Контрольный список по преобразованию майнинговой фермы в центр обработки данных для ИИ

A mining site can have 5 MW of power on paper and still fail an AI deployment review.

The problem is usually not the headline power number. It is the gap between utility capacity and usable AI capacity: transformer loading, power quality, rack-level distribution, liquid cooling, network connectivity, site security, commissioning procedures and long-term maintenance all have to work together.

The fastest path from mining to AI is not to reuse every existing asset. It is to identify which power, cooling, control and site systems are actually reusable.

This mining farm to AI data center conversion checklist explains how operators can evaluate an existing site before investing in a cooling retrofit or GPU infrastructure upgrade.

1. Start With A Reuse Audit

The first step is to separate existing assets into three categories:

A mining container may have been optimized for ASIC density, airflow and simple power distribution. An AI data center requires a different balance between electrical stability, cooling topology, network performance and serviceability.

The existing site should therefore be treated as a starting point, not a completed AI facility.

Site Data To Collect Before Engineering

Operators should prepare:

• Country and project location

• Utility service capacity and permitted load

• Actual continuous power available at the site

• Transformer rating, voltage and impedance

• Main switchgear and feeder configuration

• Existing PDU or distribution cabinet information

• GPU platform and expected rack power

• Number of racks and expansion phases

• Summer and winter design temperatures

• Elevation and humidity

• Water quality and freezing conditions

• Existing containers, buildings and foundations

• Road width, crane access and equipment delivery route

• Fiber availability and network carrier options

• Fire protection, drainage and physical security information

Without this data, a supplier can quote equipment, but the project team cannot reliably confirm whether the site is ready for AI.

Совет от профессионала: Ask for a site data sheet before asking for a final cooling quotation. A serious feasibility review begins with operating conditions, not only the requested equipment name.

2. Available Power Is Not The Same As AI-Ready Capacity

Mining operators often start with a statement such as “the site has 5 MW available.” That figure needs to be broken down.

Four Power Numbers Should Be Separated

1. Utility or interconnection capacity
   The maximum capacity approved or contracted with the local utility.

2. Transformation capacity
   The rating of installed transformers and their operating limits under local ambient conditions.

3. Distribution capacity
   The actual capacity of switchgear, busbars, feeders, PDU systems, protection devices and cables.

4. Deployable AI capacity
   The continuous and reliable power that can be delivered to GPU racks while maintaining operating margin, redundancy and maintenance access.

These four numbers may be very different.

For example, a site may have a 5 MW utility connection but only 3 MW of installed transformation capacity. Facilities might feature 3 MW of transformers yet lack sufficient low-voltage distribution for high-density racks. Alternatively, distribution capacity could be adequate while completely lacking maintenance redundancy.

IT Load And Facility Load

AI infrastructure planning should distinguish between:

• IT Load: GPU servers, CPUs, storage and network equipment

• Cooling Load: CDU pumps, dry cooler fans, chillers or other mechanical equipment

• Electrical Distribution Losses: Transformers, switchgear, cables and PDUs

• Facility Load: The total power consumed by IT and supporting infrastructure

A simple planning relationship is:

Facility Power = IT Load x PUE

If a project has 1 MW of IT load and an estimated PUE of 1.30:

1 MW x 1.30 = approximately 1.30 MW facility demand

This is a planning example, not a final design value. Actual PUE depends on ambient conditions, cooling architecture, pump and fan efficiency, electrical losses, controls and the measurement boundary.

A mining site that previously operated at a high utilization rate may not have enough margin for AI workloads, even when its nameplate capacity appears sufficient.

Power Quality Matters More At High Density

GPU systems can create stricter requirements for:

• Voltage stability

• Фазовый баланс

• Short-circuit protection

• Harmonic performance

• Grounding and bonding

• UPS or ride-through strategy

• Generator coordination

• Maintenance bypass

• Dual power paths

• Protection selectivity

A transformer sized only for average load may not be adequate for startup events, transient behavior or future expansion.

Совет от профессионала: Do not use the old ASIC miner count as the basis for the GPU electrical design. Start with the selected GPU platform, rack power, power path and expansion schedule.

3. ASIC Miner Loads And GPU Rack Loads Are Different

An ASIC mining farm and an AI/HPC facility may consume a similar amount of electricity, but their infrastructure requirements can be very different.

