Converting bitcoin mining sites into AI factories: what carries over and what has to be built

In the United States, and increasingly elsewhere, bitcoin mining sites are being re-purposed to host AI compute. The reasoning is simple. AI capacity is constrained by power, and a mining operator already holds what a new data center waits years for: an energised grid connection, a substation, land and permits. The engineering is less simple. A mining site and an AI factory draw similar megawatts, and there the resemblance largely ends. This article sets out what transfers from one to the other, what has to be added, and how to plan the conversion so that the capacity promised to a tenant is capacity that can be delivered.

Why mining sites attract AI capacity

Grid capacity is the scarce resource of this cycle. Interconnection queues run to years in many markets, and the equipment behind a connection is no quicker: a 2026 survey of manufacturer order books puts substation transformers at 75 to 110 weeks and generator step-up transformers above 50 MVA at 100 to 150 weeks or more (Terrapin Construction Group). A site with power already flowing starts with an advantage measured in years.

The market has acted on it. Core Scientific has contracted about 590 MW of critical IT load across six of its sites to CoreWeave, including a 70 MW expansion at Denton, Texas, announced in February 2025. Galaxy delivered the first 133 MW of critical IT at its Helios campus in West Texas to CoreWeave on 6 July 2026, on a site bought from a mining company in 2022, with 526 MW committed over three phases. Riot signed a lease with AMD in January 2026 for 25 MW of critical IT at Rockdale, Texas, delivered in phases between January and May 2026. TeraWulf and Cipher have leased hundreds of megawatts at their mining sites to Fluidstack under agreements that Google backstops, and Cipher signed a 300 MW, fifteen-year lease with Amazon Web Services in October 2025. IREN built a liquid-cooled AI building beside its mining load at Childress, Texas.

What a mining site is

A mining site is designed for one objective, the lowest cost per hash, and that objective shapes every system.

Power is single-path. Each block of miners typically sits behind one transformer and one low-voltage board. Mining load is interruptible by design, so UPS, batteries and generators are usually absent.
Load is flat and curtailable. Mining runs at near-constant power and can be switched off within seconds, which makes it a useful grid resource and an undemanding load.
Cooling is air or immersion. Air-cooled halls move outside air through the racks. Immersion sites submerge the miners in a dielectric fluid and reject the heat through dry coolers, with the fluid often returning at 35 to 45 °C.
Buildings are light. Floors, doors and lifts are sized for miners of a few tens of kilograms each, and fire detection, security and monitoring are basic.

A typical immersion line is one transformer, one low-voltage board or busway, a row of tank positions and one dry cooler with a single pump: a megawatt or two on one path, with no UPS, no generator, no coolant distribution unit and no spare pump anywhere in the chain. The drawing below shows what becomes of that line.

A mining line after conversion, UBIK reference drawing The grid supply, the transformer pair, the feeders, the line transformer and the dry cooler are reused after tests; the board, the pumps and the headers are retrofitted; rack power, UPS and generators, chillers, coolant distribution units, the secondary loop and the manifolds are new.

What an AI factory needs

An AI factory concentrates far more power per rack and asks more of every system around it.

Rack power of 100 kW and above. NVIDIA's facilities reference design for its current platform covers cabinets from 198 to 330 kW (NVIDIA DSX Facilities Infrastructure Design Guide v2.0, August 2026).
Direct liquid cooling to cold plates, through coolant distribution units and a technical secondary loop, with a smaller air-cooled share for networking and power supplies. NVIDIA designs the Vera Rubin platform for coolant at up to 45 °C (NVIDIA), which suits dry coolers.
Redundant power distribution, so that the loss of one transformer, board or busway leaves every rack energised.
A defined availability model for each class of load. Training clusters, inference platforms, storage and the network core carry different consequences of an outage, and the reference design treats them differently: the network core and the mechanical plant are on UPS and N+2 generators, the GPU compute load is not on UPS, and the GPU cooling plant rides through an outage on the thermal mass of its loop until the generators start.
Structure and logistics for racks weighing well over a tonne, delivered fully integrated.

The gap, system by system

Power distribution and redundancy

The transformers and medium-voltage feeders usually transfer, after condition testing. What changes is the topology. Single-path lines are regrouped into a distributed redundant scheme, for example four lines serving each room so that any three carry the full load, with every rack fed from all four. Redundancy uses capacity: lines regrouped in this way yield about three quarters of their rating as IT power, and the cooling plant takes its share of the gross power before that. Galaxy's first phase at Helios shows the arithmetic at site scale: 133 MW of critical IT from about 200 MW of gross power. A site sold at its grid number delivers noticeably less IT capacity once the plant and the redundancy are in place, and the commercial plan should start from the IT figure.

The same check applies upstream. A pair of grid transformers sized for the mining load holds N+1 only up to the rating of one of them, so a site that grows past that point needs a third transformer or a load-shedding scheme for the compute load, and the grid study that goes with either.

Backup power

Mining sites rarely carry UPS or generators. The conversion adds them, or the tenant accepts a different availability model. NVIDIA's reference design keeps UPS and N+2 generators for the network core and the mechanical plant and leaves the GPU compute load without UPS. Many enterprise and sovereign tenants still expect UPS and generators on all IT load, and the current platform reference designs for that case feed each rack from four independent trains, eight cords per rack, each train with its own transformer, generator and UPS, and give the coolant distribution units and fan walls the same ride-through as the GPUs. The choice moves the capital cost and the programme by large amounts, and it belongs in the tenant agreement, settled at the start.

Power block for liquid-cooled AI racks, UBIK reference drawing Four identical trains of transformer, generator, switchgear, UPS and power panel; any three carry the load, and every rack takes eight cords, two from each train. The mechanical chain has its own UPS, so the coolant distribution units and the fan walls ride through with the GPUs.

