Coal Is Powering the AI Revolution — And Nobody Put That in the Keynote

Published as part of iQ-LOOP's Insights series.

Every AI keynote this year has featured the same slide: soaring inference volumes, falling cost-per-token, a roadmap toward some greener, more efficient future. What none of them mention is where the electricity behind those tokens is actually coming from. Increasingly, in the United States, the answer is coal.

The number nobody expected to go back up

US coal-fired electricity generation rose 13% in 2025, according to the Energy Information Administration. That's not a rounding error — it's a reversal of a decade-long decline, and outside the western US, nearly every state generated more coal power in 2025 than it did the year before. The knock-on effect is visible in the emissions data too: carbon emissions from the US power sector rose 4% last year, roughly double the 2% increase across the economy as a whole. The EIA points to the same underlying cause: a resurgence in coal burn, driven in significant part by the buildout of large-scale data centers.

Why data centers, not steel mills

The intuitive assumption is that industrial demand — manufacturing, steel, chemicals — is what's straining the grid. It isn't. Data centers now account for roughly 4.4% of total US electricity consumption, and unlike most industrial load, they run continuously at high, predictable draw, regardless of whether wind and solar are producing. That profile makes them a poor match for intermittent renewables and a good match for whatever baseload capacity is sitting on the grid already — which, in large swaths of PJM territory (the mid-Atlantic and Midwest grid operator covering the largest concentration of US data centers), still means coal. Virginia, home to the world's largest concentration of data centers, has seen utilities nearly double their coal generation to keep up with demand from the state's data center clusters. Nearby coal states — Pennsylvania, West Virginia — have followed the same pattern.

The retirements that didn't happen

Before the AI buildout accelerated, the US coal fleet was on a clear glide path to retirement. That path has stalled. Utilities have postponed the announced retirement of at least 15 coal-fired plants, which together emitted almost 65 million metric tonnes of greenhouse gases in 2023 alone. Only 2.6 gigawatts of an anticipated 8.0 gigawatts of coal capacity actually came offline in 2025 — the smallest amount of coal retirements in 15 years. Behind that number is a blunt policy instrument: the Department of Energy has invoked emergency authority more than 40 times since May 2025 — 43 orders and counting — directing specific coal units to keep running past their planned shutdown dates, citing grid reliability and price stability. Several of those orders are now being challenged in court by state governments who argue the DOE has exceeded its authority, but for now, the plants stay open.

What's landing on the bill

None of this capacity is free, and the cost is showing up fastest in PJM's capacity auction — the mechanism by which the grid operator pays generators to guarantee they'll be available when needed. The auction for the delivery year that began this June cleared at $329.17 per megawatt-day, up from $28.92 just two auction cycles earlier and roughly 10 times the price set in the 2022 auction. PJM's independent market monitor, Monitoring Analytics, has gone further, estimating that data-center load growth is responsible for $29.4 billion of the $63.6 billion in total capacity charges across the last four auctions — a 46% share. That attribution is the market monitor's methodology and hasn't been independently verified or endorsed by PJM, but even PJM's more conservative figures put data centers at 38–40% of costs in individual auctions. Those capacity charges flow through to every ratepayer on the PJM grid, data-center operator or not, which is why data-center-driven price increases have become a live political issue across the region.

The blind spot in enterprise AI contracts

Here's what most enterprise AI buyers miss: a “cloud AI transformation” has a physical footprint that never appears in a vendor pitch deck. Every inference call is served from a physical data center, drawing from a physical grid, where the marginal unit of generation—the plant that spins up to meet the next increment of demand—is increasingly a coal plant scheduled to retire. None of this shows up in a corporate ESG report. None of it appears on a cloud vendor's sustainability page, which typically speaks in terms of renewable energy credits and long-term power purchase agreements rather than the marginal grid mix actually serving a given workload at a given hour.

Where iQ-LOOP fits

This is the problem we spend our time on at iQ-LOOP, and our answer is deliberately different from better accounting. Instead of measuring the grid's carbon intensity more precisely, we're building a way to sidestep the marginal-coal problem entirely. Our platform stores renewable electricity in aluminium — a stable, transportable solid with an exceptionally high energy density — and releases that stored energy on demand, at the point of use, as power, heat, hydrogen, or cooling. Because the energy is discharged from stored metal rather than drawn live off the grid, a facility running on our system isn't exposed to whatever happens to be marginal on the grid that hour, and it isn't a line item in a capacity auction that just cleared at $329 a megawatt-day.

The fit with data centers is closer than it might first appear. Cooling typically accounts for 30–40% of a data center's total power draw, and dispatchable cooling — delivered from the same stored aluminium as power and heat — is one of the configurations we are actively developing. We're validating the platform at 1 MW scale through a pilot in Ottawa before making any claims about hyperscale deployment, so this is a direction we're building toward, not a finished product being sold today. But the direction is the point: the PJM crisis described above is precisely the failure mode that a behind-the-meter, non-grid-dependent alternative is designed to avoid, and several of the US states we consider priority markets — Pennsylvania, Ohio, Michigan — sit inside or adjacent to the grid region absorbing the brunt of data-center-driven capacity costs.

There's a governance angle here too, not just an economic one. If the real question for a CIO signing an AI contract is “what's actually generating the electricity behind my tokens,” a system that stores renewable power and discharges it on-site, decoupled from the grid's marginal generator, is a more direct answer than a renewable energy credit purchased somewhere else on the grid.

The question for every CIO

Every enterprise signing a significant AI contract this quarter is, whether they realize it or not, making a decision about the energy behind that contract. The question worth asking before signing is simple: do you know what's actually generating the electricity behind your tokens—and would your board sign off if they saw it?

Sources

1. EIA — CO2 emissions from US power sector rose 4% last year

2. Grist — Data centers are slowing America's shift away from coal

3. Bloomberg — Data Centers and Coal Helped Drive Up US Emissions in 2025

4. EIA — US coal-fired generating capacity retired in 2025 was the least in 15 years

5. POWER Magazine — DOE's Section 202(c) Emergency Orders Since May 2025: 43 and Counting

6. IEEFA — Projected data center growth spurs PJM capacity prices by factor of 10

7. ZeroHedge / Monitoring Analytics — Data Centers Were Responsible for 46% of PJM's Last Four Capacity Auction Costs

8. Utility Dive — Data centers drove $6.3B in PJM capacity auction costs: market monitor

9. Congress.gov CRS — PJM's Electric Capacity Market: Background and Current Issues