TL;DR
Over the past few years, GPUs have been the cornerstone of AI infrastructure—those who secured more H100s and B200s had greater computing power. But with the advent of Rubin and subsequent platforms, investors need to look one step further: Can the GPUs fit into racks? Can the racks receive stable power? Can the heat be effectively dissipated? And can full-load testing be performed before shipment?
800VDC is now being discussed in the market, and here’s why. Since 2025, NVIDIA has consistently advocated for the 800VDC architecture, integrating it into the design direction of next-generation AI factories and high-power cabinets. On the surface, this represents a shift from traditional low voltage to high-voltage DC. Dig deeper, and AI servers are no longer merely assemblies of circuit boards and chips—they are increasingly becoming a form of power engineering.
Regular investors can think of it this way: an entire rack of AI servers is like a small building with extremely high power consumption. Past power supply methods could support cabinets drawing tens or even hundreds of kilowatts, but as power demands continue to rise, the issue is no longer just “how many GPUs to buy”—it’s about how to deliver the electricity, how to dissipate the heat, and how to ensure the machines can run at full load continuously before they even leave the factory.
High-power cabinets are approaching the low-voltage power supply limit.
The starting point of this shift is the rapid increase in power consumption of AI cabinets. Traditional server cabinets typically consumed only a few kilowatts to around ten kilowatts, but with NVIDIA’s GB200 and GB300 generations, the power consumption per cabinet has reached the hundred-kilowatt level. According to Tom’s Hardware, the GB200/GB300 NVL72 configuration consumes approximately 120–140 kW.
After Rubin, power density may continue to rise. Some supply chain and industry estimates suggest that the Rubin NVL72 could reach approximately 180–220 kW. This range is not an official figure from NVIDIA and should still be regarded as a third-party estimate, but the direction is clear: cutting-edge AI racks are becoming higher-density power units.
Electrical issues can be explained by a formula: power equals voltage multiplied by current. To transmit the same 600 kW of power, if the voltage is lower, a larger current must be used. The greater the current, the thicker the cables and busbars need to be, the more heat is generated, and the higher the energy loss in the transmission lines.
Traditional low-voltage power delivery is like slowly supplying a large volume of water through a very thick pipe. It works, but the pipe keeps getting thicker, heavier, and taking up more space. The rack space, which should be reserved for GPUs, memory, networking, and cooling systems, is instead consumed by power racks, cables, and busbars. At power levels of hundreds of kilowatts or higher, continuing to rely on low-voltage, high-current solutions becomes increasingly uneconomical.
The concept of 800VDC is to increase voltage, delivering power more efficiently to the vicinity of the racks, then stepping it down locally for GPU use. It’s like increasing water pressure to deliver the same volume of water through narrower pipes. According to NVIDIA’s official materials, 800VDC can reduce current, copper usage, cable volume, and conversion stages, improving efficiency by up to 5% and lowering TCO by up to 30%. Some third-party and partner analyses also suggest copper usage may decrease by approximately 45%, with actual benefits depending on the data center and rack integration design.
This isn't just about saving copper. For NVIDIA, the core value of 800VDC is enabling the next generation of AI cabinets to further increase computational density. Low-voltage systems aren't unusable, but in the highest-density AI factories, they begin to approach engineering limits.
NVIDIA redraws infrastructure responsibilities with a reference architecture
The significance of NVIDIA lies not just in proposing a voltage solution, but in redefining ecosystem roles through a reference architecture. Since 2025, NVIDIA has repeatedly publicly introduced the 800VDC architecture and demonstrated design directions for high-power systems such as Rubin and Kyber rack on technical blogs and at OCP events.
According to NVIDIA’s official blog, its 800VDC ecosystem partners include Delta Electronics, Schneider Electric, Vertiv, Infineon, STMicroelectronics, as well as ABB, Eaton, GE Vernova, Hitachi Energy, Siemens, Navitas, and Texas Instruments. A more accurate description is ecosystem collaboration and compatibility—not confirmation that orders have been finalized.
The 800VDC transition involves a comprehensive chain of changes, spanning from data center power distribution and rack power supplies to backup batteries, power components, connectors, and full-rack integration. In the past, power conversion was distributed across multiple stages, such as UPS systems, PDUs, server power supplies, and motherboard power delivery. Under a high-voltage DC architecture, power is delivered closer to the rack, where it is then stepped down by modules located within the rack or near the GPUs.
The weights within the value chain also shift accordingly. In the era of traditional servers, investors focused more on GPUs, CPUs, memory, and contract manufacturing of complete systems. In the era of high-power AI cabinets, power racks, busbars, connectors, power semiconductors, liquid cooling systems, and cabinet-level validation capabilities have all become part of delivery capability.