Cooling

Warm-water liquid cooling is where a mining site carries over best. Dry coolers rejecting heat at 35 to 45 °C sit inside the temperature window of current GPU platforms, and the existing pumps and pipework can often be reused after inspection, with standby pumps, ring headers and thermal storage added for continuity. The new elements are the coolant distribution units, the secondary loop with its filtration and leak detection, and a separate chilled-water plant for the air-cooled share. The governing check is the peak summer condition: the supply temperature on the hottest days against the platform's limit, with the design point taken from the extreme annual values of the climate file rather than the percentile figures used for comfort cooling.

The numbers run close to the limit. Dry-cooler water that leaves the coolers in the high thirties on the hottest days reaches the rack 2 to 4 K warmer after the approach of the coolant distribution unit, at or above 40 °C. That is inside a platform limit of 45 °C with a few kelvin to spare, and outside a platform limit of 40 °C on peak days, which then needs trim cooling for those hours or a different platform. The choice of platform is a cooling decision as much as a compute one, and it is made on the climate file, not on the typical day.

Cooling for a liquid-cooled AI room, UBIK reference drawing Dry coolers on a ring header with standby pumps, thermal storage and filtration; coolant distribution units N+1 between the facility water and the secondary loop that serves the racks; and a separate chilled-water loop with fan walls for the air-cooled share.

Structure and logistics

Floor ratings, door heights and lift capacities are frequently the binding constraint. Mining halls are often rated at 5 to 10 kN/m² where a current liquid-cooled hall needs 25 to 30 kN/m², and a logistics door built for miners can be lower than an integrated rack on its base. A structural survey comes before any layout. Where a hall cannot carry AI racks, the usual answers are steel support frames, ground-floor placement, or a new prefabricated block on the same energised site, which is the route IREN took at Childress.

The grid interface

AI training draws power in a pattern closer to a square wave than to a steady load: short peaks above rated power, a plateau, then near idle, repeating over milliseconds to tens of seconds (NVIDIA on power smoothing in GB300 NVL72). Grid operators have noticed, for a related reason. On 10 July 2024 a transmission fault in Northern Virginia caused about 1,500 MW of data center load to drop off the grid at once, across 60 load points, as sites transferred to their UPS and did not return for hours (NERC incident review, January 2025). NERC issued a Level 3 alert on large computational loads in May 2026 (NERC). Texas approved ride-through requirements for large electronic loads of 75 MW and above in July 2026, effective 1 August (ERCOT NOGRR 282), and Ireland's regulator approved a Grid Code modification in September 2026 that requires transmission-connected demand facilities to ride through transient faults (Arthur Cox).

AI training power profile, UBIK drawing, illustrative At the rack the load swings between a plateau, a short peak above rated power and an idle during the exchange of gradients. At the site boundary, storage in the racks and power caps flatten it.

For a converted site this makes the grid interface a design deliverable: UPS settings for ride-through and power smoothing, power caps at the racks, medium-voltage battery storage sized in a grid study, and an electromagnetic transient model of the facility for the grid operator. A mining load that could be switched off at will becomes a load that has to stay connected.

Cost and schedule

Published figures show why conversions attract capital, and why the figures need careful reading. Riot reported US$89.8 million for the retrofit of its first 25 MW at Rockdale, US$3.6 million per MW of critical IT (Riot, January 2026). IREN put Horizon 1, a liquid-cooled building of 75 MW gross and 50 MW of IT on its energised Childress site, at US$300 to 350 million, US$6 to 7 million per MW of IT (DCD). Turner & Townsend's 2026 index, which prices a typical 60 MW new build across 62 markets, has Tokyo at the top at US$18.28 per watt and reports costs in North America and Europe up 13% in a year (Turner & Townsend). A conversion that reuses the grid connection, the transformers, the heat rejection and the shell sits below a new build on a like-for-like basis, provided the comparison is made per megawatt of IT and with the same backup scope.

Two cautions apply. First, published conversion figures rarely state their backup scope. A fully redundant backup system, UPS and generators for all of the IT load, costs in the order of US$2 to 3 million per MW at 2026 rates (Datacentres.com Research, September 2026), close to the whole of Riot's retrofit figure. Second, the programme is set by long-lead electrical equipment. 2026 manufacturer slot data puts diesel generators of 1.5 to 3 MW at 50 to 78 weeks, larger sets at 90 to 110 weeks, and UPS systems at 30 to 60 weeks depending on size (Terrapin Construction Group). With equipment bought new, generator delivery sets the date of the first block, and a realistic programme lands first capacity in the second year from project start.

A method for conversions

In UBIK's feasibility work a conversion follows the same sequence as a new site, with the diagnostic in front.

A technical diagnostic first. Survey the power chain from grid to rack, the condition of transformers and switchgear, floor ratings, the performance of the heat rejection plant at the summer peak, water quality, and the grid operator's requirements for a large load. Every later decision depends on these facts.
Agree the availability model with the tenant for each class of load: compute, network core, storage and mechanical plant.
Fix the redundancy topology and state the capacity in MW of IT, room by room.
Plan a first block of one or two rooms that proves the design, the commissioning process and the operating procedures, then repeat it.
Reserve manufacturing slots for generators, UPS and switchgear at project start.
Build the operations organisation in parallel: procedures, staffing and monitoring for a mission-critical facility, which a mining operation rarely has in place.

A mining site offers an AI factory the hardest asset to obtain, energised grid capacity. Everything inside the fence, from the distribution topology to the continuity of cooling to the way the site behaves on the grid, needs engineering for a very different load. Owners who treat the conversion as a design project, starting from a diagnostic and an agreed availability model, deliver the capacity they promise. Those who start from the grid number alone discover the difference during commissioning.