The next revolutionary direction of AI infrastructure: 800VDC The latest article from Semianalysis points out that AI data centers are undergoing a very low-level power architecture upgrade: gradually moving from traditional AC power systems to 800V high-voltage DC, which is 800VDC. This change may seem like a power engineering issue, but fundamentally it will affect the cost of AI computing power, data center design, and the entire power equipment industry chain. Let's start with the simplest logic: AI cabinets are consuming more and more electricity. The power of server cabinets in the past was relatively low, and they could still cope with traditional 48V or 54V low-voltage power supply. But when the AI cabinet moves towards 600kW or even higher power, if the voltage does not increase, a very large current is required. The higher the current, the thicker the copper wire and copper bar, the more severe the heat generation, the higher the power loss, and the less space left for GPU, network, and cooling system in the cabinet. The function of 800VDC is to reduce current by increasing voltage. After the current drops, copper consumption, heat generation, and conversion losses will all decrease, and the power supply system will also be easier to support high-density AI cabinets. The article estimates that 800VDC can reduce facility level electricity consumption in data centers by about 5%. If it is a data center with a 1GW IT load, this is equivalent to saving or releasing over 50MW of continuous power, and the value is very considerable. On a deeper level, 800VDC is not a single component upgrade, but a reconstruction of the entire data center power chain. The article divides migration into four steps. The first step is not to change the original computer room, but to add a high-voltage DC power cabinet next to the AI cabinet to convert AC power to 800V DC power. The second step is that with the new generation of AI cabinets starting to natively support 800VDC, 800V power supply has shifted from pre layout to a mandatory requirement. The third step is to widely use 800V DC power distribution within the data center, which will weaken the role of many traditional AC power distribution equipment. The fourth step is that new equipment such as solid-state transformers may further integrate transformer, rectifier, energy storage, and protection functions, making the power supply chain shorter and more efficient. The industrial impact is also significant. The most direct beneficiaries in the early days were high-voltage direct current power cabinets, also known as HVDC Power Racks. It is responsible for converting traditional AC power into 800V DC power, while integrating battery backup, supercapacitor, power management, and protection functions. The article estimates that the single cabinet price of such power cabinets can reach 400000 to 500000 US dollars, far higher than the level of about 40000 US dollars for traditional AC power cabinets. This means that for every new batch of high-density cabinets added to AI data centers, the value of the power side will significantly increase. The second type of beneficiaries are power conversion and power electronics manufacturers. In the 800VDC architecture, AC-DC and DC-DC conversion will become more critical, and power conversion will become increasingly closer to GPU blades and computing trays. This will drive the demand for high-efficiency power modules, power semiconductors, connectors, magnetic components, and thermal management solutions. The third type of beneficiaries are energy storage and transient power management devices. The power consumption of AI cabinets fluctuates greatly, and GPUs experience millisecond level power spikes. Traditional centralized America may not be suitable for handling such rapid fluctuations, so cabinet level BBU batteries and supercapacitors will become more important. BBU is responsible for short-term power outage endurance, while supercapacitors are responsible for absorbing instantaneous power shocks. The fourth category of beneficiaries is DC distribution and protection equipment. After entering the facility level 800VDC, the data hall requires new equipment such as DC busbars, DC protection modules, and solid-state circuit breakers. The technical threshold here is higher because direct current does not have the natural zero crossing like alternating current, making it more difficult to extinguish arcs and requiring higher protection systems. Under pressure are some traditional AC distribution chains, including low-voltage America, AC switchgear, AC PDU, cabinet PDU, and traditional bus ducts. They won't disappear immediately, as a large number of data centers still need to be compatible with old equipment, CPU cabinets, storage, and network devices. But in the most advanced AI clusters, the added value will gradually shift towards high-voltage DC power cabinets, centralized rectifiers, DC distribution, energy storage, and solid-state transformers. Simply put, 800VDC is a new stage for AI infrastructure to move from buying more GPUs to delivering every watt of electricity more efficiently to the GPUs. The cost of future AI inference not only depends on chip performance, but also on whether power can be delivered to cabinets and chips in a low loss, stable, and high-density manner. For the industrial chain, this is a round of value redistribution for power equipment. The companies that truly benefit may be those that can upgrade from a single component supplier to a high-voltage direct current system platform.