The Curtailable Connection: How Flexibility Became the Price of Grid Access for AI Data Centres

Beneath the consensus that the AI data-centre boom is a power-supply problem demanding new generation, US and British regulators are quietly re-pricing grid access around demand flexibility, turning large loads into dispatchable resources and exposing hyperscalers, utilities, operators and regulators on a 2026-2030 horizon.

The consensus on AI and electricity is a supply story: data centres are outrunning the grid, so the answer is to build generation faster. Beneath it, a quieter pivot has taken hold: what now decides whether a large load gets connected, and how fast, is not how much new power can be built but how much the load will give back. US and British regulators and grid operators are now conditioning fast interconnection on curtailability, treating data centres as dispatchable demand. The inflection is 2026, as this moves from pilot to rule. The question for boards: are you negotiating for firm power, or for the right to be interrupted?

Signal Identification

This is a regulatory-and-market reframing rather than a single rule: the price of grid access is being rewritten around flexibility. It surfaces in interconnection dockets, reliability assessments and connection reforms, beneath the build-more-generation narrative. The signal is the conversion of large loads into curtailable, grid-interactive resources as the condition of getting connected.

What's Changing

The federal rulebook is being rewritten around curtailability. On 16 April 2026 FERC said it would act by June on its large-load docket (RM26-4), asking whether loads above 20 MW that agree to be curtailable should clear interconnection studies faster, potentially within 60 days, after reviewing over 3,500 pages of comments (FERC, 16/04/2026). The economic case is on the table: a Nicholas Institute analysis at Duke puts the system savings from large-load flexibility at $40 billion to $150 billion over the next decade (Nicholas Institute, Duke, 17/02/2026).

The shift shows up in the numbers that govern reliability. NERC's 2026 Summer Reliability Assessment cut its ERCOT net-internal-demand forecast by 3.7 GW, or 4.6%, "because more data centers can be curtailed", and lowered total internal demand by 2.3%; Texas now requires large loads of 75 MW or more interconnecting from 2026 to accept mandatory curtailment during firm load-shed events (NERC, 19/05/2026). The capability is proven: EPRI's DCFlex demonstrations sustained roughly 25% load reduction without interrupting AI service, with a reference design that could free up around 100 GW nationally (EPRI, accessed 16/06/2026).

The market has read the signal, and Britain is moving in step. In supply-constrained regions, an ICF analysis argues, connection now comes with a choice: bring your own power or bring flexibility; layering flexibility lets PJM absorb an extra 30 GW without new generation, and retail demand response can be cheap, at $32/kW-year against $100+/kW-year for behind-the-meter generation (Utility Dive, 09/06/2026). In Great Britain, DESNZ is weighing a mandatory minimum level of demand flexibility for very large users, and 81% of data-centre projects say they would accept non-firm or phased connections for an earlier date (CMS, 18/03/2026).

Disruption Pathway

The pathway runs in three stages. Gating: data centres queue for firm connections and operators build new supply to serve them. Conditioning: regulators trade a faster path for curtailability, as FERC and Britain now do, so the load that accepts interruption jumps the queue. Default: flexibility becomes the baseline, written into tariffs and, in Texas, into law, leaving the firm connection the costly exception.

Stresses concentrate at three points: the clash between AI service-level guarantees and curtailment; cost allocation, as commissions decide who pays for the grid the boom demands; and enforcement, since a curtailment promise is worth no more than its telemetry and penalties. Two adaptations follow. Operationally, workload orchestration and on-site batteries let a data centre meet a curtailment call as a battery discharge, not a service cut. In policy, curtailability is priced into connection terms and reliability planning, so a megawatt of avoided peak competes with a megawatt of new generation.

Why This Matters

For data-centre boards, utilities, operators and regulators, the signal moves the binding question from how much new power can be built to how much load will flex. The assumption to revise is that a data centre is firm demand to be served at any cost: the quicker, cheaper route to capacity is the load that agrees to be interrupted, and the connection contract is becoming a flexibility contract. Hyperscalers should treat curtailability as a design parameter, not a concession, and investors should test whether projects underwritten on firm power survive when flexible peers connect years sooner and cheaper.

Decision-action posture for this signal: Prepare — the rules, mandates and forecasts are converging now, but flexible connection is not yet the default; build the capability to be curtailable, and commit on a trigger such as FERC's final rule or curtailment terms entering your next connection offer.

Implications

This looks like a durable change in how grid capacity is allocated, not a passing response to one demand spike. The inflection window is 2026-2030, as FERC's rule, Texas's mandate and Britain's reforms set precedents others copy. Those who gain are utilities that can dispatch flexible load and operators that engineer curtailability cheaply; those who lose are projects underwritten on firm power. The deeper shift is that a unit of flexible demand is becoming a substitute for a unit of new generation, and grid access is priced accordingly.

Early Indicators to Monitor

• FERC issues a final large-load interconnection rule that formalises an expedited path for flexible, curtailable loads.
• A second US state or ISO follows Texas in mandating curtailment for large new loads above a defined threshold.
• DESNZ confirms a mandatory minimum demand-flexibility requirement for data centres in Great Britain.
• A hyperscaler signs a flexible or non-firm connection agreement and discloses it as standard practice.
• A utility launches a tariff paying large loads for curtailability as firm-capacity equivalent.

Disconfirming Signals

• FERC's final rule drops or dilutes the flexible-load fast lane after industry objections.
• Data centres decline non-firm offers at scale, and queues stay dominated by firm-power requests.
• Curtailment events trigger AI service-level breaches that push operators back to firm connections.
• Cost-allocation rulings shift the burden so heavily onto large loads that self-supply beats flexibility.
• The data-centre slowdown deepens, easing grid stress and the pressure to mandate flexibility.

Strategic Questions

• Do you engineer curtailability into new data centres now, or bid for firm power and risk a slower, costlier connection?
• Should utilities procure flexible load as a capacity resource, or keep building generation to serve it as firm demand?
• When does a flexible connection move from optional discount to the only offer on the table?

Keywords

Data centre flexibility; demand response; curtailable load; large-load interconnection; FERC RM26-4; grid headroom; AI data centres; non-firm connections; ERCOT curtailment; DESNZ demand connection reform; dispatchable demand; speed to power

Bibliography

Source tiers: Tier 1, governments, regulators and intergovernmental bodies. Tier 2, think-tanks, academic institutes, major consultancies and quality data providers. Tier 3, quality journalism and specialist trade press. Tier 4, vendor, company and practitioner sources, used only as directional corroboration.

• Tier 1 FERC to Act on Large Load Interconnection Docket by June 2026 (RM26-4). Federal Energy Regulatory Commission (16/04/2026).
• Tier 1 2026 Summer Reliability Assessment. North American Electric Reliability Corporation (19/05/2026).
• Tier 2 Flexibility Will Go a Long Way Toward Managing the Grid of the Near Future. Nicholas Institute, Duke University (17/02/2026).
• Tier 2 DCFlex: Data Center Flexible Load Initiative. Electric Power Research Institute. Evergreen reference page, accessed 16/06/2026.
• Tier 3 How load flexibility buys time for America's data center boom. Utility Dive (09/06/2026).
• Tier 3 Spring 2026 update on further Demand Connection Reform for Data Centres. CMS (18/03/2026).
• Tier 3 More than 75 data center build-outs worth $130 billion blocked in 2026. Tom's Hardware (08/05/2026).