Curtailment Is Now the Price of Data Center Grid Access

For most of the past decade, the hardest question in a data center power deal was “when can you deliver megawatts?” In 2026, grid operators are adding a second question that matters just as much: “will you turn some of them off when we ask?” PJM, ERCOT, and FERC are all moving, in parallel, toward a model where fast grid access is conditional on a data center’s willingness to curtail load during system stress — and buyers who treat that as fine print rather than a deal term are going to get an unpleasant surprise in their first emergency event.

Why Grid Operators Are Suddenly Asking Data Centers to Power Down

The shift is a direct response to demand growth that outpaced every forecast on the books. PJM’s own planning shows peak demand rising roughly 32 GW between 2024 and 2030, with about 30 GW of that coming from data centers, and its board has warned new large-load demand could add another 70 GW by 2038 against roughly 15 GW of generation retired from the footprint since 2022. The supply side hasn’t kept up: PJM’s 2028/2029 Base Residual Auction cleared 6,831.3 MW short even at the $325/MW-day price cap, and the 2027/2028 auction cleared 6,623 MW short at $333.44/MW-day. In plain terms, the region is now paying the maximum allowed price for capacity and still can’t buy enough of it.

ERCOT is watching a similar curve from a different angle. The grid operator projects data center load could reach 187 GW by 2030 — more than double its current peak of about 91 GW — and NERC’s 2025 Long-Term Reliability Assessment raised its 10-year U.S. summer peak demand forecast by 69% over the prior year’s projection. NERC issued a Level 3 Alert on large-load reliability in May 2026, and in June FERC ordered all six U.S. grid operators to justify or rewrite their large-load interconnection rules, a process we covered in detail when FERC set its August 17 deadline. In July, FERC went further and directed NERC to draft mandatory computational-load reliability standards by year-end.

PJM’s New Rules: IRAS, BYONC, and What Happens If You Say No

PJM’s answer, filed with FERC on August 13 (docket ER26-3515), is the clearest template for where this is headed. The proposal creates an Interim Resource Adequacy Service (IRAS), a Large Load Registry, and new capacity-market mechanics — all targeted at any “Large Load,” defined as end-use demand with a cumulative peak of at least 50 MW at a single site. PJM wants FERC sign-off by October 12 so IRAS can start June 1, 2027.

IRAS inserts a new step into PJM’s emergency response sequence, ahead of calling on existing firm customers: before touching legacy load, PJM would first dispatch generation and economic demand response, curtail non-firm interchange, and tap non-firm transmission service. As PJM’s Christopher Pilong put it, “in real-time, IRAS provides a new step, at PJM’s disposal, prior to impacting existing customers” — and reductions only happen when real-time conditions actually require them, not automatically.

The mechanism that matters most for a buyer’s term sheet is Bring Your Own New Capacity, or BYONC: a large load can cut or eliminate its IRAS exposure by financing qualifying new generation, but that capacity has to be genuinely new and must offer into RPM auctions as a price taker for 10 consecutive delivery years. Skip BYONC and get curtailed, and compensation is capped at 50% of PJM’s Non-Performance Charge Rate — a rate that can be set lower, or waived to zero by the customer. Starting with the 2029/2030 delivery year, any large load’s uncovered incremental demand gets pulled out of the capacity auction’s demand curve entirely until it secures BYONC or backstop capacity. Pennsylvania has already moved to codify a version of this at the state level: Governor Josh Shapiro’s August 18 directive instructs regulators to require utilities to curtail data centers ahead of other customers during emergencies, tighten permitting above 25 MW, and assign backstop costs directly to data center tariffs.

