PJM Plans Data Center Power Cuts as Grid Pressure Mounts

PJM Interconnection plans to curtail some data centers starting in June 2027, despite their growing importance to cloud computing and artificial intelligence. The policy applies to facilities drawing at least 50 megawatts, according to the curtailment plan. This techcrunch data center report captures a sharp change in how the largest US grid treats major computing loads.

The change does not mean operators will casually switch off servers. Curtailment would occur during power shortages, with affected customers compensated for reducing demand. The deeper reversal is harder to ignore: electricity access is becoming conditional for facilities built around continuous availability.

PJM serves 67 million people across all or parts of 13 states and the District of Columbia. Its territory includes Northern Virginia, the world’s largest concentration of data centers. The operator must now balance household reliability against an industry adding large electrical loads faster than generators and transmission infrastructure can arrive.

That conflict already moved beyond planning documents in 2026. PJM sought emergency federal authority to move data centers onto backup generation during periods of extreme grid stress. It also warned utilities in July to prepare for that action, although PJM ultimately did not need to use it.

The June 2027 policy turns an emergency concept into a more durable operating tool. It also sends cloud providers, data center developers, utilities, regulators, and enterprise customers the same message. Computing capacity can grow only as fast as the power system can support it, unless operators accept interruptions or supply more energy themselves.

TechCrunch Data Shows What PJM Is Changing

PJM is converting large data centers from protected growth engines into loads that can be managed during a supply emergency.

The reported policy covers data centers and other large electricity users with demand of at least 50 megawatts. A facility at that threshold can consume as much power as a sizable industrial operation. Clusters containing several buildings can place far greater demands on local transmission networks.

Curtailment means temporarily reducing or disconnecting an electrical load when available supply cannot safely satisfy demand. It sits near the end of the grid operator’s emergency toolkit. PJM can first call additional generators, purchase energy, activate contracted demand response, or request voluntary conservation.

Under the planned approach, affected customers would receive compensation for reducing consumption. That structure resembles demand response programs long used by manufacturers, commercial buildings, and other large customers. Participants trade some operational flexibility for payments and, in some cases, earlier access to constrained infrastructure.

Notice can vary with the emergency and program design. Some events provide meaningful preparation time, while fast-developing conditions leave less room to respond. Data center operators must therefore decide which workloads can move, pause, or run from another power source.

The effective date matters. PJM does not expect the new curtailment arrangement to begin until June 2027. That gives operators time to negotiate contracts, test backup systems, revise workload schedules, and determine whether on-site generation can carry critical computing equipment.

It also gives PJM time to run another capacity auction. A capacity auction pays generators to remain available during a future delivery period. It is designed to secure enough dependable supply before the electricity is actually needed.

The previous auction did not deliver PJM’s full target. PJM said resources procured for the 2027 and 2028 delivery year fell 6,600 megawatts, or 5.6 percent, below its Installed Reserve Margin target. That margin represents the supply cushion maintained above forecast peak demand.

This shortfall does not predict an automatic blackout. Grid reliability depends on weather, generator performance, transmission conditions, actual demand, imports, and emergency resources. However, it leaves less space for unexpected failures or rapid growth.

PJM had already directed stakeholders to design an interim “connect-and-manage” framework. Under that model, new large loads can connect before sufficient generation catches up, provided someone accepts responsibility for the resulting supply risk. PJM’s integration framework says curtailments should be infrequent, limited, and used only to prevent broader system impacts.

That distinction is important. The policy is not a blanket judgment that data centers deserve unreliable service. It is an allocation rule for moments when keeping every large customer fully supplied would increase blackout risk for everyone else.

The immediate change is contractual, but its effects will reach facility design. A project’s electrical connection can no longer be treated as an unlimited resource that arrives before its supporting generation. Developers must attach an operating plan to every large power request.

Why the Largest US Grid Needs Relief Now

Data center construction is compressing an infrastructure cycle measured in years into a computing cycle measured in months.

PJM coordinates a regional transmission system containing more than 88,000 miles of lines. Its footprint stretches from Virginia to Illinois and includes major metropolitan, industrial, and cloud-computing markets. That diversity normally allows the region to share supplies across a wide area.

