Bid farewell to the 54V era and move towards 800V. A power revolution is sweeping through data centers.
As artificial intelligence applications such as ChatGPT, Claude, and DeepSeek sweep across the globe, the power demand of major ultra-large-scale AI data centers worldwide has become increasingly huge. With the exponential growth of artificial intelligence workloads, data centers are being pushed to the critical point of power demand.
The power consumption of global AI data centers is rapidly jumping from the traditional 20 - 30 kW per rack to the level of 500 kW or even 1 MW. The power of a single NVIDIA AI GPU server approaches 1 kW, and the power of a fully equipped NVL AI server cabinet easily exceeds 100 kW. The planned 1 MW AI Factory rack cluster to be mass-produced in 2027 poses disruptive requirements for the power supply system.
As the computing power density of high-performance computing continues to rise, the power supply architecture of data centers is accelerating towards the 800V DC (or ±400V) HVDC high-voltage system. The industry generally believes that the 800V architecture can significantly reduce the energy loss in the power supply and distribution network, improve the overall energy efficiency, and provide technical support for the large-scale deployment of megawatt-level cabinets.
01
Traditional Rack Power System VS 800V HVDC
As the AI workload grows exponentially, the power demand of data centers also surges. The traditional 54V in-rack power distribution system is designed for kilowatt-level racks and can no longer meet the power supply requirements of megawatt-level racks in modern AI factories.
Currently, the racks in AI factories generally rely on 54V DC power supply, transmitting power from the rack-mounted power supply rack to the computing tray through thick copper busbars. When the rack power exceeds 200 kilowatts, this power supply method gradually faces physical limitations:
Space limitation: Taking the equipment equipped with NVIDIA GB200 NVL72 or GB300 NVL72 as an example, up to 8 power supply racks are required to supply power to the MGX computing and switch racks. If the 54V DC power distribution is still used, under the megawatt-level power demand, the Kyber power supply rack will occupy up to 64U of rack space, leaving no installation space for computing equipment. At the 2025 GTC conference, NVIDIA demonstrated an 800V sidecar solution that can supply power to 576 Rubin Ultra GPUs in a single Kyber rack; another alternative is to configure a dedicated power supply rack for each computer rack.
Copper cable overload: Using 54V DC power supply in a 1-megawatt single rack requires up to 200 kilograms of copper busbars. In a 1-gigawatt data center, the amount of copper required for the rack busbars will be as high as 500,000 tons. Obviously, this power distribution technology is difficult to sustain in future gigawatt-level data centers.
Inefficient conversion: The repeated AC-DC conversion in the power transmission chain not only consumes energy but also increases the risk of failures.
The power distribution system of traditional data centers is prone to low efficiency and increased complexity of the electrical system due to multiple voltage conversions. By using industrial-grade rectifiers to directly convert the 13.8kV AC grid power into 800V high-voltage DC (HVDC) at the edge of the data center, most of the intermediate conversion links can be eliminated. This simplified solution can minimize the energy loss in the multiple rounds of AC-DC and DC-DC conversion processes.
At the same time, this solution can also significantly reduce the number of power supply units (PSUs) with fans in the power chain. Fewer PSUs and fans can not only improve system reliability and reduce heat dissipation pressure but also improve energy efficiency, making HVDC power distribution an efficient solution for modern data centers while significantly reducing the overall number of components.
02
NVIDIA Leads the Formation of 800V HVDC Alliance to Layout Future AI Data Centers
In May 2025, at the Taipei International Computer Show (COMPUTEX), NVIDIA made a major move - officially announcing the establishment of an 800V high-voltage DC (HVDC) power supply supplier alliance. Its core goal is clear: By 2027, build the next-generation AI data center that can support a single rack with a power of 1 megawatt (MW).
