Why AI Data Centers Are Sized in Megawatts

When Samsung SDS announced on November 2, 2026 that it had started building an AI data center in Gumi, an industrial city in southeastern South Korea, the company gave both the site area and the power capacity. The initial power capacity is 60 megawatts, and the design allows expansion in stages to 120 megawatts. Here is why AI data centers are described in megawatts, and what that number really means.

The Samsung SDS project

Samsung SDS is the IT services company of the Samsung group. Its Gumi data center sits on roughly 46,000 square meters of land and is built for power-hungry AI and GPU servers. Construction began after the city approved it in late August, and completion is planned for the end of 2028. Besides running its own GPUs, the company will offer colocation, where customers bring their own servers or GPUs and rent space, power and cooling. Samsung SDS described the project as the first result of the Samsung group's plan, announced in July, to invest about 60 trillion won (about $45 billion) in the southeastern Yeongnam region.

What a megawatt measures

A watt measures how fast electricity is used at a given moment. One kilowatt is 1,000 watts, and one megawatt is 1,000 kilowatts, or one million watts. At the unit level, 60 MW is 60 million watts; a project's rating describes its stated power capacity rather than its measured electricity use at every moment. As simple arithmetic, that is enough to light 600,000 bulbs of 100 watts each at the same time.

Quick definition: Colocation means a data center operator rents out space, power and cooling, while customers install and run their own servers or GPUs.

Server capacity depends on available space, power and cooling. A big building with a small power supply can hold only a few servers. Samsung SDS says the Gumi center has an initial power capacity of 60 MW. Do not confuse megawatts with megawatt-hours: a megawatt is a unit of power, while a megawatt-hour is a unit of energy. Running 1 MW for one hour uses 1 MWh, or 1,000 kilowatt-hours.

The maximum power of one GPU

The NVIDIA H100 SXM GPU has a maximum thermal design power (TDP) of up to 700 watts. Counting GPU power alone, 60 MW divided by 700 W is about 85,700, so more than 80,000 GPUs.

That figure is purely theoretical. Real servers also contain processors, memory and networking gear, and a data center spends electricity on cooling and power conversion. NVIDIA lists server configurations with four or eight H100 SXM GPUs. The PCIe-based H100 NVL is rated at 350 to 400 watts, so the math also depends on which product is used. In practice, the number of GPUs a 60 MW site can hold is far lower than the simple calculation.

Why cooling matters

GPUs that use more power also produce more heat. Servers that run too hot slow down or fail, so cooling is as important as power in an AI data center. Samsung SDS said the Gumi center will use hybrid cooling that combines air cooling with liquid cooling, and that its server rooms are designed for racks with different power densities, including very high-power racks in the future. A rack is the metal frame that holds a stack of servers, and racks packed with AI servers draw more power and give off more heat.

Turning a power rating into an energy calculation

Suppose, as a simple illustration, a load stayed at exactly 60 MW for one hour. It would use 60 MWh, or 60,000 kilowatt-hours, of energy. If it stayed at that level for 24 hours, the calculation would be 60 times 24, or 1,440 MWh. These examples translate power into energy; they are not an estimate of the Gumi facility's daily electricity use. Actual consumption depends on the equipment installed and how heavily it is running.

The distinction also matters when reading the expansion plan. Moving from 60 MW to 120 MW would double the stated power-capacity figure. It would not, by that number alone, establish how many GPUs the operator has ordered, how many servers have been installed or how much computing work the center will perform. Different hardware configurations require different amounts of power, networking and cooling support.

The project thus has a capacity figure, a hardware plan and a completion timetable. Those describe different aspects of readiness: available electrical capacity is not the same as installed computing capacity, and neither establishes that a site is already serving customers.

Key takeaways

AI data centers are sized in megawatts because server capacity depends on available space, power and cooling. Samsung SDS's Gumi center starts at 60 MW and is designed to grow to 120 MW, and with a single H100 SXM GPU rated at up to 700 W TDP, that number gives a rough sense of scale. When you read about a new AI data center, look at the megawatt figure first.

Source: Samsung SDS, NVIDIA