The lever snapped somewhere between the PJM 2025/2026 capacity auction and a hyperscaler's quarterly earnings call. Actually, the lever had been broken for years — we just refused to look. The narrative of the past decade was a tidy one: coal dies, renewables rise, the grid transitions through orderly retirement schedules and predictable infrastructure spending. Clean. Linear. Almost Victorian in its faith in progress. Then artificial intelligence showed up with a demand curve that doesn't read obituaries.
Here's the anomaly nobody modeled: coal plants scheduled for decommissioning are receiving economic "retirement exemptions" worth more than their salvage value. Capacity prices in PJM, the largest US grid operator, cleared near $270 per megawatt-day — several multiples of their historic range. And underneath the headlines about clean energy, a bidding war is raging for a resource we assumed would never again be scarce: dispatchable electrons.
When the lever breaks, the story begins.
I spent 2020 building a Python scraper that pulled 1.5 million Uniswap swap logs in three weeks, hunting for the pulse of DeFi Summer. That experiment taught me a lesson I've carried through the Terra collapse, the ETF approval cycle, and now the AI-energy convergence: sentiment moves faster than price, and infrastructure moves slower than both. The American grid is the infrastructure. AI is the sentiment. And coal is what happens when the two collide.
Coal's Second Act Requires No New Believers
Let's establish the baseline, because context is everything. Coal generated roughly 45% of US electricity in 2010. By 2023, that share had eroded to about 16%. Installed coal capacity fell from roughly 300 gigawatts to around 180. The decline narrative was written, ratified, and monetized by every ESG fund on earth. Then AI data centers arrived with a load profile that borders on pathological: 24/7 baseload demand, annual growth between 15% and 25%, and extreme geographic concentration in places like Northern Virginia's Data Center Alley.
The problem is structural, and it maps to three decaying pillars. Baseload capacity has been shrinking faster than replacement capacity can be built. Dispatchable reserves are thin: gas generates roughly 43% of US power today, but a combustion turbine ordered now takes three to four years to deliver, and grid interconnection queues in many regions stretch past 2029. And renewables, wind and solar together around 18% of generation, cannot sign the 99.999% uptime contracts hyperscalers demand. The reserve margin — the industry's cushion against peak demand — has been quietly draining for a decade.
In the short window of one to three years, the only generation sources that can physically respond are the ones that already exist. That means old coal units running past their retirement dates, gas peakers pushed to maximum duty cycles, and an uncomfortable number of diesel generators in data center parking lots. Coal's renewed demand isn't ideological. It's the exhausted grid reaching for whatever is already spinning.
The policy machinery compounds the effect. The Inflation Reduction Act promised hundreds of billions in clean energy credits and an 80% clean grid by 2030. The EPA's 2024 power plant rule demanded carbon capture retrofits or retirement for coal by 2032. Then the AI imperative landed on top, and the priorities began to cannibalize each other. States citing "reliability needs" are delaying coal retirements. Utilities like Duke and FirstEnergy have openly revised shutdown schedules. The EPA's reliability exception — designed for genuine emergencies — is becoming a standing legal backdoor. In the struggle between climate goals and AI competitiveness, regulatory silence is the loudest lobbyist.
The crypto mining angle sharpens the picture further. Miners and AI data centers are the twin loads chasing the same substations, and the competition between them — the dual compute race — is already lifting regional electricity prices. But the two loads react to scarcity differently. Miners relocate or curtail; hyperscalers write checks and appeal to governors. That asymmetry is the canary in the coal mine, and it's singing.
This is where most market commentary stops, satisfied with a clever "dirty energy is back" take. But the forensic work is just starting. The question worth answering isn't why coal is resurging. It's what mechanism allowed a structurally declining asset class to capture a bidding war — and what breaks when that mechanism reverses.