Data centers are changing the demand profile. Unlike a house or conventional office, a large computing campus can draw substantial power throughout the day. AI training, inference, storage, and cooling equipment create loads that do not necessarily fall when residential demand eases.

The challenge is not electricity consumption alone. It is the speed, scale, location, and concentration of new requests. A generator can take years to permit, finance, connect, and build. Major transmission lines often face similarly long reviews and construction schedules.

A data center developer can plan facilities much faster. The mismatch becomes acute when several projects request service in the same constrained zone. Generation elsewhere cannot solve the problem if transmission lines cannot deliver it safely to that location.

PJM’s May 2026 summer assessment forecast a typical peak of about 156,400 megawatts. It listed approximately 180,200 megawatts of generation capacity and 7,800 megawatts of contracted demand response. The operator nevertheless prepared for plausible demand as high as 169,100 megawatts.

Those totals should not be read as a simple subtraction exercise. Some generating resources may be unavailable, limited by fuel, affected by weather, or trapped behind transmission constraints. Operators also need reserves that can respond when a large power plant or line unexpectedly fails.

PJM recorded its all-time summer peak of 165,563 megawatts in 2006. During June 2025, demand reached approximately 161,300 megawatts on one day and 160,900 megawatts on another. PJM used demand response to help control those peaks.

Senior Vice President of Operations Michael Bryson described a “new reality” in PJM’s summer outlook. He said data center load growth was outpacing the addition of generation, tightening reserves and increasing risk.

Extreme weather exposes that narrowing cushion. Air-conditioning demand rises sharply during prolonged heat, just as equipment can experience temperature-related stress. A generator outage or transmission failure can then transform a manageable forecast into an operating emergency.

PJM faced such conditions in May 2026. The operator forecast about 135,000 megawatts of demand while 40,400 megawatts of generation was expected to be unavailable for maintenance or other reasons. It projected fewer than 5,800 megawatts of reserves for the May 18 peak.

The resulting federal request described a specific sequence. PJM wanted utilities to disconnect data centers from grid power and transfer them to backup generation after most other tools had been exhausted. That step would occur before firm load interruption, which can include involuntary outages.

July brought another stress test. During the July 2 evening peak, generators tripped offline while heat and humidity drove demand higher. PJM activated emergency demand response and warned some utilities to prepare for possible data center curtailments.

The grid stabilized without that final action. Yet the warning proved the option was operational, not theoretical. PJM again secured temporary federal authority to direct large facilities with backup generation away from utility power before voltage reductions or broader load shedding.

This techcrunch data story therefore reflects accumulated operational pressure. The new policy follows repeated moments when PJM needed large loads to become flexible. It is not based solely on a distant forecast about AI electricity use.

The Core Tradeoff Is Uptime Versus Shared Reliability

PJM is asking data centers to surrender a small portion of guaranteed grid access so the wider system can preserve reliability.

Data centers sell availability. Cloud platforms, financial systems, communications services, healthcare applications, and AI products depend on servers remaining reachable. Even a short interruption can affect transactions, automated processes, and users far beyond the facility.

Grid operators have a different obligation. They must continuously balance production and consumption across an interconnected system. If that balance breaks, frequency and voltage can move outside safe ranges, potentially causing equipment to disconnect automatically.

The planned policy places these responsibilities in direct tension. A data center may view uninterrupted utility service as a basic requirement. PJM views temporary reduction at a few exceptionally large sites as preferable to uncontrolled or rotating outages affecting communities.

The distinction between a power cut and a service outage is crucial. Many facilities use uninterruptible power supplies, batteries, and generators to bridge a utility interruption. Their computing services can remain online even when their grid connection is curtailed.

However, backup systems were primarily designed for local failures, not routine participation in regional electricity markets. Operators must confirm fuel availability, emissions permissions, maintenance schedules, switching behavior, and runtime under real emergency conditions.

