First, the facts: According to the International Energy Agency (IEA), the global energy demand of data centers currently stands at approximately 500 terawatt-hours (TWh) per year. In total, this corresponds to roughly 1.5 to 2 percent of global electricity consumption. To put this into perspective: 500 TWh/year is more than Germany’s total annual electricity consumption.
And the trend is clear: global electricity consumption is growing steadily. According to the IEA, in 2025 alone, the energy demand of data centers worldwide rose by about 17 percent. Against this backdrop, international market research firms such as Gartner predict that global electricity consumption is highly likely to double to over 900 TWh by 2030. The IEA estimates a figure of around 945 TWh.
Looking at the global distribution of data center electricity consumption, the Canadian data analysis and forecasting firm Incorrys concludes that the lion’s share of global data center electricity consumption is accounted for by the U.S. and China—together nearly 70 percent. IEA figures, which still refer to 2024, put the U.S. share at 45 percent and China’s at 25 percent. Europe’s share in this area stood at 15 percent at that time.
The dynamic nature of developments in recent years is evident from the fact that the average annual growth in electricity consumption in the data center industry has been around 12 percent since 2017. This means that growth in this segment is four times higher than the general global increase in electricity demand. Given the global hype surrounding AI, current growth rates are unlikely to remain at these levels. The IEA therefore currently expects global electricity consumption to reach around 1,200 TWh by 2035.
If 100 percent of the electricity used were converted into “computing power,” that might still be acceptable. But even modern data centers have an average Power Usage Effectiveness (PUE) of only 1.3 to 1.5. Specifically, this means that between 25 and 35 percent of the total energy consumed is lost as waste heat. This waste heat is primarily generated by the cooling required for the facility and the uninterruptible power supplies (UPS) needed to compensate for power fluctuations and outages. Modern Hyperscale data centers operated by major cloud providers currently achieve PUE values of 1.1 to 1.2, setting the standard for best practices.
The boom in energy-intensive AI processors is gradually pushing traditional cooling technology to its limits, as the use of AI chips generates a great deal of heat in a very confined space. Switching to more efficient water cooling would be one way to address the waste heat problem, or tackling the issue more radically by changing the system architecture in data centers. So far, these centers have relied on 480-V AC architectures and primarily silicon-based semiconductors. Increasing the supply voltage would enable the transmission of high power with lower losses. This would, on the one hand, reduce the weight of the required cabling and, on the other, avert the looming thermal overload.
In concrete terms, this means that in 800-V DC systems, less current flows at the same power output due to the higher voltage. Since losses are known to increase quadratically with current, waste heat is significantly reduced. To operate as efficiently as possible with the required power electronics, the era of AI has finally ushered in the age of wide-bandgap components such as SiC and GaN in modern data centers. Manufacturers have already gained experience with 800-V systems in recent years in the e-mobility sector, where 800-V architectures have also been making waves in terms of efficiency and power for a relatively short time.
As a technology leader in the AI sector, Nvidia demonstrated nearly two years ago how the challenges associated with 480 VAC at the architectural level and the 54-V standard at the rack level can be resolved. Currently, data centers perform AC-to-AC conversions in several steps before the power is converted to direct current in the server rack. In the future, there will be only a single AC-to-800-V DC conversion. This 800 VDC will then be distributed within the data center and used directly by the server racks.
To bring this project to fruition, Nvidia is collaborating with nearly 30 partners worldwide, including plant engineering firms such as ABB, Schneider Electric, and Siemens, as well as gas turbine manufacturers like GE Vernova. However, the majority of these partners are semiconductor specialists such as Analog Devices, EPC, Infineon Technologies, Innoscience, Mitsubishi Electric, MPS, Navitas, onsemi, Power Integration, Renesas, Rohm Semiconductor, STMicroelectronics, and Texas Instruments.
The following articles from Infineon Technologies, Nexperia, and STMicroelectronics, for example, illustrate how power semiconductor specialists are specifically adapting to the new era of 800-V data centers and what benefits their products offer for the AI boom. eg
