AI's power squeeze: PJM's backstop auction and Google's nuclear deal

On 31 July 2026 PJM Interconnection, which operates a grid serving more than 67 million people across 13 US states and the District of Columbia, asked federal regulators to approve a one-time auction for new power capacity. Its regular capacity auction, with results published on 14 July, had come up 6,831 MW short of the level PJM needs to meet its reliability standard. A second filing on 13 August went further. It proposed that, in a grid emergency, new data centres which do not bring their own supply should have their demand reduced before ordinary consumers, including households, are cut off.

In Finland, a different answer to the same problem arrived on 9 September. Google announced €13 billion of investment in the country across 2027 and 2028, and the utility Fortum said Google had signed a 22-year agreement to buy power from the Loviisa nuclear plant, covering up to half of its capacity.

Together, these moves show where the binding constraint on AI infrastructure increasingly sits. Accelerators can be ordered; firm electricity in the right place at the right time is harder to secure, and grid operators are moving to make large new users carry more of the cost and risk. Both of PJM's proposals still await decisions from the Federal Energy Regulatory Commission (FERC), and the curtailment plan has already drawn a formal protest from a power-supplier group.

How a capacity market works

Most people think of electricity markets as paying for energy: the kilowatt-hours actually consumed. PJM also runs a second market that pays for capacity. That is, in effect, a promise that enough power plants and flexible customers will be available on the most demanding days of a future year, arranged well before those days arrive.

The mechanics are simpler than the jargon. PJM forecasts peak demand, adds a reserve margin and sets a reliability requirement. It then holds an auction in which resources offer to be available for a future delivery year, with prices quoted in dollars per megawatt per day. Resources that clear the auction are paid for being available, and those payments ultimately flow through to electricity bills, although customers hedged through self-supply or bilateral contracts are not exposed to the auction price.

PJM's benchmark is a standard of one reliability event in ten years. In the auction for the 2028/29 delivery year it secured 138,318 MW of capacity, and even with capacity from areas that arrange their own supply, the total fell 6,831 MW short of the requirement. PJM says a shortfall of this kind does not mean it will be unable to serve customers. It means operating with thinner reserves and more risk.

Prices were held down as well. PJM, working with the governors of all 13 of its states and with FERC, set a price floor and cap for four auctions to limit volatility, and this auction cleared at the cap of $325 per MW-day across the whole footprint. PJM said the two most recent auctions were the first in its history in which the entire region fell short, following a gap of about 6,500 MW for 2027/28.

A capped price and a persistent gap point to the same conclusion: the regular market is not bringing on new supply quickly enough. Only 525 MW of new generation and uprates cleared in the July auction. PJM's chief executive David Mills summarised the problem in the release: 'demand for electricity continues to grow faster than electricity supply'.

PJM's backstop: a one-off purchase of new supply

The Reliability Backstop Procurement filed on 31 July is PJM's attempt to fill that gap with new resources. According to PJM, it would secure capacity for terms of up to 15 years, with a maximum willingness to pay equal to a megawatt-weighted average of $555 per MW-day. That is well above the $325 cap in the regular auction. Utility Dive reported that bidding would run from 30 September to 21 October, with results due by 2 December.

The contract length is the important design choice. A developer weighing a new plant is unlikely to build on the strength of a single year's capacity payment. A commitment lasting up to 15 years looks much more like the long-term offtake agreements that make energy projects financeable.

Bilateral deals sit alongside the auction. PJM began facilitating long-term agreements between large users and generation providers on 9 June, often running for ten years or more and covering new generation, storage or demand-side resources. If data centres sign such contracts before the auction, the amount PJM needs to buy can be reduced, which avoids procuring the same capacity twice.

Who pays is only partly settled. Utility Dive reported that costs would be spread to PJM zones in proportion to their share of the procurement target, then to utilities and other load-serving entities within each zone. PJM notes that states have authority over how those costs reach retail customers. That leaves open how much of the bill lands on data centres and how much on everyone else.

The pressure behind all of this is load growth. PA Environment Digest's account of PJM's August filing attributes 30 GW of the 32 GW of forecast demand growth between 2024 and 2030 to data centres. Utility Dive reported a PJM estimate that data-centre and other large-load demand could rise by as much as 70 GW by 2038.

Bring your own power, or be curtailed first

The 13 August filing, for what PJM calls Interim Resource Adequacy Service, deals directly with the users driving the shortfall. It would apply to new large loads that have neither brought their own new capacity nor had their needs covered through the backstop procurement. Utility Dive reported that the rules would affect data centres coming online after 1 June 2027.

