NVIDIA Updates 800VDC White Paper: From Technical Validation to Execution

<p>The rapid expansion of AI data centers is driving computing infrastructure beyond chips, servers, and networks, further into the power system. On August 11, NVIDIA officially released Version 2 of its white paper, "<em>800 VDC Architecture: Industry Alignment &amp; Execution.“</em></p><p>The first edition, released in October 2025, mainly demonstrated the technical necessity and feasibility of adopting 800V high-voltage DC power supply in AI data centers. This update focuses further on system architecture, equipment interfaces, deployment methods, safety protection, validation systems, and industry collaboration.</p><p>The white paper clearly states that work related to 800VDC has moved from "concept evaluation" to "coordinated execution." Over the next 12 months, the industry will focus on equipment development, system performance validation, pilot deployment, and standards and regulatory work to accelerate the development of 800VDC solutions that can be deployed in practice.</p><h2>AI Racks Are Moving Toward Megawatt-Level Power</h2><p>The growing attention to 800VDC architecture is first related to the rapid increase in AI rack power. As the scale of large language models expands, more and more GPUs need to be connected through NVLink to form large-scale computing domains. To reduce the computing cost and energy consumption of each token, more GPUs need to be deployed within a limited space, driving single-rack power from several hundred kilowatts gradually toward the megawatt level.</p><p>According to the rack power roadmap disclosed in the white paper, Gen 1 power racks support up to approximately 145kW; the Gen 2 architecture for Vera Rubin NVL72 increases single-rack power capability to approximately 330kW; and Gen 3 will adopt fully liquid-cooled power modules, busbars, and connectors, with each rack position supporting up to approximately 570kW.</p><p>Going further, NVIDIA plans to transition to a native 800VDC rack architecture in Gen 4, providing a more scalable power supply approach for computing racks approaching the megawatt level.</p><p>At present, NVIDIA's existing and near-term computing systems will still convert 800VDC to 54VDC inside the rack to remain compatible with mature power architectures. However, as rack power continues to increase, the current required for low-voltage power delivery, as well as copper, connectors, and thermal management space, will all increase significantly.</p><p>Compared with 54VDC, 800VDC can reduce current when delivering the same amount of power, reduce the space occupied by the power supply system, and eliminate some power conversion stages. Therefore, NVIDIA regards it as an important technology route for the continued expansion of next-generation high-density AI racks.</p><h2>800VDC Will Not Immediately Replace Existing AC Systems</h2><p>The white paper also emphasizes that 800VDC is not intended to immediately replace existing 415V or 480V AC power supply systems, but rather to serve as an additional deployment option that will coexist with AC architectures over the long term.</p><p>Currently, a large number of existing and under-construction data centers still use mature AC distribution systems. If 800VDC can only be applied to newly built data centers, the speed of technology adoption will be limited, and it could also result in wasted investments in existing land, power connections, and distribution equipment.</p><p>Therefore, NVIDIA proposes several optional deployment approaches in the white paper. Operators can choose to introduce 800VDC at the rack level, cluster level, or data hall level according to existing facility conditions, construction schedules, computing density, and investment plans, without having to upgrade the system step by step in a fixed sequence.</p><h2>Three 800VDC Deployment Architectures</h2><h3>▍<strong>Option A: Rack-Level Power Rack</strong></h3><p>Option A involves configuring a dedicated Power Rack next to the computing racks. The existing data center continues to receive power through a 415V to 480V AC system, while the Power Rack performs AC-to-800VDC conversion near the racks and supplies power to adjacent computing racks through standardized DC interfaces.</p><p>The initial Power Rack has a capacity of approximately 660kW and can supply power to one or two computing racks. This approach can make use of existing AC infrastructure without requiring large-scale modifications to the upstream electrical room, making it suitable for early deployment of 800VDC computing equipment in existing data centers, test facilities, and laboratories.</p><p>According to the timeline disclosed in the white paper, the Power Rack is planned to enter production in the third quarter of 2026, making it the 800VDC solution currently closest to practical deployment.