The Convergence of Power Markets, AI, and Bitcoin Mining

In September 2024, Microsoft signed a 20-year power purchase agreement to restart Three Mile Island Unit 1 — a reactor that had been shut down five years earlier because it couldn't compete economically in the wholesale market. Amazon bought a datacenter campus wired directly into Talen's Susquehanna nuclear plant and has since expanded the arrangement toward nearly two gigawatts. And CoreWeave, an AI cloud provider, agreed to acquire Core Scientific — a Bitcoin miner — in an all-stock deal valued around $9 billion, in large part to acquire its power-dense facilities and grid interconnections.

Three different transactions, one message: electricity has become a direct input to value creation, and everything that generates, moves, or consumes it is being repriced. A reactor that was uneconomic selling into the grid at wholesale is worth restarting when a hyperscaler will underwrite twenty years of demand at a premium. That spread — between what the grid pays and what compute pays — is the whole thesis.

Two new demand curves on the same generation base

Energy to intelligence. Large-scale AI training and inference require enormous, sustained power loads, and every major hyperscaler is now competing for dedicated generation — not grid access, but generation itself. The Microsoft and Amazon deals above are the visible tip; behind them sits a queue of announced datacenter campuses whose combined power requirements are measured in gigawatts per site.

Energy to digital value. Bitcoin mining converts electricity into a globally liquid asset with no counterparty. Mining economics are simple at the surface — acquire cheap power, convert it to BTC, sell or hold — and the marginal buyer of mining infrastructure is increasingly not another miner. Core Scientific's sale to CoreWeave, and TeraWulf's pivot to hosting AI workloads backstopped by Google, mark the convergence point: the facilities, cooling, and interconnections built for mining are fungible with AI hosting, and they trade at AI-hosting valuations now.

Both demand curves share the same stack of bottlenecks — fuel sourcing, generation capacity, transmission, and hardware availability — and each constrains throughput independently.

The optionality stack for generators

Power generators historically sold into a single market: the wholesale grid. Margins were thin, pricing was regulated or commoditized, demand was predictable.

A generator with flexible capacity now has four routes for the same megawatt-hour:

Sell to the grid at prevailing wholesale rates.
Contract directly with AI datacenters at premium, long-duration PPAs.
Mine Bitcoin during periods of low grid demand or curtailment, monetizing otherwise wasted generation.
Arbitrage between all three on real-time pricing signals.

That is genuine optionality, and it changes what a generator is: less a commodity producer, more an arbitrageur of its own output.

The optionality only compounds when the infrastructure between generation and consumption actually works. The geographic mismatch between good generation sites and datacenter locations, renewable intermittency, and the transmission capacity required for gigawatt-scale clusters all create friction. Generators with co-located or directly interconnected load — the Susquehanna model — hold a structural advantage because they bypass the transmission bottleneck entirely.

Why nuclear wins the density argument

Not all generation is equal for this purpose. Solar requires large land footprints relative to output and needs storage to firm it. Fuel cells work as backup but struggle at the hundreds-of-megawatts scale AI campuses demand. The U.S. nuclear fleet, by contrast, runs at capacity factors above 90% and delivers consistent baseload on a fraction of the acreage — which is exactly the demand profile of a training cluster that runs at full load around the clock.

The supply side is being rebuilt in real time. Gen IV developers like Oklo and NuScale are moving through licensing, SMR entrants keep joining the pipeline, and subsidy and permitting decisions are actively shaping which technologies scale first. Policy is a load-bearing input in this thesis: shifts in regulatory posture reprice the generation layer faster than any change in the underlying engineering.

The uranium bottleneck

If the highest-conviction exposure to this convergence isn't in tech or crypto equities, it may be in the fuel itself. Spot uranium went from roughly $30 per pound in 2020 to over $100 in early 2024 as reactor demand, SMR announcements, and supply discipline collided; it has settled back but remains multiples of the levels that prevailed through the 2010s.

Two products matter: low-enriched uranium (LEU) for conventional reactors and high-assay low-enriched uranium (HALEU) for advanced Gen IV designs. They are not interchangeable — they feed different reactor architectures — and their supply chains are diverging while both face rising demand.

HALEU is the tighter constraint. Centrus Energy ($LEU) is the only NRC-licensed commercial HALEU producer in the United States, which amounts to a near-monopoly on the domestic enrichment step. Further upstream, Energy Fuels ($UUUU) holds an existing foothold in domestic uranium mining — tighter correlation to spot prices, direct capture of the supply-demand imbalance — and is pursuing expansion into enrichment. If that executes, it becomes a vertically integrated player spanning extraction through enrichment, a rare combination in a supply chain this thin.

The broader point: the uranium supply chain is geopolitically fragile, domestically concentrated, and facing a demand curve that bends upward from here. Companies building optionality across multiple nodes of that chain accumulate the most asymmetric exposure.

The hardware layer

Between energy and intelligence sits a physical supply chain: lithium, nickel, copper, and silicon flowing through fabrication into GPUs, ASICs, memory, and networking gear before populating the facilities where energy becomes computation.

This is where mining and AI infrastructure converge most visibly, because they share upstream dependencies — power-dense facilities, specialized silicon, industrial cooling. The convergence is operational, not theoretical: the same facility architecture serves both, which is why miners keep becoming AI hosts when the relative economics flip.

The hardware chain also has its own bottlenecks — fab capacity, rare-earth sourcing, export controls — that bind independently of energy. A generator with abundant cheap power still can't serve AI demand if the GPUs aren't available. The bottlenecks are multiplicative rather than additive, which is what makes the whole pipeline fragile and each chokepoint valuable.

What would weaken this thesis

The honest version of this argument names its own failure modes. The premium generators earn from compute demand compresses if datacenter buildouts slow, if announced campuses are cancelled, or if efficiency gains in training reduce power draw per unit of capability faster than demand grows. The nuclear leg weakens if SMR timelines keep slipping past their announced dates, or if permitting reform stalls. The uranium leg weakens if Kazakh and Canadian supply responds faster than expected, or if HALEU demand disappoints because Gen IV deployments do. Any of these is observable — announced deals, licensing milestones, spot prices — so the thesis can be tracked rather than believed.

The takeaway

The energy market is undergoing a structural repricing, and it isn't confined to generators. It propagates through a pipeline that converts fuel into intelligence and digital value, then recirculates capital back through the system — hyperscaler PPAs funding reactor restarts, mined assets collateralizing further buildout, policy reshaping each layer as it goes.

Energy is no longer just an input cost. It's becoming the medium through which intelligence and value get created, and the firms positioned across that pipeline — fuel, generation, interconnection, hardware — rather than at a single node are the ones accumulating durable exposure to the shift.

Sources: company announcements and public filings — Microsoft/Constellation PPA (Sept 2024), Amazon/Talen Susquehanna agreements, CoreWeave/Core Scientific acquisition (2025), NRC licensing records, spot uranium price history. This is not financial advice; these are observations on structural trends, not recommendations to buy or sell any security.