Typical ASIC Mining Site Characteristics

An ASIC mining site often includes:

• Many individual mining machines

• Repetitive rack or shelf layouts

• Air cooling or direct liquid cooling

• Simple east-west network traffic

• Limited storage requirements

• High operating hours

• Strong focus on cost per kilowatt-hour

• Fast replacement of individual machines

The electrical system may be distributed across many similar loads. The cooling design often focuses on total heat removal, airflow organization, dust control or miner-level liquid connections.

Typical AI/HPC Characteristics

An AI or HPC deployment may require:

• High-density GPU racks

• High-speed GPU-to-GPU interconnects

• Large east-west data flows

• High-performance storage

• Dedicated management networks

• Higher rack-level thermal density

• Direct-to-chip liquid cooling or another liquid-cooling architecture

• Strict temperature and flow control

• More complex commissioning and maintenance procedures

• Stronger physical and cybersecurity controls

NVIDIA describes the GB200 NVL72 as a rack-scale, liquid-cooled design connecting 36 Grace CPUs and 72 Blackwell GPUs. This type of architecture demonstrates why a traditional mining layout cannot be evaluated only by total megawatts or floor area.

The important question is not:

“Can the existing site supply enough electricity?”

The better question is:

“Can the existing site deliver stable power, heat rejection, network performance and maintainability to the selected AI racks?”

Why Rack Density Changes The Design

Mining containers may distribute power and heat across many machines. AI racks can concentrate a much larger thermal and electrical load into a smaller footprint.

That changes:

• PDU and busway sizing

• Cable routing

• Rack clearances

• Cooling-loop flow rate

• Pressure-drop calculations

• CDU capacity

• Pump selection

• Heat exchanger sizing

• Fire protection zoning

• Service access

• Network pathway design

The GPU OEM’s rack and server documentation should define the actual load, cooling interface and environmental requirements. Generic “AI-ready” labels are not enough for engineering approval.

4. Can A Traditional Air-Cooled Mining Site Support High-Density GPU Systems?

Sometimes it can. Often it needs a major retrofit.

The answer depends on the selected GPU platform and the existing site conditions.

Assets That May Be Reused

A traditional mining site may provide useful infrastructure such as:

• Land and site access

• Utility connection

• Transformer yard

• Existing buildings or containers

• Foundations

• Internal roads

• Drainage

• Perimeter fencing

• Security gates

• Fiber entry routes

• Operations staff

• Existing remote monitoring

These assets can reduce project cost and deployment time if they pass inspection.

Assets That Usually Need Revalidation

The following items should be reassessed before conversion:

• Transformer continuous rating

• Low-voltage distribution

• PDU and cabinet configuration

• Cable tray capacity

• Grounding system

• Cooling plant location

• Outdoor dry cooler space

• Pump station location

• Очистка воды

• Drainage and leak containment

• Fire detection and suppression

• Network room and fiber pathways

• Physical access control

• Maintenance clearances

Air cooling may still be suitable for lower-density GPU systems or hybrid workloads if the selected equipment remains within the site’s thermal envelope. It should not be assumed that the same fans, filters and container layout can support a high-density liquid-cooled rack.

A Practical Conversion Decision

A site is more likely to be suitable when:

• Power capacity is documented rather than estimated

• Transformers and switchgear have tested operating margins

• The site has stable fiber connectivity

• There is room for dry coolers and pump stations

• The ground is suitable for equipment foundations

• Roads support heavy equipment delivery

• The climate and elevation are included in heat-rejection calculations

• Water, glycol and filtration requirements can be managed

• The operator can create a secure AI operating zone

• The electrical and cooling systems can be expanded in modules

A site is less suitable when its only advantages are cheap electricity and empty land.

Совет от профессионала: For a first AI conversion, consider retaining the existing mining operation while building a separately controlled 200 kW pilot block. This creates a real operating test before the full site is converted.

5. Understand The CDU Primary Side And Secondary Side

A CDU, or Coolant Distribution Unit, is commonly used to separate and manage two cooling circuits.

The exact design depends on the GPU platform, coolant type, heat load and facility architecture, but the basic concept is as follows.

Primary Side: Facility Water System

The primary side is connected to the facility heat-rejection system. It may connect to:

• Dry cooler

• Cooling tower

• Chiller

• Primary pump

• Expansion tank

• Water treatment system

• Outdoor piping

• Heat exchanger inside the CDU

The primary loop carries heat away from the CDU to the outdoor heat-rejection equipment.