</p><p>However, the Power Rack requires additional data center space and also needs to accommodate a relatively large number of AC input cables. As computing scale expands, the number of power racks and their space requirements may become limiting factors for further expansion.</p><h3>▍<strong>Option B: Cluster-Level Power Center</strong></h3><p>Option B concentrates rectification equipment at the end of a row of racks or a computing cluster, forming a unified DC Power Center. Each Power Center can provide up to approximately 2MW of output capacity, which is then distributed to multiple computing racks through 800VDC overhead busbars, tap boxes, or remote distribution panels.</p><p>Compared with the rack-level Power Rack, this solution reduces the number of power supply devices and power cables located next to computing racks. It can also selectively provide 800VDC to high-density computing clusters without changing the upstream AC distribution room.</p><p>According to the white paper, Option B is expected to be deployed as early as the third quarter of 2027. At present, multiple equipment manufacturers are developing and certifying 800VDC rectifiers, DC busbars, tap boxes, connectors, and protection devices.</p><h3>▍<strong>Option C: Data Hall-Level DC Power Block</strong></h3><p>Option C is designed for future ultra-high-density AI factories, further concentrating power conversion and distribution at the data hall level. This architecture uses an approximately 4.8MW DC Power Block as a standardized power unit, supplying 800VDC power to the entire data hall through high-capacity DC switchgear and busbars. Four scalable units can form an approximately 20MW standardized deployment unit, which can then be replicated across a campus to support AI factories ranging from hundreds of megawatts to gigawatt scale.</p><p>The DC Power Block can also adopt a modular and outdoor containerized design, allowing it to be assembled and tested at the factory before being transported to the project site, thereby reducing on-site construction work, equipment room footprint, and commissioning complexity. In the longer-term architecture, NVIDIA hopes to directly convert the 34.5kV medium-voltage AC commonly used in data center power distribution into 800VDC, further reducing intermediate low-voltage conversion stages. In the future, a single power block may scale to 10MW or more, while the next-generation solid-state transformer technology roadmap points toward around 2029.</p><p>However, the white paper also points out that most commercially available solid-state transformers currently mainly target approximately 15kV inputs. To adapt to 34.5kV data center distribution systems, issues including insulation coordination, fault protection, thermal management, and grid connection still need to be addressed. Therefore, this part remains a medium- to long-term development direction.</p><p><img src="https://energyplanet.oss-cn-shenzhen.aliyuncs.com/uploads/images/20260814162023_94e5393e.png" alt="" loading="lazy" decoding="async" /></p><h2>Energy Storage Enters Different Layers of AI Data Centers</h2><p>In this white paper, energy storage not only serves the traditional backup power function, but is also incorporated into rack-level load management, data center power regulation, and DC microgrid architectures.</p><h3>▍<strong>First is short-duration energy storage at the rack level</strong></h3><p>AI computing loads are characterized by rapid changes and a high degree of synchronization. When large numbers of GPUs start simultaneously or adjust computing tasks, rack power may change significantly over a short period of time. The white paper proposes that small-capacity, fast-response energy storage devices, such as electrolytic capacitors, can be deployed inside or near computing racks to reduce instantaneous power peaks, smooth loads, and control the rate of power changes.</p><p>In the initial Vera Rubin Power Rack design, each power module is also equipped with 1,200J of energy storage to improve the AC input current under dynamic loads. The Power Rack can also optionally be equipped with a BBU consisting of four 20kW modules, providing a total of 80kW of power for approximately 60 seconds to provide short-duration power support for the rack.</p><h3>▍<strong>Second is large-scale battery energy storage systems at the data center level</strong></h3><p>In the data hall-level DC Power Block architecture, energy storage, DC UPS systems, renewable energy, and behind-the-meter generation facilities can be integrated with the 800VDC system to further form a DC microgrid. Since batteries themselves output DC power, directly connecting them to the DC bus has the potential to reduce repeated AC-DC conversion and synchronization stages. In the future, large-scale BESS can undertake load smoothing, grid interconnection support, on-site generation coordination, and enhanced power resilience.