Workload flexibility offers another path. Some computing jobs can shift between regions or run later without harming users. Model training, batch analytics, media processing, and certain maintenance tasks may tolerate delay better than live inference or database transactions.

Geographically distributed cloud providers can route work toward facilities with more available electricity. That approach requires spare computing and network capacity elsewhere. Moving demand also cannot help when several regions face the same heat or supply shortage.

Smaller operators face a different calculation. They may lack multiple campuses or sophisticated workload orchestration. A 50-megawatt threshold limits direct exposure to large facilities, but tenants inside those buildings can still inherit operational and contractual consequences.

Enterprise buyers should examine how cloud and colocation providers define force majeure, availability, and curtailment. A service-level agreement can promise application uptime while leaving the provider broad freedom to choose its energy source. Customers need to understand both layers.

The policy may also change where developers build. Locations with stronger transmission, nearby generation, or clearer interconnection rules become more attractive. Constrained markets may retain their network advantages while losing some appeal for new power-intensive campuses.

Northern Virginia illustrates the conflict. Its dense fiber networks, established cloud ecosystem, skilled workforce, and proximity to major customers make it difficult to replace. Those benefits also concentrate demand in the part of PJM already managing exceptional data center growth.

A July 22 disturbance showed why grid planners worry about concentration in both directions. More than 3 gigawatts of data center demand suddenly disconnected after a transmission-line event in Northern Virginia. That represented about 3 percent of PJM demand at the time, according to a grid disturbance account.

Dominion Energy said facility controls transferred the affected data centers to backup power. PJM reported a frequency change but no reliability impact. Operators stabilized the system, demonstrating that rapid load loss can be managed.

Yet the event revealed a second risk. Grid discussions often focus on demand becoming too high. A huge block of demand disappearing at once can also disturb the balance because generators cannot instantly reduce output without coordinated controls.

Planned curtailment offers more control than an automatic mass disconnection. PJM can choose timing, communicate with utilities, and account for the expected load reduction. That coordination becomes increasingly valuable as individual computing campuses approach the scale of traditional power plants.

The main opponent in this story is therefore not data centers versus households. It is unconditional computing growth versus system reliability. Households, businesses, cloud customers, and data center operators all lose when the grid cannot control either rising demand or sudden disconnections.

Backup Power Solves One Risk and Creates Another

Moving a data center off the grid can protect regional reliability while transferring pollution, fuel, and operational risks to the facility’s neighbors.

Diesel generators remain common because they start quickly, can store fuel on-site, and have a mature service network. Those traits make them useful during an unexpected utility failure. They do not make diesel an ideal source for recurring grid support.

Combustion produces nitrogen oxides, particulate pollution, greenhouse gases, and noise. The effects become more significant when many generators operate simultaneously across a dense data center corridor. Nearby communities can experience the local costs of a decision made to protect a much larger grid.

Federal rules distinguish emergency generator use from certain non-emergency operations. According to the source report, applicable regulations can allow limited annual operation for demand response and additional hours for emergencies and maintenance. Actual permissions depend on equipment, permits, jurisdiction, and event classification.

The environmental concern is not an argument for accepting blackouts. It is a reason to scrutinize what replaces grid electricity. A curtailment program can reduce one system risk without delivering a clean or durable supply solution.

Natural-gas engines and turbines offer another on-site option. They can run longer than many diesel systems and may emit less of some pollutants. They still require fuel infrastructure, permits, maintenance, and careful integration with utility equipment.

Batteries can switch almost instantly and avoid local combustion during operation. Their practical duration depends on system size and facility demand. A battery designed to bridge the seconds before generators start cannot necessarily support a campus through hours of regional scarcity.

Some facilities may combine batteries, generators, and software-managed workloads. Batteries handle the transition, flexible computing reduces total consumption, and generators support the remaining critical load. This layered design can limit fuel use while preserving service.

It also adds capital and operational complexity. Every component must work during the same stressful conditions that triggered the grid emergency. A control error could disconnect too much load, fail to shed enough, or produce an unsafe reconnection.