Secondary Side: Technology Cooling System

The secondary side serves the IT equipment. It may connect to:

• GPU cold plates

• Server manifolds

• Rack manifolds

• Technology cooling supply and return headers

• Secondary pump

• Фильтры

• Expansion components

• Обнаружение утечек

• Temperature, pressure and flow sensors

The CDU transfers heat between the two sides while controlling the secondary coolant conditions supplied to the racks.

Basic Cooling Chain

A simplified liquid-cooling path is:

GPU Cold Plate → Rack Return → CDU Heat Exchanger And Pump → Rack Supply → GPU Cold Plate

The facility-side path is:

CDU Primary Return → Dry Cooler Or Other Heat Rejection → Primary Supply → CDU

This separation can help the operator manage coolant quality, pressure, temperature and maintenance requirements on the IT side.

For a detailed explanation of the cooling architecture, see CDU для центров обработки данных искусственного интеллекта.

What Must Be Defined In The Specification

The project specification should define:

• Primary-side supply and return temperature

• Secondary-side supply and return temperature

• Flow rate on each side

• Pressure range

• Pump head

• Heat exchanger capacity

• Maximum allowable pressure drop

• Тип охлаждающей жидкости

• Materials compatibility

• Filtration level

• Expansion and filling method

• Leak detection and isolation method

• Control protocol

• Redundancy strategy

• Maintenance bypass

A CDU quotation without operating conditions is incomplete.

6. Size The Cooling System From Heat Load, Not Site Power Alone

A dry cooler or CDU should not be selected only because the mining site has 1 MW, 5 MW or 10 MW of electrical service.

Cooling design should start with the actual heat that must be rejected.

Basic Heat-Rejection Framework

A practical planning model is:

Required Cooling Capacity = IT Heat Load + Pump Heat + Fan Heat + Distribution Losses + Design Margin

Most electrical power consumed by IT equipment eventually becomes heat. However, not every facility electrical load enters the same cooling loop.

The engineering team should clarify:

• IT load

• Percentage of IT load handled by liquid cooling

• Air-cooled auxiliary load

• CDU pump power

• Dry cooler fan power

• Room or container heat

• Outdoor design temperature

• Required redundancy

• Future expansion margin

Example Flow Calculation

For an illustrative 200 kW liquid-cooling load using water with a 10°C supply-return temperature difference:

Flow = Heat Load / Specific Heat / Temperature Difference

The estimated flow is approximately:

200 kW / 4.18 kJ/kg·K / 10 K = 4.8 kg/s

This is roughly 17 m³/h for water.

At the same assumptions:

These are engineering examples, not final product specifications. Glycol concentration, fluid temperature, density, heat capacity, altitude and required design margin will change the result.

Температура на входе

Approach temperature describes the temperature difference between the fluid leaving the heat exchanger and the ambient or facility-side condition used for heat rejection.

A smaller approach temperature can require:

• Larger heat-exchanger surface area

• More airflow

• More fan power

• Lower fluid temperature

• Larger dry cooler footprint

A high ambient temperature can reduce dry cooler capacity. A system sized for a mild climate may not deliver the same heat-rejection capacity during peak summer conditions.

Read Сухой охладитель для жидкостного охлаждения центров обработки данных before finalizing the outdoor heat-rejection strategy.

Совет от профессионала: Request dry cooler performance at the project’s actual summer design temperature, not only at a favorable laboratory condition.

7. Check Coolant, Water Quality And Freeze Protection

Liquid cooling is not only a matter of adding pipes to a mining site.

The quality of the working fluid can affect:

• Heat-transfer performance

• Corrosion risk

• Pump life

• Filter loading

• Производительность теплообменника

• Cold-plate reliability

• Warranty compliance

• Long-term maintenance cost

The final fluid specification should follow the requirements of the GPU, server, CDU and piping materials.

Water Quality

The project team should define:

• Проводимость

• Твёрдость

• pH

• Corrosion control

• Particulate limits

• Microbiological control

• Filtration method

• Sampling frequency

• Makeup-water procedure

Do not select a coolant only because it is inexpensive or locally available.

