In the past, when we talked about data centers, we talked about racks, fiber, servers, and air conditioning. In the AI era, the core problem of data centers has changed—it is no longer an IT engineering problem, but an energy and civil engineering problem.
Put simply, the more GPUs you buy, the thing that eventually chokes you is not the chips, but where the power comes from, how to get it into the server room, and how to carry the heat away.
First, look at the scale of power consumption.
A traditional enterprise rack is generally 5 to 10 kilowatts, which is already considered high density.
NVIDIA’s latest GB200 NVL72 full rack, with 72 Blackwell GPUs and 36 Grace CPUs packed into one rack, has a rated power consumption of 120 kilowatts, and at full load it can surge to 130 to 132 kilowatts—more than ten times that of a traditional rack.
If you line up several hundred such racks in a single data center building, the electricity demand of one building alone immediately reaches the hundred-megawatt scale.
At a global scale, Gartner estimates that worldwide data center electricity consumption in 2026 will grow 26% to 565 TWh, with AI as the main incremental driver, while the peak power consumption of AI chips themselves is expected to reach 38 gigawatts by the end of the year. That number is already close to the electricity consumption of a medium-sized country.
When power consumption gets this high, air cooling simply cannot handle it.
The GB200 NVL72 explicitly requires direct liquid cooling into the chips, and the supporting CDU units from vendors are sized at the 250-kilowatt level.
Liquid cooling is not as simple as adding a water pipe. It means the entire system has to be rebuilt: the underlying piping in the server room, the coolant loop, secondary-side heat exchange, and leak detection.
Old data centers cannot be retrofitted; they have to be newly built. And new construction is not just about finding any piece of land—it has to be close to a water source, close to a substation, and close to a suitable climate zone. With all these conditions combined, there are actually very few viable locations.
The real challenge is the power interconnection itself.
The cycle for buying GPUs is a few months, and the cycle for building a liquid-cooled data center is more than a year. But pulling a high-voltage dedicated line at the hundred-megawatt level, or expanding a 500 kV substation, often takes three to five years. The waiting queue for the large grid has already been pushed all the way to the end of this century. Nuclear power construction cycles start at ten years, which is far too long for companies trying to keep up with the pace of AI.
So you can see that large companies are now shifting their choice from “hook into the big grid” to “solve it ourselves”: solar plus storage, self-built gas-fired power plants, or even long-term PPAs for small modular nuclear power. The logic for data center siting has also changed. In the past, it followed fiber and users; now it follows power and water. Places like Texas, Iowa, and Inner Mongolia—where electricity is cheap, land is inexpensive, and the climate is suitable—are beginning to become the main battlefield.
After the power comes in, there is still the issue of distribution. In traditional data centers, electricity passes through multiple stages of transformation before finally reaching the servers, with a lot of loss and bulk along the way. With AI data centers’ extremely high rack power density, the industry is pushing the entire electrical architecture toward DC. The idea is to use solid-state transformers (SSTs) to convert medium-voltage AC directly to a DC bus, reducing steps, improving efficiency, and freeing up data center space. This is a reconstruction that extends all the way from chips to the power grid.
So when we look at AI data centers today, they no longer look like an IT project; they look more like a power and infrastructure project. The parts belonging to semiconductors—GPUs, HBM, networking—have delivery cycles measured in quarters, while the parts belonging to energy and civil works—power, distribution, liquid cooling, substations, civil construction, permits—have delivery cycles measured in years. When the speed of compute expansion collides with the construction speed of the physical world, the latter becomes the real constraint determining how fast AI can run. If you want to understand the next stage of AI infrastructure, you have to look at electricity; otherwise, you cannot see the full picture.