I. Changes in Power Supply Architecture in the AI Era
Over the past few years, data center power supply systems have evolved along a relatively clear path: AC UPS systems have continuously improved in efficiency, become modularized, and optimized for redundancy, forming a highly mature infrastructure foundation in the era of general computing and cloud computing. However, the emergence of AI computing power, especially the high-density deployment of intelligent computing centers , is putting structural pressure on this path.
This “pressure” is first reflected in the significant increase in power density. From traditional 300-watt servers to today’s 600-kilowatt or even 1-megawatt racks from Nvidia, the overall power has increased more than a hundredfold. This means that continuing to rely on AC power will quickly reach its limits in terms of line losses, power efficiency, and space occupation.
Secondly, there’s the change in the power supply itself. The diversification of power supply, from the power grid to chips, aims to match diverse computing power demands. When discussing AI, high power density and DC power supply have become essential topics. When the power density of a single cabinet exceeds 300 kilowatts, the power supply mode will shift from AC to DC, and further towards 800V DC or even higher voltage levels.
II. 800V DC is unlikely to dominate the market.
800V DC is often described as the “ultimate form” of future data centers, but the reality of the industry is far more complex than this vision. This is because most data centers will long exist in a state where various types of servers and different forms of computing power coexist. For example, the server hosting requests received by Medtronic Technology in the past year have included high-density intelligent computing, low-density general computing, and edge computing servers coexisting in one room. This dictates that the power supply architecture of data centers must have multiple technical approaches, and a one-size-fits-all approach is not feasible.
It’s important to understand that data center lifecycles typically exceed ten years, while GPU servers and chip platforms only iterate two to three generations. This creates a natural mismatch between the “slow variables” of the power supply system and the “fast variables” of computing equipment. Some companies have been forced to use methods like converting 800V to 240V DC or 800V to 380V AC to ensure compatibility with existing servers.
A more pressing issue is the fragmentation of standards. Globally, multiple technical routes exist surrounding 800V and ±400V, and different standards organizations and chip manufacturers have differing choices. For example, should the PICU be placed in a rack or recessed to the node level? How can potential sparking issues in board-level maintenance be avoided? These questions involve complex power supply chips and system design, and a consensus has yet to be reached.
Therefore, Medtronic believes that 800V DC is more likely to exist as a “special power supply solution” for high-density computing servers in the long term, rather than a unified infrastructure base for all scenarios. What the industry really needs may not be to bet on a certain voltage level, but rather the ability to build a power supply system that is compatible with different types and densities of servers.