[Can our power grid accommodate 260,000 GPUs?]

[Can our power grid accommodate 260,000 GPUs?]

(This is not intended for political or purpose. This is about the technical issues of the grid.)

This is not to say that you can't use GPUs because you don't have enough electricity. There is enough electricity. I'm not saying that we should talk about the lack of power grids again. Let's say that AI data centers are built in provinces with enough power grids and power generation facilities. Can they actually operate?

A data center with a capacity of 50MW or more is called a hyperscale center, and the first data center of this size to focus on AI modeling is xAI's Colossus Center. It is a famous anecdote that mobile generators were mobilized across the United States to supply electricity due to a lack of initial power systems.

The current consumption load of the colossus center is up to 200MW, which uses enormous ESS. This is because problems that put a lot of load and impact on the electrical grid occur in the process of learning the AI model.

The 154kV substation handles a maximum load of 500MW, but assumes a load of 400MW at most due to safety factor. However, the 200MW Colossus Center starts learning the model, and the load goes from 10MW to 200MW in one second, and if the learning is stopped, it drops back to 10MW in one second.

If half of the total capacity of the 154kV substation is frequently turned on and off, the frequency and voltage difference occur frequently, and the substation cannot withstand it. Therefore, xAI delays load fluctuations by using ESS, because if the local power company does not do this, it will cut off electricity.

The power load of substation units in many provinces in Korea is at most 200-300 MW. Therefore, large AI data centers cannot be attached right here, and centers can be installed in substations in GW units or systems that are already in demand.

For example, to build a 200MW AI data center in Haenam, it is difficult to handle with a general KEPCO substation, and special facilities must be supplemented or scaled down. In addition, if there are many volatile renewable energy facilities in the area, there are more to consider. Instead of a traditional alternating current-electric-based system facility, it should be composed of a completely different method of facility or an independent network in which load fluctuations in AI data centers do not affect the surroundings.

The core of the discussion above is the actual load pattern and characteristics of the AI data center, but there is no public information at all. Before discussing whether there is enough electricity or enough power grid, there are many tasks to be done, such as how giant AI data will affect the power grid in real time, what will happen to the load pattern, and how to guide modeling and service composition.