[DigitalToday reporter Jinju Hong (홍진주)] The power infrastructure of artificial intelligence data centres is increasingly likely to shift from an alternating current (AC)-centred model to a direct current (DC)-based one. As surging AI computing demand exposes the limits of existing power structures, DC architecture is emerging as an alternative in terms of efficiency and cost.
According to online outlet Gigazine on March 26 (local time), next-generation AI chips and infrastructure designs unveiled at Nvidia GTC 2026 are accelerating the shift. Major power companies such as Vertiv, Eaton and Delta Electronics are also responding to market changes by announcing DC-based data centre designs in succession.
Existing data centres have been designed around AC power supply. Power is supplied as medium-voltage AC, converted to low-voltage AC, then passed through uninterruptible power supply (UPS) equipment, repeatedly converted between DC and AC, and finally converted again to DC inside servers before being delivered to chips. This structure repeats multi-stage conversions, increasing energy loss, heat generation and facility complexity.
This structure is showing its limits in the AI era. That is because power density has surged sharply, with AI data centres consuming as much as 1 megawatt (MW) per rack. According to Nvidia, a 1 MW rack requires up to 200 kg of copper, and in large data centres that scale increases to hundreds of thousands of kilograms. Losses during power conversion and rising current are cited as the main causes.
The industry is therefore turning to high-voltage DC architecture. For example, converting 13.8 kV AC to 800V DC can remove multiple conversion stages, reducing power loss and improving system efficiency. Analysis also suggests it can cut copper use by about 45 percent by delivering more power through the same conductor, and improve overall energy efficiency by about 5 percent. Some assessments also say gigawatt (GW)-class data centres could cut total cost of ownership (TCO) by up to 30 percent.
A DC-based structure also has advantages on the facilities side. It can boost system reliability by reducing the number of fans and power supplies, and it could also reduce heat and save space. Racks apply a method that uses small converters to transform power into voltages suited to GPUs and CPUs.
The shift is already becoming reality in some regions. According to global market research firm Omdia, China has already introduced high-voltage DC data centres, and in the United States, Meta, Microsoft (MS) and the Open Compute Project are cooperating to test racks based on 400V DC.
Companies are also moving faster. Vertiv unveiled 800V DC infrastructure integrated with Nvidia Vera Rubin, while Eaton is developing a semiconductor transformer for DC power distribution. Delta Electronics also introduced a DC rack solution that includes battery backup.
Challenges remain substantial. Insufficient capacity to produce DC-only equipment, the need to expand supplies of semiconductors and materials, and uncertainty over long-term demand still persist. The industry has pointed out that the pace of transition could be limited unless large-scale investment and a supply chain overhaul proceed in parallel.