The Auction Mechanism No One Walks Through
The "bidding war" in the original report isn't a metaphor. It's a description of how American wholesale electricity markets price scarcity. Most regional transmission organizations — PJM, MISO, ISO-New England — run layered markets: energy for megawatt-hours delivered, capacity for the promise to be available on peak days, and ancillary services for the grid's balancing needs. For a decade, capacity prices hovered near zero because the system was structurally oversupplied. That's what killed coal economically, not politics: a market signal that said "your electrons aren't needed on demand."
AI flipped that signal. When the PJM auction cleared near $270/MW-day, it told every generator owner in the region: availability is now the most valuable thing you can sell. The capacity market just re-priced uncertainty, and the beneficiaries are any dispatchable assets with a pulse, regardless of fuel. That's why a 40-year-old coal plant with a scrubber retrofit is suddenly a strategic asset. It's not because coal became clean. It's because "available" became scarce.
The price signal is generating three consequences simultaneously. First, existing coal, nuclear, and gas assets switched from "losing money, schedule retirement" to "profitable, extend life." Second, new gas plant negotiations exploded, but those plants take four to five years to reach commercial operation — a time scale that does nothing for an AI cluster demanding power in 18 months. Third, the hyperscalers themselves started bypassing utilities entirely, signing direct PPAs with merchant generators at premiums that look aggressive even by crypto standards. Microsoft's deal to resurrect Three Mile Island's nuclear unit with Constellation is the clearest example.
The merchant generators — the unregulated ones — are the pure plays on this scarcity. Regulated utilities, by contrast, must ask state commissions for permission and settle for a capped return. The difference in strategic maneuverability is the difference between a cheetah and a draft horse.
This is also where crypto-native watching gives me an edge — or, as my 2022 Terra post-mortem taught me, the gap between narrative and mechanism. When Terra died, the narrative was "algorithmic money failed." The mechanism was a death spiral of collateral and confidence. Same shape here: the narrative is "AI wants coal." The mechanism is a capacity auction that fails to differentiate between a plant that can run for 30 years and a plant that can just make it to 2032. When the lever breaks, the story begins. But the auction already knew the lever was cracked.
The Elasticity Mirage
Here's the detail the mainstream coverage keeps omitting. Even if demand rose another 20% tomorrow, American coal supply cannot respond. Production peaked at around 1.2 billion tons in 2008; by 2023 it had collapsed to roughly 580 million tons. Mines have closed, and they don't reopen quickly. Rebuilding a mine takes 12 to 18 months minimum, and that's before we talk about rail logistics — BNSF and Union Pacific shrank their coal-hauling capacity aggressively as the industry declined. The workforce rebuilt itself into other industries. The equipment supply chains withered. Coal extraction is now a memory in most of Appalachia, and memories don't dig.
The cruel operational detail: active US producers like Peabody and Arch prefer selling metallurgical coal into Asian and European export markets, where steelmaking demand and freight arbitrage pay premiums domestic utilities cannot match. The leftovers — the low-grade production that costs the most to mine — land on unit trains for domestic power plants at prices that no longer make sense. So we face the possibility of a domestic coal shortage precisely at the moment of AI-driven demand. The price response will be ugly, and it will bleed directly into electricity bills.
The supply chain has contracted, and contraction is not the same as pausing. Restarting is more expensive than doubling down ever was. Based on my experience tracing token supply shocks through on-chain data, I've learned the most dangerous variable isn't demand — demand is visible, it's the headline. The dangerous variable is the shape of the supply curve, because that's where shocks hide.
The gas story is equally strained. New gas plants get contracted and look excellent on paper, then take four to five years to build. Small modular reactors, the modular nuclear savior, won't materialize at commercial scale until the 2030s at best. And renewables, for all their deflation, cannot seal a bankable 24/7 contract without storage attached. In the gap between time-to-market and demand arrival sits the entire coal comeback.
Storage: The Silent Variable That Wants This Job
Now the contrarian thread, and honestly, the one that deserves more words than coal will get. The bidding war is for dispatchability, not for fuel. And dispatchability just got dramatically cheaper.