Защита от замерзания

Outdoor piping and dry cooler systems may face sub-zero temperatures. The project may require:

• Glycol mixture

• Система обогрева трубопроводов

• Изоляция

• Drain-down design

• Recirculation during standby

• Сигналы тревоги при низкой температуре

• Freeze protection interlocks

Glycol can change viscosity, heat capacity and pressure drop. Pump sizing and flow calculations must be updated when glycol is introduced.

Leak Detection

A conversion project should consider:

• Point leak sensors

• Cable-type leak detection

• CDU drip trays

• Rack-level isolation valves

• Automatic pump shutdown

• Alarm escalation

• Local visual alarms

• Remote notifications

• Emergency drainage

• Spare seals and hoses

The purpose is not only to detect a visible leak. It is to identify abnormal moisture or flow behavior before it damages computing equipment.

8. Build Controls Around Operations, Not Only Installation

A working cooling system needs clear operating visibility.

A practical monitoring system may include:

• Температура подачи

• Температура на выходе

• Differential temperature

• Supply pressure

• Return pressure

• Differential pressure

• Расход

• Pump speed

• Состояние насоса

• Скорость вращения вентилятора

• Ambient temperature

• Coolant level

• Filter differential pressure

• Leak status

• Door or access status

• Power quality

• История срабатываний сигнализации

A PLC can manage local sequences and interlocks. A BMS, DCIM or remote monitoring platform can provide higher-level visibility, trend analysis and alarm management.

Alarm Logic Should Be Defined Early

The project team should specify:

• Warning thresholds

• Trip thresholds

• Alarm delays

• Sensor-failure behavior

• Pump-failure response

• Fan-failure response

• High-temperature response

• Low-flow response

• Leak response

• Communication-loss response

• Manual override rules

• Restart procedure

Remote monitoring is useful only when the alarm has a clear owner and a defined response.

The project should also consider cybersecurity, user permissions, secure remote access and network segmentation. Connecting a PLC to the internet without access control creates a new operational risk.

9. Network, Fiber And Physical Security Requirements

Mining farms may operate with relatively simple network requirements. AI/HPC systems can require high-bandwidth, low-latency communication between servers, switches and storage.

A conversion plan should assess:

Network Infrastructure

• Fiber availability at the site

• Carrier diversity

• Entrance pathways

• Backbone topology

• High-speed Ethernet or InfiniBand requirements

• Top-of-rack switch location

• Management network

• Out-of-band access

• Storage network

• Redundant network paths

• Fiber patching and labeling

• Electromagnetic separation from power cables

NVIDIA’s current rack-scale AI architecture materials emphasize that networking is a fundamental part of large-scale AI performance. The cooling plant may be technically excellent, but the deployment can still fail if the network cannot support the workload.

Physical Security

Operators should review:

• Site perimeter

• Vehicle and visitor control

• Camera coverage

• Access logs

• Rack or container access

• Network-room access

• Emergency exits

• Fire detection

• Fire suppression

• Cable protection

• Backup power for security systems

• Local regulations and insurance requirements

A mining site can be physically remote and still require data-center-grade access control once it hosts customer AI workloads.

10. Plan Expansion In 200 kW, 500 kW And 1 MW Blocks

Phased expansion reduces technical and financial risk.

These values should be treated as planning bands for IT load unless the project specification states otherwise.

Phase 1: 200 kW Pilot

The 200 kW phase can be used to validate:

• GPU platform compatibility

• Rack-level liquid cooling

• CDU operation

• Dry cooler performance

• Water quality

• Обнаружение утечек

• PLC alarms

• Network performance

• Maintenance procedures

• Actual PUE

The pilot should be designed with room for measurement and troubleshooting. It should not be installed as an isolated experiment that cannot connect to future blocks.

Phase 2: 500 kW Expansion

At 500 kW, the project should formalize:

• Modular CDU capacity

• Dry cooler module arrangement

• Резервирование насосов

• Power distribution blocks

• Network expansion

• Запасные части

• Service access

• Maintenance bypass

• N+1 strategy where required

N+1 means that the system has the required number of operating modules plus one additional module. The correct application depends on the customer’s uptime target and maintenance philosophy.

Phase 3: 1 MW Production Block

At 1 MW, the project should evaluate:

• Separate power and cooling zones

• A/B power distribution

• Multiple CDU modules

• N+1 pumps or fans

• Heat-rejection redundancy

• Fire zoning

• Network path redundancy

• Spare transformer or feeder strategy

• Planned maintenance procedures

• Full factory and site acceptance testing

A single oversized cooling machine may look simple on a quotation. Several coordinated modules may provide better serviceability and expansion flexibility.

