AI Data Centres Drive 76% Power Price Surge as Australia Watches

The AI revolution has a power bill, and it landed in the first quarter of 2026. Wholesale electricity prices on the PJM Interconnection — the largest US electric grid, serving 67 million people across 13 eastern states — jumped 76 per cent year-on-year to an average of $US136.53 per megawatt-hour, driven almost entirely by the ballooning energy appetite of the data centres that train and run large language models.

The numbers are stark. Capacity costs alone surged 398 per cent in the quarter, adding roughly $US13 billion ($19.3 billion) to customer bills across the last two PJM capacity auctions. The grid’s independent market monitor, Monitoring Analytics, did not soften the finding in its quarterly State of the Market report.

The price impacts on customers have been very large and are not reversible.

That phrasing — “not reversible” — was delivered in the context of a demand trajectory the watchdog expects to persist through at least 2028. The report pinned the surge squarely on data centre load growth, calling it “the primary reason for recent and expected capacity market conditions, including total forecast load growth, the tight supply and demand balance, and high prices.”

But Monitoring Analytics’ diagnosis of the cause is not universally shared. SemiAnalysis, the semiconductor and infrastructure research firm, argues that PJM’s capacity auction mechanism — the Base Residual Auction, a simulation-heavy construct run three years ahead of delivery — is itself the root problem. In a detailed March analysis, SemiAnalysis contended that the BRA produces artificial scarcity, and that poor market design, not AI demand alone, is what converts load growth into ratepayer cost shocks.

The cost pass-through mechanism is unusually blunt. Under PJM’s tariff, the cost of new generation and transmission built to serve data centre load is socialised across all customers in the relevant zone. When Maryland’s Office of People’s Counsel filed a complaint with FERC in early May, it calculated that Maryland households would pay roughly $US1.6 billion in additional transmission costs for projects driven by out-of-state data centres — a bill that arrives regardless of whether those households ever use an AI chatbot.

The capacity-price trajectory tells the story in numbers. In the 2024/25 delivery year, PJM capacity cleared at $US28.92 per megawatt-day. For 2026/27, it hit $US329 — an 11.4-fold increase — and data centres accounted for 63 per cent of the price rise in the 2025/26 auction alone, adding $US9.3 billion to ratepayer costs. The grid’s interconnection queue now holds roughly 1,200 gigawatts of generation projects waiting for study — a backlog that can stretch five years — and Google has told Reuters that grid connection is now its single largest infrastructure obstacle.

FERC Chair Laura Swett crystallised the regulatory frustration at the commission’s annual meeting in mid-May, calling PJM “too big to function” and announcing a July 2026 technical conference to examine market-design reforms. Pennsylvania Governor Josh Shapiro has gone further, threatening to withdraw his state — the second-largest load centre in PJM — from the grid operator entirely if governance changes are not adopted.

Pennsylvania is no longer going to be held captive to PJM. We put forth some very specific proposals I wanted to see them do to reform themselves. They have not yet adopted those, and I’ve been very clear that they’re either going to adopt them or they’re going to lose Pennsylvania.

— Josh Shapiro, Governor of Pennsylvania

The political temperature around data centres is rising on other fronts, too. A Gallup survey published in May found that 71 per cent of Americans oppose data centre construction in their area — a higher opposition rate than for nuclear power plants. Maine has already passed a data centre moratorium, and at least 12 other states are considering similar bills. In the first quarter of 2026 alone, an estimated $US41.7 billion in data centre projects were cancelled.

For Australian readers, the PJM story is not an American curiosity. The Australian Energy Market Operator’s data, compiled by Oxford Economics Australia, shows data centre electricity consumption reached 3.9 terawatt-hours in FY25 and is forecast to hit 12 TWh by FY30 under AEMO’s Step Change scenario — a compound annual growth rate of 25.1 per cent. Victoria alone recorded a 94 per cent year-on-year increase in data centre demand in the first quarter of 2026, reaching 187 megawatts.

AGL Energy has warned that Australian planners are sharply underestimating the surge. The Australian Financial Review reported in early May that AGL’s internal modelling shows demand growth running well ahead of AEMO’s central forecast, raising the prospect of supply crunches in the National Electricity Market that look structurally similar to PJM’s. The difference — and it is a significant one — is that Australia’s energy-only NEM has no capacity auction to amplify the price signal. Whether that market structure insulates households, or merely delays the same cost shock through a different mechanism, is the question Australian regulators are only beginning to ask.

The emerging regulatory consensus in the US is coalescing around a “Bring Your Own Generation” model — a requirement that hyperscale data centre operators finance and build dedicated generation capacity rather than drawing from the common pool and socialising the infrastructure cost. Monitoring Analytics has demanded BYOG mandates for new large loads. CSIS notes that the core puzzle is cost allocation: proposed backstop auction rules would assign 100 per cent of new plant costs to data centres “whether they show up and use the power or not,” creating a financial incentive to forecast demand honestly.

For Australia, where Equinix and NextDC are expanding aggressively and the Greens have launched a Senate inquiry into data centre energy consumption, the question is whether the regulatory framework is being built fast enough to get ahead of the load. The PJM experience suggests the answer is no — and that by the time the bill arrives, it is already too late to reverse.