Coal's AI Comeback Is a Time Arbitrage, Not a Transition Reversal

The data does not care about narratives. It never has. In the PJM 2025/2026 capacity auction, clearing prices reached roughly $270 per megawatt-day, breaking every record in the modern organized wholesale market era. Utility executives who spent five years announcing coal plant retirements quietly extended them. Independent power producers with dispatchable assets watched their equity curves invert from slow bleed to vertical climb. A bidding war for American power plants is underway, and the asset being bid on is not clean energy, carbon-neutral baseload, or strategic reserves. It is time. This is a blockchain publication, so let me frame it the way I frame everything I trade: the power market just experienced a short squeeze. AI data centers are the new leveraged longs. Existing coal and gas plants are the trapped shorts. The margin call arrived in the form of a capacity auction, and the energy transition narrative absorbed the loss. I have spent 21 years watching markets discover scarcity. In 2017, I manually audited early-stage smart contracts during the ICO boom and flagged reentrancy vulnerabilities that would have cost projects $4.2 million. In 2022, I spent three weeks tracing the Terra/Luna death spiral on-chain, tracking the exact block sequence where the algorithmic peg broke. The lesson from all of it is identical: trust is a technical variable, not a marketing claim. So when the headlines say "coal demand returns because of AI," my first instinct is not to ask whether coal is back. It is to ask what mechanism is paying for it, and what that mechanism will do when the imbalance fills. Context: AI Load Breaks the Grid's Planning Assumptions AI data centers present a load profile that breaks every assumption in US power planning. First, they require continuous baseload operation; training clusters do not ramp down at night. Second, their growth rate is steep: 15% to 25% annual load growth, versus the 1% to 2% linear growth utility planners have budgeted for decades. Third, they are geographically concentrated. Northern Virginia's Data Center Alley, Ohio, and Texas are single nodes consuming city-scale power, and the interconnection queue for new substation capacity in Virginia runs past seven years. The US grid has three structural deficiencies that this load profile exploits. The coal fleet has fallen from roughly 300 GW to about 180 GW over the past decade, with retirements outpacing replacement capacity. Natural gas is the dominant source at roughly 43% of generation, but a new combustion turbine ordered today takes three to four years to deliver. Wind and solar add about 18% of the mix, but neither can be contractually counted as dispatchable in a capacity agreement. Every one of these deficiencies is a latency problem. And latency is precisely what AI demand cannot tolerate. The result is a classic supply-demand mismatch with a distinctive pricing signature. In normal power markets, price reflects marginal fuel cost. In a capacity-scarce market, price reflects the cost of standby availability: the option value of having a megawatt ready to run when the load arrives. That option value is exploding across PJM, MISO, and ISO-NE right now. The Route Competition Is a Timing Competition The technical route competition for AI supply, coal versus gas versus nuclear versus renewables plus storage, is really a competition over deployment speed. Over the one-to-three-year window, the only assets that can respond are existing coal units running past their retirement dates and gas peakers pushed to full utilization. Coal's share of US generation fell from roughly 45% in 2010 to around 16% in 2023. The current revival is a reflection of the short-term supply curve, not a permanent re-rating of coal. The average US coal unit is over 40 years old; extending its life requires capital spending on environmental controls and, if the EPA has its way, carbon capture. The economics only work if the scarcity premium lasts more than a decade. I have never seen a scarcity premium last a decade. In the three-to-five-year window, the competition shifts to two candidates: nuclear, particularly small modular reactors, and hybrid configurations of renewables plus storage plus gas backup. The hyperscalers are voting with contracts. Microsoft's agreement with Constellation to restart Three Mile Island, and the shelf of SMR purchase agreements signed by Google, Amazon, and Microsoft, show where long-term capital believes the answer lies. Coal is the emergency bridge. Nuclear is the destination. Storage is the buffer that determines how long the bridge has to stay open. This ordering matters for asset allocation. If storage costs keep falling, the bridge shortens. If SMR timelines slip, and they have been slipping for fifteen years, the bridge extends. The market is pricing coal as if the bridge is permanent. The contract data says the bridge is temporary. Coal Supply Is Inelastic: The Revival Narrative Ignores the Funnel Here is the part the bidding-war headlines miss. Even if AI demand pulls coal off the retirement block, the US coal supply chain cannot respond in kind. Production has fallen from a 2008 peak of roughly 1.2 billion tons to about 580 million tons in 2023. High-cost mines are closed. The remaining capacity is concentrated among a handful of large producers, including Peabody and Arch, who spent years under ESG pressure deliberately halting new mine development. That is a strategic choice, not a passive decline. Logistics compounds the constraint. Coal moves on rail, and the rail network was rightsized for a shrinking industry. BNSF and Union Pacific are not holding spare capacity for an AI-driven coal renaissance. Skilled