PJM Power Curtailment 2026: What UK Buyers Should Do

What PJM has actually proposed

PJM Interconnection, which coordinates wholesale electricity across 13 US states and Washington DC, has proposed curtailing supply to certain new large power users, including AI data centres, during grid emergencies from June 2027. The trigger is size and preparedness: PJM defines a Large Load as an end-use customer drawing at least 50 MW at a single site within a one-mile radius. New Large Loads that fail to bring their own generation, or otherwise secure supply, by that date will face curtailment ahead of PJM's standard Pre-Emergency Load Management steps.

PJM is also building a Large Load Registry, tracking sites, service areas and whether each facility brings its own power supply, so operators can manage load more precisely during capacity shortages. The scale of the pressure behind this is stark: PJM forecasts new large loads growing by roughly 70 GW by 2038, while around 15 GW of generation has retired since 2022 — a widening gap that the operator says is now a reliability issue, not just a planning one.

Why a US grid decision matters to UK infrastructure buyers

No UK regulator has proposed anything equivalent yet. But the underlying pressures — AI-driven demand growth outpacing generation and transmission build-out — are the same forces already straining the UK data centre power and energy landscape. PJM's move shows a mature, well-funded grid operator concluding that voluntary backup arrangements are no longer sufficient and that conditional access is now a viable regulatory tool.

That is the real signal for UK buyers: not that PJM has ordered blanket curtailment today, but that a major grid operator is actively building reliability mechanisms, faster interconnection routes and new cost-allocation rules specifically because of data centre load. Ofgem and National Grid ESO face structurally similar constraints on transmission capacity and connection queues, and precedent from one major market tends to inform policy debate in others.

Who is exposed first

According to Pareekh Jain, CEO at EIIRTrend & Pareekh Consulting, the facilities most at risk are new builds that haven't yet contracted their own power supply, particularly smaller or newer operators without the capital to self-fund backup generation. Hyperscalers such as Amazon, Google and Microsoft are comparatively insulated because they can afford dedicated on-site power. The highest-risk locations are grid clusters already under strain — Northern Virginia, plus growing regions in Ohio, Pennsylvania and Maryland.

Jain also flags a downstream risk for anyone renting rather than owning capacity: "If a company rents computing power from a data center instead of owning one, their service could get shut off during a power emergency, and if a bunch of data centers are all in the same region, they could all get hit at the same time." That is directly relevant for UK enterprises leaning on US colocation or cloud regions concentrated in the same power-constrained corridors.

Stress-testing DR and multi-region strategy now

Building dedicated on-site generation before mid-2027 is largely unrealistic. Jain notes that ordering generation equipment can take three to five years given backed-up factory order books, before permitting and grid interconnection delays are even considered. The only credible near-term levers are incremental: battery backup, and formally agreeing to participate in demand response curtailment during emergencies.

UK buyers should treat this as a planning trigger rather than a US-only concern. Practical steps include asking providers directly whether a specific facility has contracted or independent power secured, rather than assuming generic backup exists; running workloads through a proper backup and DR sizing calculator to understand real failover capacity; and reviewing backup and disaster recovery solutions against scenarios where an entire regional cluster, not just a single site, loses priority access to power.

It's also worth separating workloads by criticality now. Jain's advice — deciding which compute tasks need uninterrupted power and which can tolerate a pause — applies equally to UK estates running storage solutions for AI and analytics or dense GPU clusters, where power headroom is already tight, as covered in our look at hosting a Blackwell AI rack in the UK.

The wider grid picture buyers should watch

PJM's curtailment proposal sits inside a much larger reliability push. The operator has reported a 6.8 GW shortfall in its 2028/2029 delivery year capacity auction and warned it may not have enough generation to meet peak demand plus reserves from mid-2027, against reporting of a possible 60 GW shortfall over the next decade. It has approved an $11.8 billion transmission expansion plan and a further $5.9 billion in projects, including a 765 kV multistate backbone through Virginia, Ohio and West Virginia, alongside Dominion Energy Virginia's proposed 525 kV, 185-mile underground transmission line.

PJM is simultaneously reworking interconnection rules: a FERC-approved fast-track process allows up to ten large interconnection requests a year, requiring at least 250 MW and delivery within three years, running to the end of 2027. It has also asked FERC to approve a 50 MW threshold for behind-the-meter and colocated generation, with a three-year transition period, meaning new loads above that size won't be able to net against on-site generation. Separately, PJM's Critical Issue Fast Path process is developing market measures ahead of the 2028/2029 capacity auction, responding to forecast data centre peak load growth of around 30 GW through 2030.

Energy intelligence firm Currence estimates that between 30% and 50% of large-scale US data centre capacity expected online in 2026 will likely be delayed, largely due to power constraints, with only around 5 GW currently under construction against roughly 11 GW still stuck at the announced stage. For anyone sourcing capacity via server configuration partners or evaluating a Dell server configurator, HPE server configurator or Lenovo server configurator build against a US-hosted deployment timeline, this delay risk should now be a standard due-diligence question.

Bottom line for UK buyers

PJM hasn't ordered broad data centre shutdowns — it has built a conditional-access framework, a load registry, faster interconnection paths and new cost-allocation rules, all directly caused by AI compute demand outrunning grid capacity. Jain expects other grid operators facing similar AI-driven demand to adopt comparable policies, which would make power availability a core factor in site selection and investment decisions well beyond PJM's footprint.

UK buyers don't need to wait for an equivalent domestic proposal to act. Verifying supplier power commitments, running DR scenarios that assume regional rather than single-site failure, and reviewing AI server research against realistic power headroom are sensible steps now, while the US precedent is still fresh and the UK's own connection queues remain under similar pressure.