Converting Bitcoin Mining Sites to AI Data Centers: Power Advantage and Gaps

Bitcoin mining built a large fleet of power-hungry sites across the United States. EIA’s preliminary 2024 estimate put cryptocurrency mining at roughly 0.6% to 2.3% of U.S. electricity consumption, and noted that miners in volatile wholesale markets cut their load when prices spike.¹⁰ Many of those sites were developed in Texas and other low-cost power regions, with grid connections of tens to hundreds of megawatts.

Two forces pushed operators toward AI. Mining revenue per unit of computing fell after the 2024 halving cut the block reward from 6.25 to 3.125 bitcoin.¹¹ At the same time, AI developers ran into a power bottleneck. In a May 2026 note, Bernstein analysts estimated that securing a gigawatt of power can take close to 50 months across U.S. markets, called power access rather than chips the main constraint on AI data centers, and described miners’ grid-connected sites as “warm powered shells” ready for GPUs.² VanEck argued in 2024 that suitable mining sites could energize GPUs in under a year, against four or more years for greenfield AI developments.¹

The resulting deals have been large. IREN signed a five-year, $9.7 billion GPU cloud contract with Microsoft, and Cipher Mining announced a roughly $5.5 billion, 15-year lease with Amazon Web Services.¹¹ Our guide to AI data center site requirements explains why power availability now leads the AI siting order.

Core Scientific’s agreements with CoreWeave are the most closely watched example. In June 2024 the companies announced about 200 MW of infrastructure to host CoreWeave’s high-performance computing, with more than $3.5 billion of projected revenue over initial 12-year terms and about $290 million a year on average. About $300 million of Core Scientific’s capital costs was to be credited against hosting payments at no more than 50% of monthly fees.⁵

CoreWeave then exercised further options, bringing contracted capacity to about 500 MW of critical IT load at six sites and $8.7 billion of potential revenue by October 2024.¹² A February 2025 expansion at Denton, Texas, added 70 MW, taking Denton to just over 260 MW and the total to about 590 MW across six sites, with $10.2 billion of projected revenue.³

The case also shows the limits. CoreWeave’s roughly $9 billion all-stock bid to acquire Core Scientific failed when Core Scientific shareholders voted it down on October 30, 2025, leaving the hosting contracts in place.¹³ Delivery has been uneven: a January 2026 report said CoreWeave had brought more than 16,000 GPUs online at Denton for OpenAI after construction delays tied to weather and a third-party developer.¹⁴ Converting a mine is construction, with construction risk.

A mining site’s main value is an existing interconnection agreement and, often, a substation already built and energized. That can save years compared with a new large-load interconnection. Galaxy’s Helios site in West Texas illustrates the scale: Galaxy said CoreWeave had committed to 526 MW across three phases under 15-year leases, backed by 800 MW of approved gross power, and it delivered the first 133 MW of critical IT load in 2026.⁶

Three cautions apply. First, gross power is not critical IT load. A mining site can put nearly all its grid power into hashing, while an AI data center loses a share to cooling and electrical losses; Cipher’s Barber Lake deal, for example, delivers 168 MW of critical IT from up to 244 MW of gross capacity.⁹ Second, expansion power still runs through the grid operator. In September 2026 Galaxy reported that ERCOT’s Batch Zero large-load review had conditionally classified Helios I (800 MW) and Helios II (830 MW) as base load, while a 1,000 MW Helios III request was classified as studied load, which remains subject to capacity allocation.¹⁵ See Texas Senate Bill 6 and ERCOT site selection.

Third, an interconnection agreement says little about the electrical path to the rack. Creative Strategies notes that mining sites often distribute at 208–240 volts, while AI systems need 480 volts or more, which means new transformers, switchgear and protection with long equipment lead times.⁴ Our guide to UPS and electrical distribution explains what that path includes.

Mining sites were designed to switch off. That flexibility is a feature for a miner: in August 2023, Riot Platforms earned $31.7 million in power and demand response credits in Texas while cutting its load by more than 95% during peak periods.⁷ An AI tenant paying for capacity expects the opposite. Few mining sites have the redundant feeds, uninterruptible power or generator fleets that a Tier III-style facility uses, and Creative Strategies cites industry practice of sizing AI backup generation at 1.2 to 2.0 times critical IT load.⁴

Cooling is often the largest change. Air-cooled mining halls are rarely built for GPU racks, and Creative Strategies concludes that retrofitting a mining shell for AI cooling can approach the complexity and cost of a greenfield build.⁴ Bitfarms’ plan for its 18 MW Washington mining site illustrates the target: liquid-cooled infrastructure for up to 190 kW per rack, with completion targeted for December 2026.⁸ In practice, many operators keep the substation and land and build new halls beside or in place of the old ones. See air vs. liquid cooling and retrofit or build new.

Fiber is the other gap. Mining needs very little bandwidth, so remote sites may have a single modest connection. AI training tolerates distance better than interactive services, but tenants still expect diverse, high-capacity routes; see fiber connectivity for data center sites.

Most conversions are long-term hosting or lease contracts in which the former miner provides land, power and the building, and an AI cloud company or hyperscaler installs the computing. Core Scientific’s contracts run 12 years,³ Galaxy’s Helios leases 15 years,⁶ and Applied Digital’s three CoreWeave leases at Polaris Forge 1 in Ellendale, North Dakota, total 400 MW of critical IT with about $11 billion of anticipated revenue.¹⁶

Credit support has become part of the structure. Google agreed to backstop $1.4 billion of Fluidstack’s lease obligations at Cipher’s Barber Lake site to support project debt, receiving warrants for about 5.4% of Cipher.⁹ Lenders and tenants underwrite the site’s power and construction plan much as they would a greenfield campus. Our guide to how data center deals are structured covers the main forms.

The screen starts with the power position and works inward to the building. A conversion makes sense when the interconnection is firm and the gaps can be closed faster than a greenfield site could be powered.

1. 01Confirm the interconnection terms: firm or interruptible, contracted MW, and any curtailment or demand response commitments. See firm vs. interruptible power.
2. 02Translate gross MW to deliverable critical IT load after cooling and losses.⁹
3. 03Inventory the substation, transformers and switchgear and price the upgrade to AI-grade distribution.⁴
4. 04Decide whether the existing halls can take liquid cooling or should be replaced.
5. 05Check fiber routes, water for cooling, and whether local zoning treats an AI data center differently from a mining operation.

A mining site with a firm, energized connection can be one of the fastest paths to AI capacity, but the parcel and building still need a full data center review. If you hold or are evaluating one, we can get a site reviewed for conversion potential.