miners have left the industry; equipment supply chains have contracted; restarting a mine takes 12 to 18 months. Meanwhile, metallurgical coal, the high-margin export grade, is fetching premiums abroad. Rational producers will allocate output to the highest bidder. The domestic thermal coal market could be tighter than expected exactly when it matters most. Here is the feedback loop the narrative ignores. AI-driven demand meets inelastic coal supply. Coal prices rise. Fuel costs for coal plants rise. The cost advantage that revived coal versus gas in the first place erodes, and the system swings back toward natural gas, which is itself constrained by turbine delivery timelines. This is not a linear revival. It is a price chain reaction across three commodities. The code does not lie, only the audits do, and the audit of US coal supply was performed during a decade of managed decline. The spare capacity simply is not there. Storage Is the Silent Variable in the Entire Trade The original framing of this story, coal versus AI, omits the asset class that is quietly reshaping the economics: battery storage. Lithium carbonate prices have collapsed roughly 85% from their 2022 peaks. Global battery storage system prices fell to roughly $0.07 to $0.10 per watt-hour in 2024. The Lawrence Berkeley National Laboratory reported that storage constitutes more than half of the US interconnection queue, at terawatt-hour scale. Tesla and Fluence have secured multiple data center backup orders. FERC Order 2328 and related rulings are opening storage participation in capacity markets. The more important shift is the revenue model. Data centers are becoming the highest-quality off-takers for storage projects, not because storage earns arbitrage spreads in the traditional peak-valley sense, but because a data center pays an enormous premium for availability. The modeling framework is moving from energy arbitrage to capacity insurance. In 2020, when I ran a $1.5 million yield farming operation across Uniswap V2 and Curve, I learned that yield is compensation for a specific imbalance. Storage paired with a data center is capturing the imbalance between grid interconnection latency and AI load arrival. That is the same trade, in a different settlement currency. If storage costs keep falling, coal's role as the bridge shortens. The lithium macro cycle is a silent tailwind that the coal-revival narrative does not price. Every forecast that treats coal as the fixed response to AI demand is missing the fastest-moving input in the system. Capacity Markets Are the Smart Contract I want to be precise about the mechanism behind the bidding war. PJM, MISO, and ISO-NE operate three-layer markets: energy, capacity, and ancillary services. For a decade, capacity prices were depressed because the grid was structurally oversupplied. Investment in new generation collapsed. The AI demand surge hit an unprepared market, and capacity prices exploded. The price signal is producing three simultaneous responses. First, existing coal, nuclear, and gas assets are becoming economically attractive to extend rather than retire; the economics flip from unprofitable retirement to profitable life extension. Second, contract negotiations for new gas plants have surged, but those plants take four to five years from signing to commissioning; they solve none of the immediate scarcity. Third, hyperscalers are bypassing utilities entirely and directly investing in generation assets. Microsoft, Google, and Amazon have all signed direct purchase agreements with nuclear and geothermal developers. ExxonMobil is in discussions to supply gas-plus-CCS directly to data centers. The compute industry is becoming a power industry. Smart contracts execute logic, not intentions. The capacity market is a smart contract run by the system operator. It does not read the Inflation Reduction Act. It does not read EPA emission rules. It reads one variable: how many megawatts are available at the reliability deadline. That is why coal is back at the table. Not because coal is clean. Because the capacity auction has no ESG column. One caution from my trading background: capacity market signals are lagging and oscillating by design. They reflect scarcity at the moment of auction, not scarcity five years ahead. If AI demand forecasts are inflated, and if the exponential curve decays because chip efficiency improves faster than anticipated, today's panic prices will be tomorrow's overbuilding. The fiber-optic bubble of 2001 followed exactly this shape. The buildout was real. The investment thesis destroyed capital anyway. Policy Is Contradicting Itself, and the Contradiction Is the Arbitrage US federal policy currently operates three commitments that cannot all be true. The Inflation Reduction Act subsidizes clean energy deployment toward an 80% clean grid by 2030. The EPA's 2024 greenhouse gas rule requires existing coal plants to install carbon capture or retire by 2032. And the AI competitiveness agenda treats compute infrastructure as a national security priority. When commitments collide, priorities rank. The EPA rule contains a reliability exception. Grid operators can declare a reliability emergency and keep coal online. Under the AI demand narrative, that exception is becoming standard operating procedure. Duke Energy, FirstEnergy, and Ameren have already delayed coal retirements citing reliability needs. If enforcement weakens further, and the political environment in 2025 through 2028 is hostile to EPA enforcement, the 2032 timeline becomes a suggestion. The coal-plus-carbon-capture story is becoming a lobbying vehicle: "no coal, no AI" is the new "coal is irreplaceable." But carbon capture at $70 to $130 per ton of CO2, against an actual carbon price near zero, does