesale spike to weather. Customers see the bill, not the cause.

Latitude's reporting cites POLITICO polling showing half of Americans now oppose a data center within three miles of home, up from 28% in January. When a customer class becomes a campaign issue, the price of its connection gets written into rules.

Firm power is the premium grade

The second shift is quieter, and it matters more for builders. Grid operators are starting to sell large loads a non-firm product.

FERC's June show cause orders to all six regional operators named five categories of reform. One gap was the lack of services tailored to flexible large loads that are willing to curtail. The suggested remedy: non-firm or contract demand service, plus a way to study nearby generation and load together. Responses are due Nov. 16, and Nov. 20 for SPP.

CAISO's board has already approved its answer. The proposal centers on a "flexible interim load interconnection." That would let a large load connect quickly, with enforceable flexibility based on grid conditions, until its network upgrades are complete. It also requires hourly forecasts, telemetry, and communication protocols to remotely disconnect large loads when reliability requires it. Texas got to a similar place by statute: under Senate Bill 6, loads of 75 MW or more must install equipment that allows remote disconnection during firm load shed, as Latitude Media explains in its reality check on flexible data centers.

The path is not settled. FERC rejected the core of PJM's plan to let large loads connect without matching new capacity if they accepted curtailment ahead of emergencies, finding that it appeared to intrude on state jurisdiction over retail rates and service. The commission's message was that states should decide who gets curtailed first. Commissioner David Rosner put it plainly: "We need state regulators to establish clear rules for which customers should be curtailed first when circumstances require it." Jefferies analysts said the ruling could keep PJM capacity prices higher for longer, which favors existing generators.

The grades are coming region by region. The details will differ. The direction will not.

Flexibility now has a price

The case for a flexible grade rests on a well-known number. Duke University's Nicholas Institute estimated that the existing U.S. system could absorb 76 GW of new load if that load could be curtailed for 0.25% of its maximum annual energy use. That rises to 98 GW at 0.5% and 126 GW at 1%. The authors call it a first-order estimate of generation headroom, not a transmission study.

Contracts are starting to put numbers on it. Under a 25-year agreement, OpenAI committed up to 1,000 MW of flexible demand response from that 3,200 MW Georgia project, which Georgia Power says lets it invest in fewer new generation resources. In PJM, Google expanded a virtual power plant deal with Voltus by 1,070 MW, to run through 2040. The same report cites the market monitor: demand response revenue in PJM jumped 240% in one year, to more than $520 million in 2025.

Flexibility is not free, and it is not universal. Latitude's reality check reports that an EPRI DCFlex field test using Emerald AI software cut a data center's power use by 25% during peak hours by orchestrating GPU clusters. It also quotes Impact ECI's Peter Hirschboeck warning that inference, now the dominant workload, can't be slowed the way training can. And PJM's market monitor has called flexibility a "regulatory fiction," arguing that without binding curtailment authority there are no guarantees. Flexible service works where the commitment is enforceable and the load can actually move.

Flexibility runs on iron

Here is where the energy lane meets the turbine lane. Much of today's real-world flexibility, Latitude's reporting notes, lives in on-site batteries and generators. A flexible data center is often one that can switch to its own power.

BloombergNEF's public summary of its July U.S. data center outlook is blunt. The most new data center demand the U.S. grid has connected in a single year is about 10 GW. If that rate does not accelerate, 48 GW of on-site gas must be built by 2035 to meet BNEF's base case. It counts 124 GW of announced on-site gas capacity for data centers, only 56% of it mapped to a commissioning date. (BNEF figures here come from its free public summary only, not subscriber research.)

Nat Bullard's June note on Halcyon's gas plant tracker shows the mix shifting toward speed. The tracker now follows more than 212 GW of U.S. gas capacity under development. Planned reciprocating-engine capacity rose to 18.3 GW from 1.85 GW a year earlier, and 31 assets are tagged as behind the meter for large loads.

Utilities are building the same bridge. FirstEnergy says it is weighing bridge generation, meaning reciprocating engines, fuel cells, and batteries that could serve a first 100 MW building and later move to another site. It has also reserved Siemens Energy turbines for 2030 delivery for a 1,200 MW plant. AEP's new plan includes $20 billion for gas generation and a 702 MW fuel cell project backed by a 20-year Google contract.

Our read: the firm grade is being built with large frames on next-decade delivery slots. The flexible grade is being built with whatever ships this decade. Both are dispatchable machines, and the flexible kind will be asked to start when the grid is stressed. That is peaking duty.

What this means for people building companies

Owners and operators. Your contract terms are now a product. Minimum bills, collateral, and curtailment rights decide which loads you can serve and when. Flexible service also lets you connect a large load before the big plant arrives. Price the backup and bridge assets behind those commitments as capacity, because to the system they are.

Independent service providers and their suppliers. Behind-the-meter engines, aeroderivatives, and small frames are a growing installed base, often owned by first-time owners. BNEF found that 52 of the 100 largest U.S. data center projects come from companies that have never built one. Machines that start on grid-stress signals need maintenance planned around starts and readiness, not just fired hours.

OEMs. Two order books are emerging. One is for firm baseload, measured in delivery slots. The other is for bridge and backup machines, measured in months to site and the ability to redeploy. Mobile, modular, and fast-start products fit the flexible grade's contracts.

Builders using AI. The flexible grade is a software problem as much as a hardware one. Forecasting, telemetry, verified curtailment, and workload orchestration are now written into tariff proposals. Show that a curtailment commitment is real, measured, and enforceable, and you are selling speed to power. If you can't verify the response, it is just a promise.

What Driftwood Club is watching

• Nov. 9: the Senate vote on the permitting bill and its large-load cost provisions, per Latitude Media.
• Nov. 10: EIA's next Short-Term Energy Outlook for the retail price trend.
• Nov. 16 and 20: the six RTO and ISO responses to FERC's large-load show cause orders, and how many offer a non-firm grade.
• March 1, 2027: the target launch of PJM's reliability backstop procurement.
• Bullard's Machine Readable and his new Nameplate podcast of utility CEO interviews, for how the people writing these contracts talk about them.

If you build, own, run, or service the machines and software behind this shift, Driftwood Club is for you. Join Driftwood Club. Tell us what you run or what you are building. We introduce members working on the same problems, one at a time, and we send the sourced reading each week.

The connection used to be a request. Now it is a contract. Read the terms.