What does a 100 MW AI data centre require?

What does 100 MW mean?

A megawatt measures power at a given moment. In this example, the computers use 100 MW.

The number usually refers to the information technology load: servers, processors, memory, storage and networking. It does not include all the equipment that keeps those computers running. Pumps move cooling water. Fans move air. Electrical equipment changes voltage and converts power. Lights, controls and security equipment also consume electricity.

The grid connection, cooling equipment and backup supply must all support that computer load. Each part needs enough capacity for the electricity entering the building and the heat leaving it.

The International Energy Agency estimates that data centres used 415 terawatt-hours (TWh) in 2024 and could use about 945 TWh in 2030. It estimates that grid constraints could put around 20% of planned projects at risk of delay if those risks are not addressed.[1]

The main test

Establish how much electricity the utility can deliver and when. Then calculate the heat the building must remove.

How does 100 MW become a 138 MW grid request?

Data centres use a measure called power usage effectiveness, or PUE. It divides all electricity entering the building by the electricity used by the computers. A perfect building would have a PUE of 1.00, but real cooling and electrical equipment need power of their own.

In our middle example, PUE is 1.20. A 100 MW computer load needs 120 MW for the full building. Adding 15% planning headroom for growth, operating limits and design uncertainty gives a grid request of 138 MW.

Step	Calculation	Result
Computers	Given design target	100 MW
Full building	100 MW × PUE 1.20	120 MW
Grid planning figure	120 MW × 1.15	138 MW

The 15% headroom is an assumption. The utility and electrical designer must set the figure for the site. In this example, a 100 MW supply cannot cover the computers and the supporting equipment.

How many high-power racks does that represent?

NVIDIA states that one GB200 NVL72 rack can use about 120 kW at full load. Each rack contains 18 compute nodes and 72 Blackwell graphics processors.[2]

Dividing 100,000 kW by 120 kW gives 833.3, or 834 rack equivalents when rounded up. A real data centre may use different equipment. The calculation gives a reference for its electrical and cooling needs.

Item	Reference amount	What the design must provide
AI racks	834 at 120 kW each	Power, network connections, liquid cooling and service access.
Full building	120 MW in the middle case	Transformers, switchgear, pumps, fans, controls and other support equipment.
Grid request	138 MW with our headroom	A firm utility offer, delivery date and agreed backup arrangement.

How much electricity does the building use in a year?

Power measures demand at a given moment. Energy measures use over time. One megawatt-hour (MWh) is one megawatt used for one hour.

Annual energy = computer load × PUE × 8,760 hours × average use.

We assume the computers average 90% of full load across the year. The table compares three PUE values.

PUE	Full-building power	Annual electricity	Cost at €100/MWh
1.10	110 MW	867.2 GWh	€86.7M
1.20	120 MW	946.1 GWh	€94.6M
1.40	140 MW	1,103.8 GWh	€110.4M

The middle case uses about 946.1 GWh per year. At the example price of €100/MWh, that costs about €94.6M each year. Reducing PUE from 1.40 to 1.10 saves about 236.5 GWh, or €23.7 million at the same price.

The electricity price is an example, not a quoted tariff. A commercial model would also need connection charges, taxes, demand charges and hedging costs.

Where does the heat go?

Nearly all the electricity used by the computers becomes heat. A 100 MW computer load therefore creates roughly 100 MW of heat that must be moved from the chips to the outdoor air.

NVIDIA's reference rack needs about 120 kW of cooling capacity and uses direct liquid cooling.[3] Water can carry much more heat than air, which is why modern AI racks use liquid close to the chips.

If water warms by 10°C while carrying away 100 MW, the basic heat equation gives a flow of about 2,388.9 litres per second. A real building divides this flow between many closed pipes, pumps and coolant distribution units.

ASHRAE reports that warm-water designs without chillers can approach PUE 1.10 and use very little cooling water in suitable climates.[4] The Open Compute Project recommends delivering coolant at 30°C or warmer as a durable design target.[5] Local weather determines whether dry coolers can reject the heat throughout the year.

What the cooling design must state

Record how much heat goes into liquid, the water temperatures, the pump power, the hottest outdoor condition, the water treatment plan and what happens after a leak or failed pump.

What happens when the grid supply stops?