not pencil out. Most extended coal units will not complete CCS retrofits by the deadline. The likely outcome is looser rules or trivial penalties. Hydrogen is the policy orphan in this narrative. Green hydrogen remains too expensive and lacks infrastructure to serve AI loads in the short window. Natural gas blending with hydrogen is a mid-term mitigation, and the DOE's regional hydrogen hub program is spending billions on demonstration projects. But none of this moves the needle inside the three-year scarcity window. These technologies matter for the decade-ahead picture. They do not matter for the auction that just cleared. The state-level picture is a subsidy war. Virginia, the data center capital, is hitting grid walls and imposing new fees on data center load. Texas offers fast ERCOT interconnection, but its grid has demonstrated fragility under extreme weather. Georgia has approved new gas turbines with costs socialized across the rate base. Iowa and Nebraska have abundant wind but are constrained by interconnection limits. Every state is bidding for AI dollars, and the bill will arrive with retail rate increases. When voters discover they are subsidizing data center power at market-beating terms, expect litigation, rate caps, and a political reset. The arbitrage window is real. Its duration is political. Who Wins the Positioning War The market competition story is about who owns existing dispatchable assets. Ranked by speed and risk, the winners are: existing nuclear plants, then existing coal, then fast-track gas peakers, then SMRs, then new combined-cycle gas plants. Independent power producers such as Constellation, Vistra, and Talen are the leverage plays on this ranking. Their stock performance in 2024 and 2025 reflected it. Regulated utilities such as Dominion, Georgia Power, and Duke are less flexible because their returns are capped by regulators. Intermittent renewable operators benefit from higher overall prices and PPA premiums, but face curtailment and queue constraints. There is a structural transformation underneath this competition. Power asset owners are shifting from a public utility identity to a technology infrastructure provider identity. Direct sales to data centers earn far more than merchant market sales. This bypasses traditional utility regulation and will draw scrutiny as it scales. The industry is being repurposed around a different customer class with a different risk tolerance and a different balance sheet. None of this is new to anyone who has watched compute infrastructure commoditize an underlying input. Crypto miners went through this exact cycle: energy access became the moat, power procurement became the strategy, and the winners were the operators who locked long-dated power at fixed prices before the demand arrived. The mining playbook is now being executed by Microsoft and Amazon, with better lawyers and worse public sympathy. Contrarian: The AI Demand Forecast Is the Weakest Link Every participant in this bidding war is modeling AI power demand as a monotonic exponential. Exponential projections are the most dangerous input in financial modeling. In 2022, I traced the Terra/Luna collapse block by block; the structural flaw was recursive token deposits requiring ever-growing inflows to sustain the peg. The AI-coal trade is not recursive in that sense, but it has a dependency of the same shape: it requires the AI demand narrative to keep growing. The bear case is underappreciated. AI chip efficiency gains, especially at the inference layer, could bend the power demand curve downward significantly. If efficiency advances faster than adoption growth, the AI power crisis narrative gets revised, and every long-dated power purchase agreement signed at panic prices becomes a liability. The capacity market's lagging and oscillating price signal makes this risk worse. What looks like a structural need today can become structural overcapacity in four years. Also worth noting: the players in this trade have aligned incentives to amplify the crisis narrative. Coal producers want policy protection and public sympathy; the "no coal, no AI" meme converts an extraction industry into a national security necessity. Data center developers want priority grid access. Utilities want rate base expansion. Every one of them benefits from a slightly exaggerated version of the story. The code does not lie, only the audits do. The market's audit of AI demand is still pending. Coal is pricing as if the audit will pass. Storage and nuclear are pricing as if the audit will be revised. They cannot both be correct. Takeaway The bidding war for American power plants is a trade on time, not on energy ideology. Coal's revival is an arbitrage on scarcity: a yield paid to dispatchable megawatts for simply existing while the grid catches up. It will be profitable for some. It will also be temporary. Watch three variables: battery system prices, nuclear PPA terms, and capacity market rule changes. Falling storage costs shorten coal's window faster than the auction curve implies. A downward revision to AI demand forecasts turns today's panic-priced PPAs into the next mark-to-market casualty. And the political backlash against subsidizing hyperscalers will eventually reshape capacity market design, just when the incumbents are least ready for it. I have spent my career watching markets pay extraordinary yields for extraordinary imbalances. The rule never changes: the imbalance fills, the yield decays, and the asset that priced in permanence gets revalued first. Coal's AI comeback is a beautiful yield trade. It is a terrible buy-and-hold conviction. The question for the reader is not whether coal is back. It is whether you know what signal will tell you when the bridge has closed.