By end of 2027, three more 250+ MW AI power deals will be struck on repurposed heavy-industry or ex-mining sites, as memory and power become the binding constraint

The prediction

In July 2026 the scarcity in artificial intelligence visibly moved from the model to the substrate. SK Hynix raised about 26.5 billion dollars in the largest US listing ever by a foreign company, on the strength of the high-bandwidth memory inside every AI accelerator, and Anthropic signed a 19-billion-dollar, 20-year lease with the former bitcoin miner TeraWulf for a 401-megawatt data center built on a decommissioned Kentucky aluminum smelter. My claim is that these are not one-off deals but the opening of a durable pattern: AI growth is now gated by memory and megawatts, and the industry will keep securing both through long-dated, physically-anchored deals.

By December 31, 2027, at least three additional data-center power deals of 250 megawatts or more each — new leases, power-purchase agreements, or site acquisitions — will be publicly announced by major AI labs or hyperscalers in which the capacity is sited on a repurposed heavy-industry site (such as a former aluminum smelter, steel mill, or coal plant) or a converted crypto-mining operation. These are counted beyond the Anthropic and TeraWulf Justified Data lease and Marathon's July 2026 two-gigawatt Texas acquisition, which serve as the baseline.

Qualifying deals must be newly announced within the window (July 12, 2026 through December 31, 2027), must be 250 megawatts or larger, must involve a material AI-sector buyer (a frontier lab or a hyperscaler building for AI), and the capacity must sit on a repurposed industrial or ex-mining site rather than a greenfield build. Ordinary greenfield data-center announcements and the named baseline deals do not count toward the three.

Why 80 percent confidence

The structural driver is unusually strong. Power has become the hard ceiling on AI growth, and new grid capacity is slow and politically contested to build, so the fastest way to acquire hundreds of megawatts is to take over a site that already has the electrical envelope — which describes exactly the decommissioned smelters, retired plants, and crypto-mining operations now being converted. The bitcoin miners bring a ready-made skill set (dense, power-hungry compute sited next to cheap electricity, run around the clock) and existing power contracts, and the capital markets rerated them the moment the Anthropic lease landed, which pulls more of them toward the same pivot. The demand side is not slowing: frontier labs are raising tens of billions and racing to secure multi-year capacity ahead of rivals. When the incentive, the supply of convenient sites, and the capital are all aligned and pointing the same direction, three more deals over eighteen months is a low bar.

Confidence is held at 80, not higher, for two reasons. First, definitional risk: some large deals will be greenfield or will not disclose the site history clearly enough to confirm the repurposed-industry criterion, which could suppress the countable total even as the underlying behavior continues. Second, a sharp slowdown in AI capex — triggered by a funding pullback, a demand disappointment, or a credit shock — could pause the largest deals within the window. The direction is clear; the uncertainty is whether three deals clearly meet the size and site-history criteria in public disclosures before the window closes.

What would prove this right

Three or more qualifying announcements, on or before December 31, 2027, each 250 megawatts or larger, each from a material AI buyer, and each sited on a repurposed heavy-industry or converted crypto-mining location — for example, another lab leasing a converted smelter or steel-mill campus, a hyperscaler acquiring a miner's power portfolio for AI, or a neocloud building AI capacity on a retired power-plant site under a long-term lab lease.

What would prove this wrong

The window closes with fewer than three additional qualifying deals. This would most likely happen if AI infrastructure capex stalls, if the largest new capacity is built greenfield rather than on repurposed sites, or if deals are struck but their site histories are not disclosed clearly enough to confirm the criterion — leaving the July 2026 cluster as an early burst rather than a sustained pattern.