For a broader modular deployment framework, see 5MW Modular Mining Farm Design Guide.

Совет от профессионала: Design the 200 kW block so it can become one complete operating zone inside the future 1 MW facility. Future expansion is easier when the first phase already has its own valves, sensors, controls, power boundaries and network identity.

11. Which Mining Sites Are Good Candidates For AI Conversion?

Strong Candidates

A mining site is a stronger candidate when it has:

• Verified utility and transformer documentation

• Sufficient continuous power margin

• Modern switchgear and protection

• Adequate grounding and bonding

• Space for CDU and dry cooler installation

• Reliable road and crane access

• Stable fiber connectivity

• Good drainage and flood protection

• Secure perimeter and controlled access

• Expandable foundations

• A manageable climate and elevation

• Skilled local maintenance support

• Clear permitting and insurance requirements

Weak Candidates

A site deserves caution when it has:

• Only a utility promise without installed distribution

• Undersized transformers

• Seasonal power restrictions

• Poor voltage quality

• No expansion room

• Weak fiber connectivity

• Flooding or standing-water risk

• Narrow roads or limited crane access

• No safe location for pumps and dry coolers

• Uncontrolled water quality

• No leak response procedure

• Weak physical security

• No local service capability

Low-cost electricity is valuable, but it does not compensate for an unsuitable facility.

12. Mining Farm To AI Data Center Conversion Checklist

13. Go Or No-Go Decision

A mining site should move forward only when the following questions have clear answers:

1. Can the site deliver the required continuous power, not only the contracted capacity?

2. Can the transformer and distribution system support the selected GPU racks?

3. Can the cooling system reject the calculated heat at the project’s worst design ambient?

4. Can the CDU maintain the required secondary-loop temperature, flow and pressure?

5. Can the site manage coolant quality, filtration and freeze protection?

6. Can the controls detect and respond to abnormal conditions?

7. Can the network support the workload and future expansion?

8. Can the site meet physical security and fire-safety expectations?

9. Can equipment be delivered, lifted, installed and maintained?

10. Can the project pass factory testing, site commissioning and customer acceptance?

If several answers are uncertain, the project is still in feasibility stage. It is too early to approve a final equipment list or promise a deployment date.

Drolin Box can review the project requirements and prepare the appropriate technical scope for modular mining containers, liquid-cooling systems, CDU integration, dry cooler deployment and phased computing-infrastructure expansion.

For project assessment, submit:

• Project country and site location

• Available and target power

• Transformer information

• GPU or server model

• Expected rack quantity

• Current mining-site layout

• Ambient temperature and elevation

• Water and freezing conditions

• Site photos or drawings

• Required deployment phase

Вы также можете ознакомиться с 500MW Data Center Site Selection Risk Guide before evaluating a larger power opportunity, or contact Коробка «Дролин» for a project discussion.

Часто задаваемые вопросы

Can every mining farm be converted into an AI data center?

No. A mining site may have useful power and land resources, but the transformer, distribution, cooling, network, security and permitting systems must be reassessed for the selected AI workload.

Is a CDU required for every GPU data center?

Not necessarily. The cooling architecture depends on the GPU platform, rack design, liquid-cooling method and facility-water strategy. A CDU is commonly used when the technology cooling loop needs separation and controlled heat transfer from the facility loop.

Can an air-cooled mining container be used for GPU servers?

It depends on the server thermal design and rack power density. Lower-density or hybrid systems may use air cooling, while many high-density GPU platforms require direct liquid cooling or another specialized architecture.

Should dry cooler capacity equal the mining farm’s total power capacity?

No. Dry cooler sizing should be based on the heat-rejection load, liquid-cooled fraction, design ambient temperature, flow, approach temperature, redundancy and future expansion.

What is the safest first step for a mining site AI conversion?

Start with a documented feasibility review and a modular pilot. A 200 kW pilot can validate power, cooling, controls, network and operating procedures before the project expands to 500 kW or 1 MW.

What information should buyers prepare before requesting a quotation?

Prepare the country, available power, transformer rating, GPU or server model, rack quantity, ambient conditions, elevation, water quality, site photos, layout drawings, delivery route and expected expansion phase.