Although high-voltage DC power transmission is not a new concept, in the past decade, limited by problems such as insufficient converter efficiency, imperfect protection mechanisms, and lack of standardized supporting facilities, it has never been able to achieve large-scale promotion in the data center field. However, in recent years, on the one hand, the progress of solid-state power supply technology and the mature development of the electric vehicle industry have provided a solid guarantee for the safety, energy efficiency, and cost control of the 800V system; on the other hand, NVIDIA has joined hands with upstream and downstream partners to build a complete collaborative network covering chips, power supplies, electrical engineering, and data center operations, clearing the obstacles for the implementation of the technology. Therefore, NVIDIA has started to cooperate with partners in the data center energy ecosystem to jointly develop the 800V HVDC architecture.
Facing the rapid growth of AI energy consumption, simply relying on increasing hardware density cannot solve the fundamental problem. NVIDIA's 800V HVDC architecture completely reconstructs the power distribution system, not only breaking through the bottleneck of energy transmission but also opening up new possibilities for the coexistence of high-density AI factories and low total cost of ownership (TCO).
Official data shows that the end-to-end energy efficiency of this architecture can be increased by up to 5%. Due to the reduction of power supply unit (PSU) failures, the labor cost of component maintenance is significantly reduced, and the maintenance cost can be reduced by up to 70%. Moreover, there is no need to configure AC/DC power supply units in the IT racks, thus significantly reducing the heat dissipation-related costs. This is undoubtedly a cost-effective and necessary technological investment for AI service providers, cloud computing platforms, and even ultra-large-scale data center operators.
In fact, in addition to NVIDIA's layout, Microsoft also released the Mount DrD Low separated power supply architecture in October last year, using 50V DC power supply, and plans to achieve 400V HVDC power supply by replacing the power supply module. Google has proposed a short-term transition plan and an ultimate plan. The ultimate plan involves directly connecting the data center to the power grid and converting it into ±400V DC power through a large rectifier device for full-site power supply to maximize energy efficiency. Meta has released a three-step high-power power supply solution, gradually upgrading to the megawatt-level HVDC solution.
03
Domestic Manufacturers Start Early Layout
Facing this new technological change, domestic manufacturers are also following the industry trend to layout relevant technologies.
It is worth noting that in the list of 800V DC power supply architecture partners updated on NVIDIA's official website on August 1st, Innoscience, as the only Chinese chip company selected this time, has officially reached in-depth cooperation with NVIDIA. The two sides will join hands to promote the large-scale application of the 800V DC (800 VDC) power supply architecture in AI data centers. This cooperation is expected to increase the computing power density of a single computer room by more than 10 times, help the power density of a single cabinet break through 300kW, and promote the global AI data centers to officially enter the megawatt-level power supply era.
Another packaging and testing company, JCET, has also launched a series of solutions.
In the primary power supply unit (PSU) link, JCET demonstrates comprehensive technical coverage: it provides high-power discrete devices based on advanced packaging forms such as TO263-7L, TOLL, and TOLT, as well as industry-leading plastic-encapsulated power modules, all of which are compatible with power devices made of third-generation semiconductor materials such as gallium nitride (GaN) and silicon carbide (SiC). Currently, these discrete devices and plastic-encapsulated modules have achieved stable large-scale mass production. In response to the requirements of the 800V DC architecture, JCET has completed the technical layout and mass production verification in advance and has mature supporting capabilities.
Intermediate bus conversion (IBC), as the core bridge connecting the 800V high voltage and the subsequent 12V/4.8V low voltage output, has strict performance requirements for high power density and extremely low PDN loss, posing a severe challenge to packaging technology. In this field, JCET can provide PDFN packaging with double-sided heat dissipation characteristics and has formed mature packaging and testing solutions for both gallium nitride MOSFETs and silicon-based MOSFETs. At the same time, the company has achieved a breakthrough in multi-layer high-density system-in-package (SiP) technology, and the relevant products have been delivered in batches for first-class server board projects.
In the point-of-load power supply (PoL) link, JCET also has significant advantages. For products such as DrMOS and multi-phase controllers, the company provides mature QFN packaging and a new type of LGA highly integrated packaging solution; relying on its independently developed multi-layer SiP process, it has successfully realized miniaturized power management modules with two-phase to eight-phase multi-channel outputs, with a maximum single-phase current of over 60A. Currently, the team has completed the research and development of a new generation of highly integrated modules, and the SiP interconnection reliability test results are excellent.