What the Economics Actually Say About Flexibility

The industry’s instinct is to treat curtailment as pure downside. Grid modeling suggests the opposite is closer to true, at least for load that’s designed to flex. Analysis from Relae using ERCOT-calibrated modeling (CD-PyPSA-USA) found that at 40 GW of data center load, a grid with no flexibility hits forced outages of 158 GWh over 81 hours — a roughly $5.5 billion annual hit at ERCOT’s $35,000/MWh value-of-lost-load benchmark. Add flexibility — modeled at 25% emergency curtailment capability, matching performance Emerald AI has already demonstrated in production — and forced shedding disappears entirely across every scenario tested, replaced by controlled curtailment events averaging under 5% of nameplate load, with the single largest event reaching 14%, and all of it confined to under 1% of annual hours.

Separately, Duke University’s Nicholas Institute found that 22 major U.S. balancing authorities could absorb roughly 98 GW of new flexible load at just 0.5% annual curtailment, or 76 GW at 0.25%. That’s the strongest argument buyers have for negotiating flexibility into a deal rather than fighting it: done right, it’s a materially cheaper way to get interconnected than waiting in a multi-year queue or self-funding a gas plant.

Early Movers Are Already Cutting the Deals

A handful of hyperscalers aren’t waiting for FERC to finish rulemaking. Google announced roughly 1 GW of demand-response contracts in March 2026 with Entergy Arkansas, Minnesota Power, and DTE Energy. Emerald AI has publicly demonstrated cutting a 256-GPU cluster’s power draw by 25% for three hours during an actual Arizona grid stress event without halting training jobs. EPRI’s DCFlex initiative is actively building standardized flexibility protocols and contract language with both hyperscalers and utilities, aiming to turn ad hoc deals into a repeatable framework. Texas has gone the regulatory route directly: SB6, signed in mid-2025, requires any new load of 75 MW or more to install remote curtailment equipment and creates a voluntary demand-response program with a minimum 24-hour notice period.

The common thread is that flexibility is becoming a currency operators can spend to buy speed. A data center that can credibly commit to shedding load — through on-site generation, battery buffering, GPU power capping, or workload shifting — is a cheaper, faster interconnection customer than one that can’t, and grid operators are starting to price that difference explicitly rather than leaving it to goodwill.

What This Means If You’re Signing a Capacity Deal in 2026

Three things belong on every buyer’s diligence list right now. First, ask your provider or utility directly whether your facility, or the campus you’re buying into, will be classified as a “Large Load” under whatever version of IRAS or its regional equivalent ends up in force, and get the 50 MW threshold and curtailment compensation terms in writing — a verbal assurance that “it won’t affect you” isn’t worth much against a FERC-approved tariff. Second, treat BYONC-style options as a genuine financing decision, not boilerplate: committing to bring new generation is a 10-year obligation with real capital behind it, and the alternative — uncompensated curtailment during exactly the hours your workloads matter most — needs its own cost estimate before you sign. Third, if your workload can tolerate some flexibility (batch training, non-latency-sensitive inference, deferrable analytics), use that as leverage in negotiations rather than conceding it for free; the Relae and Duke findings both suggest utilities and RTOs have real money to pay for it, and early movers like Google are already capturing that value instead of leaving it on the table.

None of this is settled law yet — PJM’s filing is pending FERC review, other RTOs are still drafting their own versions under FERC’s June order, and state-level rules like Pennsylvania’s are directives, not finished tariffs. But the direction is no longer ambiguous: the price of fast grid access in 2026 and beyond increasingly includes a flexibility clause, and the buyers who negotiate it on their terms will come out well ahead of the ones who find out about it during their first curtailment notice. For a structured way to work through power, siting, and contract risk before you sign, start with the TechInfraHub Data Center Buyer’s Toolkit.

Written by

Raajeev Ratra

Data Center Infrastructure Expert | 15+ Years in DC Design, Operations & Project Management

Raajeev is a seasoned data center professional with hands-on experience in hyperscale facilities, colocation design, power & cooling infrastructure, and global DC operations. He shares practical insights to help engineers and IT leaders build better infrastructure.

Connect on LinkedIn →