In practice, the proposal sets an order of priority for grid emergencies. When supply falls to dangerously low levels, a new emergency procedure would have utilities curtail or move demand from those new large loads before any action that would disconnect ordinary consumers, households included. A large load ordered to cut back could receive credit for supporting reliability.

PJM would also maintain a registry covering every large load, old and new, which states and utilities could draw on when deciding whether to apply the service. From the auction for the 2029/30 delivery year, which PJM has scheduled for December, new large loads without their own new supply would not be counted when PJM calculates how much capacity to buy. PJM asked FERC to accept the filing within 60 days.

If FERC approves it, the message to developers is plain. A data centre that arrives with contracted new supply joins the grid on familiar terms. One that does not may still connect, but it accepts a curtailment risk that other customers do not carry.

Finland: buying into existing nuclear output

Google's announcement of 9 September describes €13 billion of investment in digital infrastructure, clean energy and local partnerships in Finland, which it calls its largest single investment in Europe. The company has run a data centre in a converted paper mill in Hamina for 15 years. It projects that construction in 2027 and 2028 will support more than 37,000 jobs and add €3.6 billion a year to Finnish GDP, and it plans €31 million of community funding in Hamina, Kajaani, Muhos and Vaala over four years.

The power arrangements are set out in Fortum's release, published as inside information. The 22-year agreement begins in 2028 at a smaller volume and reaches 50% of Loviisa's capacity for 2030 to 2049. Loviisa supplies about 10% of Finland's electricity, roughly 8 terawatt-hours a year. Fortum said the contract gives it the revenue certainty to keep the plant running until its operating licences end in 2050, and that it enables a new 10 MW power increase on top of a 38 MW uprate already expected in 2028. The release does not disclose a price.

Google is also contracting a 94 MW battery system, which it says will help steady prices during cold, windless spells. ESG Dive reported that the battery is due to connect to the grid near one of Google's data centres in late 2027. The same report noted that Google has blamed its AI build-out for an 18% increase in its carbon emissions in a single year, set against a 2030 goal of running on carbon-free energy around the clock.

The logic mirrors PJM's. Keeping an existing reactor running delivers firm, low-carbon output far sooner than new construction can, and a buyer willing to sign for decades makes that extension investable. The large user underwrites supply instead of drawing on whatever spare capacity the grid happens to have.

In both markets, large new loads are being asked to pay for the supply they depend on.

What is disputed or still pending

Neither PJM proposal is in force. Both need FERC approval, and the backstop timetable, with bidding due to open on 30 September, depends on the commission acting in time. Until FERC rules, the auction dates, the $555 cap and the curtailment order should all be read as proposals.

The curtailment plan faces organised opposition. On 3 September the Electric Power Supply Association, which advocates for competitive power supply, filed a protest. It said it supports tackling the shortfall and does not oppose the registry, but argued that requiring new large loads to bring new capacity treats new and existing power plants unfairly, would distort capacity prices and could raise consumer costs over time. It proposed instead that utilities short of capacity designate load or contracted demand response for curtailment.

Some of the numbers are softer than they look. The 70 GW figure is an upper estimate, and PJM's own peak forecast for 2028/29 was about 2,000 MW higher than the one used for the previous auction. Whether a $555 cap attracts enough new supply to close a gap of this size in time is untested, given how little new generation cleared in July.

In Finland, the contract is signed but the investment programme behind it is not complete. Fortum said about 80% of the projects, and €700 million of the capital spending needed to keep Loviisa running to 2050, still await investment decisions. The jobs and GDP figures in Google's announcement are the company's own projections, and one trade report gave a larger figure for the additional uprate than Fortum's release; this article uses Fortum's number.

What this means for teams running AI and automation

For organisations that deploy AI agents, robotics or automation in real operations, grid constraints show up indirectly: in where providers can add capacity, what it costs and how reliably it runs under stress. Several practical steps follow.

• Ask where your workloads run. Capacity and pricing are likely to diverge by region as grid operators tighten terms for new data centres, so know which regions your providers use and what fallback exists.
• Check the resilience terms. If PJM's curtailment proposal is approved, some new facilities could be cut back ahead of other customers in emergencies; ask providers how that would affect service levels and failover.
• Separate urgent from deferrable work. Interactive voice agents and control systems need steady capacity, while evaluation runs, reindexing and batch analytics can be scheduled away from peak demand.
• Make efficiency a requirement. Smaller models, caching and fewer redundant calls cut the energy used per task and reduce exposure to regional power constraints.
• Keep critical functions local where it matters. Robots, vehicles and site systems that must keep working through outages or network loss benefit from on-device or edge inference, with the cloud as a supplement.
• Read energy deals as timelines. The Loviisa contract starts in 2028 and reaches its full contracted share in 2030, so it does not change near-term capacity; plan around the dates rather than the headlines.