U.S. data centre electricity use is expected to rise to four times the current level by 2035, accounting for 20 percent of total power generation, a forecast showed. As demand for artificial intelligence computing surges, data centre expansion is accelerating faster than expected, but power grids are not keeping up, making supply bottlenecks a key variable in competition over AI infrastructure.

Tech news outlet TechCrunch reported on Monday that BloombergNEF (BNEF), in its latest report, projected a sharp rise in power consumption by U.S. data centres over the next 10 years as demand for AI training and inference surges.

The report estimated that U.S. data centre installed capacity will reach about 200 gigawatts within the next 10 years. Nearly half of that is expected to be used for computing for AI training and inference.

Demand for AI chips is also expected to concentrate in the United States. BNEF predicted that, based on power demand, the United States will absorb 64 percent of the world's AI chips by 2033. That means the centre of AI computing infrastructure will remain in the United States for the time being.

The latest projection was sharply revised up from last year. BNEF's forecast for data centre power demand in 2035 was 83 percent higher than its projection in December last year. It said the faster-than-expected pace of data centre construction in the United States also raised the power demand outlook.

Other organisations are making similar projections. The Electric Power Research Institute (EPRI) more than doubled its forecast for data centre power demand last year, and S&P raised its related projections by more than a third between October last year and April this year.

The problem is that many new data centres are concentrating on regional power grids that already lack spare supply capacity. BNEF expected that most data centres built over the next 10 years will be connected to areas where the burden on power supply is heavy.

At PJM Interconnection, the largest U.S. grid operator spanning Virginia to Illinois, 34 percent of total electricity is projected to be supplied to data centres. ERCOT, which operates the Texas grid, was also estimated to have data centres use 22 percent of generation capacity.

PJM is already one of the United States' largest data centre clusters. But grid connections have been delayed as applications to connect power plants and large power users flooded in at once. PJM even halted applications for new generation resources to connect to the grid for four years because of the burden.

It resumed the connection process in April, but the supply-demand imbalance is still continuing. U.S. power company American Electric Power (AEP) even mentioned the possibility of leaving PJM as the situation worsened. A 76 percent rise in power prices in the region over the past year also shows supply pressure is increasing.

Despite pressure on the grid, data centre demand is not easing. In PJM's recent capacity auction, data centres accounted for 38 percent of total power demand bids. That means investment by AI data centre operators is continuing despite grid congestion and rising costs.

Even if the United States remains the hub of AI computing, data centre power demand is expected to expand rapidly worldwide. BNEF estimated that if AI adoption continues at the current pace, the additional electricity newly required by data centres worldwide will reach 1,935 terawatt-hours per year by 2033. That is comparable to India's annual power consumption.

The market is producing analysis that the key to competition over AI infrastructure is shifting from building data centres themselves to securing power. As demand for AI computing surges, whether generation capacity and transmission and distribution networks are expanded is expected to emerge as a key variable determining the pace of data centre investment.