Throughout the three subsystems of PSU, IBC, and PoL, facing the board-level application requirements of the 800V large voltage difference, JCET has established a technical pattern of "equal emphasis on discrete and integrated, parallel development of single-chip and modules" in the packaging process, and the mass production rhythm is synchronized with market demand. At the same time, through in-depth cooperation with multiple material, equipment, and system integrators, the company has established a stable collaborative network in the upstream and downstream of the industrial chain and can provide customers with full-process value-added services from thermal simulation, reliability testing to performance optimization.
04
Why Choose GaN?
The core reason why Innoscience can enter NVIDIA's supply chain lies in the tight global supply pattern of gallium nitride (GaN). TSMC announced the closure of its GaN production line last month, and the underlying reason is the scarcity of GaN material supply - even an industry giant like TSMC faces difficulties in obtaining it, so the supply capacity of other manufacturers is naturally more limited. Against this background, Innoscience's selection is, to some extent, an inevitable choice under the market supply pattern.
Looking deeper, as an important member of the third-generation semiconductors, GaN has unique advantages compared with silicon carbide (SiC). In high-voltage scenarios, although there are two technical paths of SiC and GaN to choose from, GaN performs more prominently: traditional power supply equipment generally has the problems of large volume and high energy consumption, while GaN devices can effectively solve this pain point.
From a technical characteristic perspective, the core advantage of gallium nitride devices (i.e., HEMT) stems from their special structure. The basic structure of a GaN HEMT consists of a GaN/AlGaN heterostructure - this special junction formed by two semiconductors with different bandgaps will form a two-dimensional electron gas (2DEG) at the interface. It is precisely this characteristic that enables GaN devices to have high electron mobility and thus achieve low on-resistance; however, precisely because of this, P-GaN HEMTs use a lateral structure, which places extremely high requirements on the accuracy of epitaxial growth and manufacturing processes. Therefore, the IDM (vertical integrated manufacturing) model becomes the optimal solution - full-process self-control from epitaxial growth to packaging and testing can better ensure product performance stability.
Different from silicon-based power semiconductors, GaN transistors form a two-dimensional electron gas for conduction through the piezoelectric effect at the interface between AlGaN and GaN materials. Since the two-dimensional electron gas conducts electricity only through high-concentration electrons, there is no problem of minority carrier recombination (i.e., body diode reverse recovery) in silicon MOSFETs, which makes the high-frequency characteristics of GaN devices particularly excellent.
The advantages brought by the material characteristics are directly reflected in the performance: GaN HEMT power devices have smaller on-resistance and gate charge, and better conduction and switching performance, so they are particularly suitable for high-frequency application scenarios and can significantly improve the efficiency and power density of converters.
The high-frequency characteristics bring three significant advantages:
Reduce transformer volume: According to the law of electromagnetic induction, the higher the frequency, the faster the magnetic flux change rate, and the volume of the required transformer can be significantly reduced, thereby reducing the overall volume and weight of the equipment;
Improve conversion efficiency: During high-frequency operation, the switching loss of the switching device is relatively lower, directly improving the power conversion efficiency;
Reduce filter size: High-frequency signals are easier to process through the filter, which can reduce the size of the filter and further optimize the equipment volume.
For the step-up scenario required by high-voltage DC power supplies, using GaN devices can not only improve conversion efficiency but also reduce the heat generation and volume of the equipment, which is also the key to its core value in the 800V high-voltage DC architecture.
With the simultaneous launch of the Kyber architecture and the 800V HVDC system in 2027, the global AI infrastructure will enter a new power era. From rack design, energy scheduling to load response management, every link will reshape its boundaries in the "clear stream" of high-voltage DC power. In the AI-driven future, computing power is not the only key; energy is the core resource for the next battlefield.
This article is from the WeChat official account "Semiconductor Industry Landscape" (ID: ICViews), author: Pengcheng, published by 36Kr with authorization.