Stories about AI usually end at the chip: which accelerator is how many times faster, which company bought how many. But you can plug the chip in and nothing happens if the power does not arrive.
Through 2026 this problem has been moving from the back of the news to the front.
The scale, in numbers
The International Energy Agency projects that electricity consumption by data centres, AI and crypto-related industries will exceed 1,000 terawatt-hours in 2026. Data centres’ share of total electricity use is estimated to double from about 2.1% in 2022 to about 4.4% in 2026.
There are domestic projections too. Per reporting in Electronic Times, Korean data centre power demand is expected to pass 1.5 gigawatts within three years.
If gigawatts do not mean much to you, try it this way: the installed capacity of one large nuclear reactor is roughly 1 gigawatt. It means supply on that scale has to be newly added for data centres alone.
Why it became a problem so suddenly
Three things overlapped.
1. Density changed
Conventional server racks drew a few kilowatts each. A rack filled with AI accelerators goes to a far higher density. Several times the power goes into the same floor area, and that much heat comes out. You cannot simply stuff it into an existing data centre building.
2. Cooling scales with it
Use a lot of electricity and you generate a lot of heat. Removing the heat takes more electricity. Past the point where air cooling cannot cope you move to immersion and liquid cooling, and then a new water problem appears.
3. The grid is a multi-year construction project
A data centre takes one to two years to build. Transmission lines and substations do not. In Korea the capital region’s grid is short of headroom, and new transmission line construction has been delayed by years through local opposition and permitting. Demand grows by the quarter; supply grows by the year.
Who pays the bill
In the US the argument has already started. As retail electricity rates rose in areas dense with data centres, people began asking whether ordinary households should share that burden. That is where the discussion of data-centre-specific tariffs begins.
The point at issue is simple.
| Position | Argument |
|---|---|
| For a dedicated tariff | Whoever created the demand surge should bear the cost of reinforcing the grid |
| Against | It weakens industrial attraction and investment incentives. Differentiated rates set a precedent for other industries too |
The same question is coming to Korea. Right now grid connection queues and siting regulations bind first, but if demand grows as projected, a debate about rate structure is unavoidable.
Reading it alongside Nvidia’s results
On 26 August 2026 Nvidia guided to $108 billion in Q3 revenue. It means the number of accelerators sold keeps climbing.
Layer one remark from CFO Colette Kress on top and the picture completes. On the cause of the memory shortage, she said it was “in significant part brought about by the build-out of AI infrastructure itself.”
Power has the same structure. AI infrastructure drives up the price of its own inputs. Memory can be expanded within a few years by building fabs; the grid is slower than that. Which is why people are saying the real bottleneck of the next few years may be power rather than chips.
What reaches individuals
- The rate structure debate. How costs are split between industrial and residential users comes back to the table
- Local conflict. Data centre siting and transmission lines are already live local issues in some places
- Cloud pricing. GPU instance rates are not unrelated to the price of electricity. It shows up in the costs of individual developers and small teams
- Renewable contracts. When large operators lock up volume with long-term power purchase agreements, conditions change for the rest of the market
In short
- On IEA projections, 2026 electricity consumption by data centres, AI and crypto passes 1,000 terawatt-hours, with the overall share rising from 2.1% in 2022 to 4.4%
- Korean data centre power demand is projected to pass 1.5 gigawatts within three years — the scale of one large nuclear reactor
- The problem is less the total than density, cooling and how fast the grid can be built. Data centres take one to two years; transmission takes several
- Data-centre-specific tariffs are under discussion in the US. Korea will find the same question hard to avoid
- The bottleneck of the next few years may be power, not chips
Frequently asked questions
If data centres use electricity, does my household bill go up?
Not through a direct chain of cause and effect. Rates are set by fuel costs, grid investment and policy decisions together. That said, if a demand surge increases grid reinforcement investment, that cost enters the rate system in some form, and how to allocate it is exactly what is being argued about now.
Does simply using AI models consume a lot of electricity?
The power used by a single query is small. The problem is scale. When hundreds of millions of people use it several times a day the total grows, and the inference share keeps rising relative to training. It is less a matter of individual thrift than of infrastructure design.
Do renewables solve it?
The generation mix addresses emissions, but grid capacity and transmission are separate problems. Solar and wind output varies by time of day, so matching them to a data centre running 24 hours requires storage or other generation alongside.
Why are Korean data centres concentrated around Seoul?
Latency, staffing and network interconnection points. But with the capital region’s grid short of headroom, policies to encourage regional dispersion are also under discussion, and in actual siting decisions the ability to secure power is becoming an ever larger variable.
Leave a Reply