</p><p>However, the white paper does not disclose the specific capacity, storage duration, battery technology route, or suppliers for large-scale energy storage, nor does it announce any specific energy storage procurement plans. What can currently be confirmed is that energy storage has been incorporated into NVIDIA's future AI factory power architecture, while the configuration of specific projects will still need to be further determined according to data center and grid conditions.</p><h2>Safety Protection and Certification Remain Key to Deployment</h2><p>Compared with mature AC systems, 800VDC has different technical requirements for fault protection. AC current has periodic zero-crossing points, making arc interruption relatively easier; DC current has no natural zero-crossing point, meaning that arcs may persist. Therefore, 800VDC systems need to place greater emphasis on fault current limitation, DC arc detection, grounding methods, insulation monitoring, connector interlocking, and rapid fault isolation.</p><p>The white paper lists several grounding options, including high-resistance midpoint grounding, high-resistance circuit grounding, floating grounding, and direct grounding. However, no single grounding method has yet been determined, and system-level evaluation is still required based on personnel safety, fault detection, equipment supply, and operational continuity.</p><p>In terms of protection equipment, molded-case circuit breakers with more mature supply chains and certification systems may be adopted first in the short term. In the long term, solid-state circuit breakers can interrupt faults at the sub-millisecond level, reducing peak fault current and arc energy, and are therefore regarded as important protection devices for high-density 800VDC systems.</p><p>NVIDIA is currently focusing on two types of products: 125A air-cooled solid-state circuit breakers and 1,250A solid-state circuit breakers, targeting rack-level internal power modules and future native 800VDC rack interfaces, respectively.</p><p>At the same time, NVIDIA is working with cloud service providers, equipment companies, and organizations such as UL Solutions to advance requirements for power quality, dynamic response, grounding protection, equipment certification, and system stability, while coordinating with standards systems including UL, NFPA, IEEE, and IEC.</p><h2>From Architectural Consensus to Industry Execution</h2><p>Compared with the first edition of the white paper, the most important change in this update is that the focus of the 800VDC discussion has shifted from "why it is needed” to “how to deploy it."</p><p>At present, the relevant architectures are gradually converging toward several major forms at the rack, cluster, and data hall levels, allowing equipment companies to develop products around relatively standardized power levels, connection interfaces, busbar systems, and protection requirements.</p><p>Over the next 12 months, the industry will continue to establish system-level performance requirements and validate the system from medium-voltage AC input all the way to the GPU chip side. Through hardware testing, system integration, and digital simulation, the industry will evaluate the compatibility, stability, and reliability of equipment from different manufacturers. At the same time, pilot deployment, equipment certification, installation and commissioning, and operation and maintenance standards will also be advanced simultaneously.</p><p>It should be noted that some of the timelines and products mentioned in the white paper are still part of the roadmap and do not mean that commercial deployment has already been completed. However, from the overall direction, 800VDC is no longer merely a conceptual architecture for AI data centers, but is beginning to enter the stages of equipment development, system validation, and industry collaboration.</p><p>As AI racks move toward higher power density, the data center power system will gradually shift from a traditional supporting facility to a core infrastructure component that affects computing scale, construction speed, and operational efficiency. The changes brought by 800VDC will also extend to multiple areas, including rectifiers, DC busbars, connectors, circuit breakers, solid-state transformers, energy storage, and energy management systems.</p><p>Future AI data centers will need not only larger battery capacity, but also faster power response, DC-side integration capabilities, fault current control, and the ability to operate in coordination with computing loads and on-site power sources. From this perspective, NVIDIA's latest white paper is not only an update on 800V power supply technology, but also an execution roadmap for the power infrastructure of next-generation AI factories.</p><p><br /></p><p><a href=" https://energyplanet.global/api/v1/community/posts/77/share-page?v=20260721-content-preview-v2 " target="_blank" rel="noopener noreferrer">《800 VDC Architecture: Industry Alignment &amp; Execution》link</a></p>