That concern became harder to dismiss after the Northern Virginia event. More than 3 gigawatts transferring away from the grid within a short period was manageable, but it was not trivial. Future controls must consider how thousands of backup systems behave collectively.

PJM’s approach also raises a fairness question. Existing customers helped fund a system designed around earlier demand forecasts. New data centers create economic activity and tax revenue, but their scale can require new generation and transmission that take years to deliver.

If ordinary customers fund those upgrades while data centers receive continuous service, political resistance will grow. If data centers pay all associated costs, developers may argue that projects become slower or less competitive. Curtailment functions as an interim compromise.

Compensation complicates that bargain. Payments can encourage useful flexibility and recognize the operational burden placed on participants. Regulators must still ensure those payments reflect genuine reliability value and do not shift excessive costs to other customers.

Transparency will matter. The public needs to know when curtailment occurs, how much demand was reduced, which resources replaced grid electricity, and what the program cost. Without those details, supporters and critics can each select numbers that favor their position.

The techcrunch data center account correctly identifies pressure toward more on-site power. What remains uncertain is which technologies developers will choose. Diesel is available now, while cleaner systems may require more planning, space, permits, and interconnection work.

The policy should not become a substitute for new generation and transmission. Backup systems are insurance against limited events. Treating them as the permanent foundation for AI growth would exchange a regional capacity problem for local pollution and reliability problems.

Cloud Providers and AI Companies Now Face a Power Constraint

The new rule turns electricity flexibility into a product, infrastructure, and procurement requirement for the AI sector.

For hyperscale cloud companies, the first response will likely involve portfolio planning. A provider with facilities across several grid regions can direct new computing deployments toward locations with greater power availability. It can also reserve constrained campuses for workloads that need their network position.

AI developers may feel the effects through computing schedules and availability. Training a large model can require sustained operation across many accelerators. Pausing and restarting that work is possible, but it can waste time and complicate cluster utilization.

Inference workloads create a different challenge. They serve live requests from users and applications, so latency and availability matter more. Providers can route traffic across regions, but doing so requires available capacity and compliant data-handling arrangements.

This makes energy awareness part of cloud architecture. Companies already design around network failures, hardware faults, and regional outages. They may now need operating policies for predictable grid stress, including heat waves and capacity emergencies.

A searchable record of decisions, incident reports, and supplier commitments becomes more important as these dependencies multiply. Teams managing such projects can use an engineering knowledge base to connect technical evidence with operational choices.

Data center developers also face new financing questions. Lenders and tenants will want to know whether a project has firm electrical service, conditional access, or a curtailment obligation. Each arrangement changes expected utilization and risk.

Power contracts may become as central as land and fiber agreements. A campus with dedicated generation, storage, or a credible flexibility plan can offer tenants greater confidence. One relying entirely on a constrained utility connection may face delays or less favorable terms.

Utilities will carry much of the implementation burden. They need accurate facility-level demand information, tested communications, and switching procedures. PJM operates the regional system, but local utilities physically connect most customers.

Generators also face pressure. Capacity prices signal that dependable electricity is scarce, yet higher prices do not build plants immediately. Developers must navigate fuel supply, equipment queues, environmental reviews, local opposition, and PJM’s interconnection process.

The data center industry can reasonably argue that computing infrastructure supports essential services and economic growth. A facility may host emergency communications, government systems, medical applications, and ordinary business operations alongside AI workloads. Curtailment rules must avoid treating every server as expendable.

Grid planners can answer that physical reliability cannot depend on how valuable each customer considers its own workload. When supply is insufficient, operators need predetermined actions. Improvised negotiations during an emergency would be slower and less predictable.

This is why compensated, rule-based curtailment is preferable to vague expectations. Operators know what response is available, while customers can prepare before accepting service. The arrangement makes the constraint visible instead of hiding it until a crisis.

The competitive effects will not be equal. Large cloud providers can invest in generation, batteries, workload routing, and specialized energy teams. Smaller data center companies may depend more heavily on utilities and commercial backup systems.

Large providers may therefore turn energy control into an advantage. Owning or contracting dedicated power can reduce exposure to curtailment. It can also accelerate access when the public grid lacks enough capacity for another fully firm connection.

