Skip to content
PIDCPisico Data Center
ENVI
On this page

Guide

What is an AI data center?

An AI data center is a building, or a campus of buildings, designed to power and cool large clusters of GPUs and other AI chips. It does the same basic job as any data center, with far more power and heat packed into the same space.

Illustrative GPU rackIllustrative GPU rack: power shelves at top and bottom, compute trays with GPUs on cold plates, switch trays in the middle, and coolant supply and return lines on the side.1234
  1. Compute trays: GPUs on cold plates
  2. Switch trays that link the GPUs
  3. Power shelves
  4. Coolant supply and return
Illustrative GPU rackNot to scale

What it does

Training and inference

Training

Thousands of GPUs work together for weeks or months to build an AI model from large amounts of data. They must sit close together on very fast networks, so training favors large, dense sites.

Inference

Running a trained model to answer requests, from chat assistants to image recognition. Inference runs around the clock and grows with the number of users, so it also needs good connections to the networks those users are on.

How it differs

The same kind of building, at many times the power

Power drawn by one rack, kW
≈ 8Average rack in data centers worldwide, 2024[1]
132One NVIDIA GB200 NVL72 rack, 72 GPUs[2]

One current-generation GPU rack draws about 16 times the power of an average rack today. That changes the electrical supply and the cooling around it.

FeatureConventional data centerAI data center
Main hardwareCPU servers and storageGPU servers linked into large clusters
Power per rackMostly under 10 kWTens of kW; about 130 kW for the newest GPU racks
CoolingMostly airLiquid piped to the chips, with air for the rest of the room
NetworkBuilt for traffic to and from usersVery fast links between GPUs inside the cluster, plus links out to users and clouds
Campus sizeA few MW to tens of MWTens to hundreds of MW
Grid connectionMedium voltage or 110 kVHigh voltage, often 110–220 kV or above

Power

Why power decides where they are built

Data centers used about 415 TWh of electricity in 2024, around 1.5% of the world's total. The International Energy Agency expects that to more than double, to about 945 TWh by 2030, with AI the main driver.[3]

Data center electricity use worldwide, TWh
4152024
≈ 9452030, IEA projection

So the first question for any AI campus is power: a high-voltage grid connection with capacity to spare, room to add more over time, and supply that holds up around the clock. Land, buildings and equipment can be arranged. A new high-voltage connection can take years.

That is why PIDC chose a site beside the 220 kV grid.See the site

Cooling

Cooling and water

GPU racks give off too much heat for air alone. Current-generation systems such as the GB200 NVL72 are liquid-cooled: coolant flows through cold plates mounted on the chips and carries the heat away.[2]

Liquid cooling loopLiquid cooling loop: coolant takes heat from cold plates on the chips to a coolant distribution unit, which passes it to a second loop that releases it outdoors through dry coolers or cooling towers.Cold plateson the chipsCDUcoolant distribution unitDry cooler orcooling towerrackoutdoors
  • Cool coolant
  • Warm coolant
Liquid cooling loopSchematic

Outdoors, that heat is released through dry coolers, cooling towers or a mix of the two. Closed-loop designs recirculate the same coolant and use little water. Evaporative designs use more water but less electricity. The right balance depends on the local climate and water supply.

Efficiency is measured as PUE, power usage effectiveness: the facility's total power divided by the power used by the computing equipment. A PUE of 1.0 would mean no overhead at all. The industry average was 1.56 in 2024.[1]

Power drawn from the grid for 100 kW of computing, kW
156PUE 1.56, industry average in 2024[1]
100PUE 1.0, no overhead
  • Computing equipment
  • Cooling and other overhead

For the province

What a campus brings to a province

  • Large investment in buildings, electrical and cooling systems, and fiber.
  • A steady local buyer for power generated in the province.
  • Skilled jobs in electrical, mechanical and network engineering and in operations. Data centers employ fewer people than factories of similar cost, but the work is technical.
  • New power and fiber infrastructure that other digital businesses can build on.

Terms

Terms you will hear

GPU
Graphics processing unit: the chip that does most AI computing.
Rack
A standard cabinet that holds servers. Power density is usually counted per rack.
MW IT
Power available to the computing equipment itself, not counting cooling and losses. Data center size is usually quoted this way.
PUE
Power usage effectiveness: total facility power divided by IT power. Lower is better.
Liquid cooling
Coolant piped to plates on the chips, which removes heat far better than air.
Colocation
Renting space, power and cooling to a customer who brings its own servers.
Hyperscaler
A very large cloud provider that runs its own fleet of data centers worldwide.
Redundancy (N+1, 2N)
Spare capacity so one failure does not stop service. N+1 adds one spare unit; 2N duplicates the whole system.
Latency
The time data takes to cross a network, measured in milliseconds.
Cable landing station
Where a submarine fiber cable comes ashore and joins land networks.

Sources

Where the numbers come from

  1. Uptime Institute, Global Data Center Survey 2024 (rack density, PUE).
  2. Supermicro, SRS-GB200-NVL72 rack specification (132 kW, 72 GPUs, direct liquid cooling).
  3. International Energy Agency, Energy and AI: Executive summary (2025).

Questions about the campus?

Contact us