PJM Files 50 MW Rule to Cut AI Data Centers First During Grid Shortages
PJM filed with FERC on July 31 to make new AI data centers the first in line for power cuts during shortages. The grid math: 70 GW of projected new demand, 15 GW of retired generation, and capacity prices up roughly 11x in two years.
The grid operator serving 67 million Americans just told hyperscale AI compute it is last in line when the lights start flickering.
Key takeaways
- PJM Interconnection filed with FERC on July 31, 2026 to make new loads of 50 MW or more, including AI data centers, the first in line for curtailment during capacity shortages, ahead of households, effective June 1, 2027.
- The grid math is broken: PJM projects roughly 70 GW of new large-load demand by 2038 against roughly 15 GW of generation retired since 2022, with capacity prices rising from approximately $28.92/MW-day (PJM 2024/25 BRA report) to approximately $329.17/MW-day (PJM Inside Lines) in two years.
- New covered loads can avoid curtailment only by securing their own generation under a "Bring Your Own New Capacity" mechanism, a mandate that materially raises the all-in cost of building a GPU cluster inside PJM's footprint.
PJM Interconnection, the grid operator covering 13 states and Washington, D.C., filed a proposal with the Federal Energy Regulatory Commission on July 31, 2026 that would formalize a new curtailment hierarchy: any new customer drawing a cumulative peak load of 50 MW or more at a single site is first in line for power cuts during capacity shortages, behind households and existing demand-response customers. The filing responds to a December 18, 2025 FERC order that found PJM's existing tariff "unjust and unreasonable" for lacking clear rules on large co-located loads.
The proposal, which PJM calls the Interim Resource Adequacy Service (IRAS), takes effect June 1, 2027 for new large loads that do not bring their own generation or secure supply. Existing facilities are not affected.
The Grid Math
PJM's own numbers explain why this filing exists. The operator projects roughly 70 GW of new large-load demand by 2038. Since 2022, approximately 15 GW of generation has retired. That gap is not a rounding error.
Capacity prices reflect the stress. The market cleared at approximately $28.92/MW-day for the 2024/25 delivery year (PJM 2024/25 BRA report). By the 2026/27 delivery year, that figure had reached approximately $329.17/MW-day (PJM Inside Lines), roughly an 11-fold increase in two years. PJM's Market Monitor has attributed a significant share of those increases to data center demand, with figures on total ratepayer cost attribution flagged for editorial verification against the underlying Monitoring Analytics report before publication.
Alongside the IRAS filing, PJM filed a separate Reliability Backstop Procurement seeking new resources for a 6,831 MW shortfall in the 2028/2029 delivery year, with contracts up to 15 years and a maximum willingness to pay of $555/MW-day, per PJM's Inside Lines.
PJM will also create a mandatory registry of all 50 MW-plus loads by location, service area, and generation status, shared with states, utilities, and regulators. One important limit on the framework: PJM acknowledged it lacks authority to curtail individual sites directly and would rely on utilities and state governments to carry out reductions.
The White House's Ratepayer Protection Pledge, introduced in March 2026 and expanded in July 2026, aligns with the IRAS model. Its language, cited in PJM's filing overview, requires new large loads to "build, bring, or buy the new generation resources and electricity needed to satisfy their new energy demands."
Rigid Load Meets an Elastic Policy
The IRAS framework encodes a distinction the grid has always known but never formally priced into policy: some load can flex down on short notice, and some cannot.
Hyperscale AI compute is rigid. A training run cannot pause mid-epoch without consequences. A GPU cluster cannot curtail 80% of its draw in two minutes and restart cleanly an hour later. The economics of AI infrastructure depend on utilization rates that leave almost no headroom for voluntary curtailment.