Lithium carbonate collapsed from roughly $80,000 per ton in 2022 to the $10,000–15,000 range in 2024–2025, and battery system prices followed. Lawrence Berkeley National Laboratory's interconnection queue data shows storage now represents over half of all pending grid-connection requests, measured in terawatt-hours. Tesla and Fluence have signed data-center backup contracts that would've been unthinkable when I was watching DeFi pools in 2020. FERC Order 2328, finalized in 2024, opened capacity markets to storage participation just as capacity prices exploded. That timing is not an accident.
Every "coal is back" article buries storage in a footnote, if it appears at all. The pulse didn't wait for the forecast. Storage costs cut themselves in half while analysts were still debating whether batteries could replace baseload. The subsequent shift in the data center power stack — from diesel generators plus UPS to storage-plus-interconnection — is the quiet rewrite of the energy playbook.
The deeper insight is about who becomes storage's best customer. AI data centers face millions of dollars per minute in downtime costs. They will pay absurd premiums for availability. That willingness has shifted the storage revenue model away from pure arbitrage — buying low, selling high — and toward something closer to capacity insurance. The data center is becoming the anchor tenant for a future grid built around energy insurance, not energy throughput. In that worldview, coal isn't a competitor. Coal is the emergency brake on a vehicle that's about to get airbags.
The New Pecking Order
Rank the generation assets by who wins this bidding war, and the order is coldly clear: existing nuclear is king, existing coal is the surprisingly resilient middle class, and brand-new large gas plants are the overfunded rookies with a terrible salary cap. Constellation Energy's Three Mile Island restart for Microsoft is the archetype: a zombie facility revived not by subsidies, but by the reliability premium a hyperscaler will happily pay. Talen, Vistra, and the other merchant generators have seen their fleet valuations repriced upward.
Nuclear is the long-term answer, and every sensible player knows it — Microsoft, Google, and Amazon have all signed nuclear PPAs, including SMR offtake agreements for the 2030s. But the 2030s are an eternity in the AI development cycle. So the winners of the next three years are existing dispatchable megawatts, whatever their source. And the losers are the loads with less money and comparable desperation.
The Crypto Miner Axe
And here's where the original Crypto Briefing angle shines, because the crypto industry is watching this conflict from the front row. Crypto miners and AI data centers want the same substations, the same electrons, the same long-term PPAs — but they occupy opposite ends of the flexibility spectrum. A Bitcoin miner can curtail in seconds; an entire mining business can relocate on the back of a semi-trailer. A GPU cluster cannot. Downtime for a hyperscaler costs millions per minute, measured in model training runs and service-level agreements, not just hardware depreciation. That asymmetry converts miners into swing capacity: the first to be curtailed, the first to be shoved aside in every renegotiation.
Based on my 2025 work tracking AI-agent transactions on Render Network, decentralized compute is trying to thread the same needle: offer cheaper, flexible compute while hyperscalers hoard silicon and electrons. The same logic that makes miners flexible load also makes them the ideal grid response asset. Some miners are already pivoting their business models from "mine crypto" to "sell interruptibility." That transition deserves more respect. In the bidding war for electricity, the ones who can flex, survive.
But the political reality is unkind. Ratepayers in states like Georgia are being asked to pay for massive new gas plants to serve hyperscale load, and the backlash to "subsidizing AI billionaires at taxpayer expense" is already forming. Virginia has floated surcharges on data centers. Texas's ERCOT forces data centers into interruptible products under its coincident peak structure. The era of silent taxpayer subsidization is ending.
The Zombie Problem
Now the contrarian reckoning, because narrative discipline requires me to attack my own frame. Coal isn't "back" in any durable sense. What we're watching is corpse-reanimation economics: generous conditions, a compliant regulatory environment, and narrative urgency keeping a terminally ill asset class on life support. The coal-plus-CCS story is the perfect illustration.
