AI Data Center Power Grids Are Shifting from AC to DC
Power in AI data centers is switching to direct current. On the 24th, global power equipment maker ABB unveiled 'Infinitus,' a direct current (DC) power solution for AI data centers, developed in partnership with Nvidia. The move comes as the power consumption of a single next-generation AI server rack (a frame housing multiple servers) is set to rise from about 200kW today to over 1MW (1,000kW), making it hard for the conventional alternating current (AC) approach, which requires repeated voltage changes, to avoid power losses, heat generation, and equipment space problems.
ABB described Infinitus as the industry's first integrated DC portfolio covering the entire power supply chain of a data center, from the grid to the server rack. The key lies in reducing the power conversion stages that were repeatedly carried out inside data centers. A typical data center receives AC power from the grid, transforms and converts it across multiple stages, and then changes it into DC that servers can use—the more conversion stages there are, the greater the energy losses and heat, and the more space taken up by transformers and power conversion equipment. Infinitus is built around a solid-state transformer (SST) that directly converts medium-voltage AC into 800V DC. The 800V DC is then distributed within the data center and delivered to server racks. The structure ties everything—from medium-voltage power supply and AC-DC conversion to DC distribution and protection and cooling optimization—into a single system.
Nvidia views the existing power supply architecture as one of the bottlenecks for next-generation AI data centers and argues that the 54V DC method used inside racks cannot handle next-generation power demands, pushing 800V DC as the power standard for next-generation AI factories. Supplying power at higher voltage lowers the current required to deliver the same amount of power, reducing cable and copper usage as well as power losses. The message: it's not enough to just swap out the GPUs.
ABB and Nvidia announced their collaboration on 800V DC power architecture development for gigawatt-scale next-generation AI data centers last October. Nvidia is also developing related standards with Google and Microsoft through the Open Compute Project (OCP), and said more than 80 equipment and infrastructure companies are currently developing products to those specifications.
ABB and Boston Consulting Group (BCG) estimated that adopting a DC distribution architecture could raise a data center's energy efficiency by more than 5 percent. Based on a 500MW data center, that translates to roughly 25MW of power that could be redirected to additional IT computing.
ABB expects data centers to move through a hybrid phase combining existing AC equipment with 800V DC before shifting long-term toward DC-centric facilities, forecasting that around 25 to 40 percent of newly built data center capacity in 2030 could use DC distribution. ABB plans to begin producing Infinitus's key components within the next 12 months and expects the first DC-native data center applying the full portfolio to go live within two to three years. It did not disclose specific customers, installation locations, or pricing.