Bitcoin mining is the opposite. Miners can curtail instantly, cheaply, and without destroying product. That property is not incidental; it is the operational foundation of every demand-response contract miners have signed with grid operators for years. PJM's IRAS hierarchy, which places voluntary flexible load above involuntary rigid load when shortages hit, is a regulatory confirmation of exactly that hierarchy. Miners already operating demand-response contracts inside PJM territory (Pennsylvania, Ohio, Virginia, West Virginia) are structurally compliant with a standard that just made their AI-compute competitors legally subordinate during grid stress.
The BYONC mandate raises the stakes further. Any new 50 MW-plus load that wants to avoid curtailment must secure qualifying capacity resources on its own. For a hyperscale operator, that means contracting for generation, building on-site power, or purchasing capacity in an already-stressed market clearing at over $300/MW-day. That cost does not appear in a GPU manufacturer's marketing materials, but it is absolutely real, and it lands on top of interconnection queues that phantom load requests have already stretched to the breaking point. OpenAI recognized the trajectory early: the company hired a power-trading lead specifically to manage energy procurement as Stargate's capacity demands became impossible to ignore.
Bitcoin mining's energy cost structure, stress-tested across a decade of adversarial conditions, looks disciplined by comparison. Miners site where power is cheap and available, including stranded gas fields, curtailed wind corridors, and hydroelectric surplus. AI cannot move a training cluster to a stranded methane site. The geographic flexibility that has always made mining look like a liability (no fixed address, always chasing cheap electrons) is now, inside PJM's new framework, a competitive advantage.
What to Watch
FERC must approve the IRAS filing before it takes effect. The agency has previously signaled it wants PJM to resolve its large-load framework, and the December 2025 order that triggered this filing was explicit on that point. Approval is not guaranteed on the current terms, and large-load operators with FERC counsel are already reviewing the tariff language. The trigger that would collapse the mining-versus-AI differentiation: if FERC or future PJM rulemaking imposes equivalent curtailment obligations on demand-response-enrolled miners, or if AI operators rapidly achieve demand-response flexibility through widespread on-site generation and storage. Neither outcome is imminent, but the Texas data center queue fight and ongoing ERCOT debates suggest every major grid region is running the same calculus PJM just formalized.
Sources
- PJM Inside Lines: IRAS Data Center Framework Filing
- PJM Inside Lines: Reliability Backstop Procurement Proposal
- PJM Inside Lines: Board Directs Action on Large Loads
- PJM Inside Lines: 2026/27 Base Residual Auction Results
- PJM 2024/25 RPM Base Residual Auction Report
- PJM FERC Filing Document, July 31, 2026 (Accession #20260731-5214)
- FERC Order, December 18, 2025
- FERC Fact Sheet: December 2025 Order
Frequently Asked Questions
The rule, filed with FERC on July 31, 2026 as the Interim Resource Adequacy Service, applies to any new customer with a cumulative peak load of at least 50 MW at a single site within a one-mile radius inside PJM's territory. Those loads become the first in line for curtailment during capacity shortages, ahead of households and existing demand-response customers, unless they secure their own qualifying generation under the BYONC mechanism. The rule takes effect June 1, 2027. Existing facilities are not affected.
BYONC is the escape valve inside the IRAS framework. A covered load that secures its own qualifying capacity resources, whether through on-site generation, a long-term power contract, or purchased capacity in PJM's market, moves out of the mandatory curtailment-first tier. For AI data center developers, that means the all-in cost of a new facility inside PJM territory now includes procuring capacity in a market that cleared at approximately $329.17/MW-day for the 2026/27 delivery year (PJM Inside Lines), on top of standard interconnection and construction costs.
A Bitcoin mining operation drawing 50 MW or more at a single PJM-territory site without its own generation or capacity contract is technically covered by the rule. In practice, miners operating existing demand-response contracts are in the preferred tier: the IRAS framework curtails non-flexible loads before calling on paid demand-response resources. Miners without supply arrangements at covered sites would be natural BYONC candidates or demand-response partners given their intrinsic curtailability, not reluctant curtailees facing the same exposure as a rigid hyperscale operator.