That strategy carries its own risk. Behind-the-meter generation, where a facility receives power without first sending it across the regional grid, can affect costs and reliability for other users. Regulators are still working through how such arrangements should pay for shared infrastructure.

The practical lesson for enterprise buyers is not to abandon cloud or AI projects. It is to treat power availability as a supply-chain dependency. Vendors should explain where critical workloads run, how facilities respond to curtailment, and what redundancy exists outside the affected region.

This techcrunch data report marks the point when that discussion becomes difficult to postpone. AI infrastructure planning can no longer assume electricity is an invisible input. Power is now a schedule, location, compliance, and resilience constraint.

Three Signals Will Show Whether the Plan Works

The policy succeeds only if PJM gains dependable flexibility without turning temporary emergency measures into a permanent substitute for infrastructure.

The first signal is PJM’s next capacity auction. Readers should watch how much dependable generation clears, whether the reserve shortfall narrows, and how prices respond. A stronger result would reduce the likelihood that curtailment becomes frequent.

A weak auction would reinforce PJM’s central warning. Demand would still be growing faster than accredited supply, leaving flexibility programs to cover a structural gap. That outcome could produce tougher connection rules for future data centers.

Auction totals need context. A megawatt of capacity is useful only when the resource can perform during stressed conditions and deliver power through available transmission. Location, accreditation, fuel security, and outage performance matter alongside the headline number.

The second signal is the final design for June 2027 curtailments. PJM and its stakeholders must define eligibility, notice, compensation, dispatch order, testing, measurement, and restoration. Those details determine whether the program is an orderly reliability tool or a source of disputes.

A credible framework will identify which facilities can switch safely and how much load they can remove. It will also coordinate reconnection, since thousands of machines returning simultaneously could create a new demand spike.

The framework should distinguish workloads and power sources without asking grid operators to manage individual applications. Data center owners remain responsible for deciding what stays online. PJM needs a reliable quantity of load reduction at the requested time.

Public reporting will help test the program’s claims. PJM should disclose aggregate curtailment duration, demand reduction, compensation, and avoided emergency actions. Environmental agencies should track generator operation and local emissions under applicable permits.

The third signal is how developers redesign new projects. Announcements involving batteries, dedicated generation, flexible interconnection, or regional workload shifting would show the market responding. Heavy dependence on additional diesel capacity would expose the policy’s environmental downside.

Real operating evidence matters more than promotional commitments. Developers should demonstrate transfer tests, fuel arrangements, battery duration, and coordinated restoration. Grid reliability cannot rest on equipment that has only been modeled or briefly tested.

Watch Northern Virginia closely. Its concentration of data centers makes it the clearest test of both load growth and coordinated controls. Another sudden multi-gigawatt disconnection would increase pressure for common technical standards.

The July disturbance also shows why flexibility must work in both directions. PJM needs the ability to reduce demand during shortages, while preventing excessive demand from disappearing without coordination. Smarter controls must solve both problems.

For developers, the choice is becoming explicit: wait for firm grid capacity, accept curtailment, or bring dependable energy resources. Each path carries different costs, schedules, and environmental effects. None creates electricity instantly.

For cloud customers, the useful next step is a resilience review. Ask providers whether a contracted service can move between regions, what happens during grid curtailment, and whether backup generation has been tested under realistic loads.

For policymakers, the central test is whether temporary flexibility buys time for durable investment. New generation, transmission upgrades, storage, and faster interconnection remain necessary. Curtailment can protect the system while those assets arrive, but it cannot replace them.

The broader techcrunch data center story is not that AI servers will suddenly go dark. It is that the electrical system has started placing conditions on their growth. That shift will influence facility locations, cloud contracts, computing schedules, and energy investments well beyond PJM.

Before approving the next major AI deployment, teams should ask one direct question: what happens when the grid cannot provide every requested megawatt? A credible answer should cover workload priorities, alternate regions, backup duration, emissions, contracts, and restoration. If a provider cannot supply that answer, its computing capacity is less dependable than its dashboard suggests.