EPA's 2024 power plant rule says coal plants can continue operating past 2032 only if they install carbon capture with at least 90% capture rates. The 45Q tax credit provides up to about $85 per ton of captured CO2. CCS deployment costs run $70 to $130 a ton depending on plant and geology. These numbers barely clear the hurdle — on paper. But engineering timelines do not cooperate. Retrofitting a 1970s-era coal plant with CCS takes years, and the queue of decision-makers facing compliance deadlines will not all make it. Some states will push the EPA's "reliability exemption" until it means whatever they want it to mean. If political winds shift further toward energy independence, enforcement evaporates entirely.
This is the classic narrative trap I documented after Terra: the market can believe a story that no longer fits the mechanism. Terra's "digital yen" positioning fell apart because the structure couldn't support the fiction. Coal's "essential infrastructure" positioning may survive on legal exceptions long after the engineering reasons vanish. But the economics are deteriorating from within — exactly the phase where smart capital exits before the retail narrative catches up.
There's a second blind spot, and it's the biggest one: the AI power demand forecast may simply be wrong. Chips have been getting more efficient at a steep curve; inference workloads migrate to specialized low-power silicon; the industry talks itself into a capacity overbuild at least as often as it underbuilds. If the 2030 demand projections turn out to be a 1.5x rather than a 3x, today's scarcity price inverts into an excess glut, and coal's extended leases look like expensive insurance nobody needed. The overbuild cycle is a feature of deregulated markets, not a bug. Witness the fiber optic bubble, the LNG contract wave, the crypto mining expansion of 2021.
When the Grid Already Knew
Falling through the floor to find the foundation. The foundation of this bidding war isn't coal, isn't nuclear, isn't even storage. It's the economic value of dispatchability — the ability to deliver electricity at a specific moment, no excuses. The capacity market is pricing that ability the way it hasn't in decades. Every technology that can sell dispatchability gets repriced:
- Coal with legs until 2032: repriced upward, temporarily.
- Storage with four-hour duration: repriced upward, structurally.
- Gas peakers with fast start: repriced upward, cyclically.
- Interruptible loads like crypto mines: repriced upward, eventually.
- Non-dispatchable generation: repriced downward, forever.
The narrative arc, mapped properly, says America's next grid is not a fight between clean and dirty. It's a fight between rigid supply and flexible demand, between 15-minute response and decade-long construction queues. The coal revival is a smoke signal — alarming, but revealing something more important than itself.
Mapping the chaos to find the hidden narrative arc: the story after coal's short encore will be about load flexibility. If AI compute keeps growing at hyperscale pace, the system will resolve around batteries, demand response, and interruptible load markets where computing power — including crypto mining — becomes a grid asset rather than a grid burden. The tokenization of interruptibility is, frankly, a Web3-native use case waiting for its first large-scale proof.
The Next Lever
The lever that breaks next won't be a coal plant's boiler. It will be the planning assumption that electricity demand stays linear in an exponential world. The narrative that survives this cycle will be the one that treats dispatchability as the scarce resource and builds the markets accordingly.
The question I keep asking as I watch the PJM charts and the hash price tickers: who profits when the grid's availability becomes the most expensive asset in the machine? Coal plant owners capture the short-term rent, but the terminal values accrue to the flexible. The bidding war's biggest winner is "any type of unrestricted power" — the market is temporarily indifferent to fuel, which is precisely why coal got a seat at the table. But capacity markets are lagging indicators. If the hyperscalers over-forecast their own demand — and the history of every infrastructure cycle suggests someone will — today's scarcity contracts become tomorrow's stranded assets. The smart money is watching the same charts, hedging both directions.
When the lever breaks, the story begins. But the next lever is already made of software, batteries, and interruptibility. The coal dust settles, and the foundation beneath is cleaner than the narrative feared. And when the next narrative forms around whatever replaces coal, I'll be watching the mempool of the grid — the sentiment data that always moves first.