Computers cannot wait several minutes for a generator to start. Batteries or another fast supply must carry the building during the changeover or give the computers enough time to shut down safely.

At 120 MW, five minutes of full-building power equals 10 MWh of delivered energy. Our model assumes 90% discharge efficiency and uses only 80% of the battery capacity. That raises the installed battery to about 13.9 MWh.

Five minutes is only a bridge. The design must specify what happens after ten milliseconds, ten seconds, five minutes and four hours. It must also cover battery recharging and the loss of part of the backup supply.

What must be proven before construction?

Before detailed design, the computer plan, grid supply, cooling equipment and backup sequence must be consistent. These are the minimum records we would ask to see.

Question	Evidence required
What will the computers draw?	Rack list, expected use, growth plan and power limits.
What will the utility deliver?	Signed capacity, connection design, delivery date and expansion terms.
Can the building distribute the power?	Electrical diagram, transformer and switchgear ratings, losses and protection study.
Can the building remove the heat?	A complete heat calculation for the hottest weather and failed equipment.
What happens during an outage?	Battery, generator or alternate-supply sequence with fuel and recharge requirements.
Can the owner afford to run it?	Electricity, grid, cooling, water, maintenance and equipment-replacement costs.

If one of these answers is still an estimate with no owner or delivery date, the site is not ready for a 100 MW commitment.

Sources and assumptions

This illustrative model uses the inputs below. The inputs below distinguish assumptions from the vendor reference. The original article was checked on 15 July 2026; the linked sources and passages below were reviewed on 28 October 2026.

Model inputs

Input	Value	Basis and limit
Computer load	100 MW	Illustrative design target.
Power usage effectiveness (PUE)	1.10, 1.20, 1.40	Assumed cases; 1.20 is the middle case.
Average use	90% of full load	Assumed annual average; the model scales facility power with computer load.
Operating year	8,760 hours	A non-leap year; no separate outage allowance is applied.
Grid headroom	15%	Planning assumption applied to facility power; the utility must set the actual requirement.
Rack power	120 kW per rack	NVIDIA GB200 NVL72 full-load reference [2]; other rack designs differ.
Heat and coolant	100 MW, 10°C rise	Assumes the full computer load becomes heat carried by water; heat split and coolant mixture may differ.
Water properties	4.186 kJ/kg·K; about 1 kg/L	Simplified water heat capacity and density for the flow estimate.
Backup bridge	5 minutes at 120 MW	Illustrative full-facility load; duration and supported loads require a site design.
Battery assumptions	90% efficiency; 80% usable capacity	Assumptions used to size installed energy capacity.
Electricity price	€50, €100 and €150/MWh	Illustrative sensitivities; the article shows €100/MWh and excludes other charges.

Source trail

[1] IEA: Energy and AI, executive summary (2025)
Location
Sections on data-centre electricity demand and grid integration.
Supports
The 2024 global demand estimate, the 2030 projection and the risk of project delays.
Limit
Global estimates and scenarios, not measurements or forecasts for this example facility.
[2] NVIDIA: Mission Control administration guide, power FAQ
Location
Example 1: configuring a power domain for a GB200 NVL72 rack.
Supports
The approximate 120 kW full-load rack, 18 nodes and four GPUs per node.
Limit
A vendor reference configuration; actual equipment and operating loads may differ.
[3] NVIDIA: GB200 NVL72 designs for the Open Compute Project (2024)
Location
Liquid cooling manifolds and floating blind mates.
Supports
The rack's direct-liquid-cooling design and approximately 120 kW cooling capacity.
Limit
A vendor description of one rack design, not a facility-wide cooling specification.
[4] ASHRAE: AI data centre integrated design principles
Location
GB200 dry-cooled architecture comparison and case studies.
Supports
The cited PUE example and warm-water, chiller-less cooling discussion.
Limit
Examples depend on equipment, climate and operating conditions; PUE 1.10 is not guaranteed.
[5] Open Compute Project: 30°C Coolant, revision 1.0 (2024)
Location
Page 25: cooling temperature and climate analysis.
Supports
The 30°C technology-cooling-system temperature discussed in the article.
Limit
The analysis uses stated operating and climate assumptions; it does not prove a chiller-less design works at every site.