The boundaries must also be clearly defined. 800VDC is more of a reference architecture for cutting-edge, high-density AI factories, not a universal standard that all data centers will immediately adopt. Many existing data centers will continue using traditional AC or hybrid architectures, while new projects will adopt 800VDC in layers based on power density, cost, owner willingness to retrofit, and safety regulations. What’s truly being traded in the market isn’t an immediate full switch to 800V this year, but rather the evolving infrastructure standards for the highest-density AI racks after 2027.
Power, connectivity, liquid cooling, and full-rack testing have been brought to the forefront.
From an investment perspective, the most direct impact of 800VDC is bringing previously back-end infrastructure components into the spotlight.
The first category includes power infrastructure companies such as Vertiv, Schneider Electric, Delta Electronics, and select power equipment manufacturers from South Korea and Taiwan. These companies are not merely selling traditional data center power equipment but are actively participating in the design of power distribution, rack-level power supply, backup battery systems, and high-voltage direct current (HVDC) systems for next-generation AI factories. According to Asia Business Daily, NVIDIA has been engaging with South Korean power equipment firms including LS Electric, HD Hyundai Electric, and Hyosung regarding 800VDC data center infrastructure. While this report is based on industry sources and does not equate to confirmed orders, it illustrates that power equipment providers are increasingly being integrated into the AI factory ecosystem.
The second category consists of power devices, such as SiC/GaN (silicon carbide/gallium nitride), which are next-generation power switches. These are better suited than traditional silicon devices for high-voltage, high-frequency, and high-efficiency applications. Previously discussed primarily in the context of electric vehicles, charging stations, and industrial power supplies, they are now expanding into AI data centers. Companies like Infineon and STMicroelectronics have thus come into investors’ focus. However, the benefits to power semiconductors depend on specific designs, market share, pricing, and yield—they cannot be simplistically equated with “800VDC theme stocks.”
The third category consists of connections and mechanical structures, including copper busbars, motherboards, high-voltage connectors, high-end backplanes, and certain thick-copper, high-layer PCBs. After voltage increases, current decreases, alleviating copper loss pressure; however, requirements for insulation, safety, connection reliability, and structural design become more stringent. Low-end copper materials do not inherently benefit—true value lies in connection and power delivery materials compatible with high-power, high-reliability cabinets.
The fourth category is liquid-cooled and whole-rack ODM. As power requirements increase, cooling is no longer a secondary concern. For servers to operate stably in customer data centers, they must undergo full-rack testing before shipment, including power supply, cooling, network, and GPU under full load stability. Companies such as Dell, Wiwynn, and Wistron, which deliver complete racks, compete not only on assembly efficiency but also on their capacity for sufficient power, space, liquid cooling testing, and system tuning.
The design direction is clear, but delivery capability still needs to be tested.
NVIDIA has outlined a clear technical direction, but supply chain execution won't automatically run smoothly. This tension is precisely what investors need to monitor.
Independent supply chain analyst Dan Nystedt has recently cited multiple reports from Taiwanese media and industry sources: AI server ODM revenues are strong, and production preparations for Rubin are progressing, but component availability, power infrastructure, and full-rack burn-in testing (full-load aging tests) are becoming practical constraints. Burn-in refers to stress testing performed on servers before they leave the factory, during which GPUs run at full load for extended periods, requiring simultaneous validation of power delivery, cooling, and system stability.
If a single cabinet requires continuous power in the 100–200 kW range, the testing facility itself must possess electrical and cooling capabilities approaching those of a small data center. This supply chain signal should not be interpreted as implying that the industry has universally adopted on-site power generation, nor should it be directly extrapolated to mean that power supply has replaced GPUs as the primary bottleneck. Instead, it serves as a reminder: delivery in the Rubin era means more than just GPU arrival and motherboard assembly—it requires the simultaneous readiness of power, liquid cooling, testing, and overall cabinet stability.
Some ODMs, power equipment, and liquid cooling companies have been revalued for the same reason: their value stems not just from participating in AI server production, but from their ability to reliably deliver high-power cabinets to cloud providers. In the future, if all companies receive the same NVIDIA reference design, the real differentiators may be testing facilities, power capacity, liquid cooling calibration expertise, and delivery yield.
For AI cloud providers like CoreWeave and Nebius, 800VDC is not a direct component beneficiary logic, but rather a variable affecting capital expenditure efficiency and deployment speed. Whether high-density racks can be deployed on time impacts compute delivery, depreciation timing, and revenue realization. Companies such as Marvell and Lumentum in the high-speed interconnect or optical module value chain are more aligned with the parallel logic of AI cluster expansion and should not be conflated with direct beneficiaries of 800VDC.
Looking at Kyber Cabinet and order fulfillment in 2027
The direction of 800VDC is now much clearer than it was a year ago: actively driven by NVIDIA, supported by partner ecosystem adaptation, constrained by physical limitations, and required by cutting-edge high-density AI factories for more efficient power delivery. However, it is still in the preparation and early deployment phase. NVIDIA has stated that full-scale production of 800VDC will align with the 2027 Kyber rack-scale systems; true validation will depend on whether subsequent products and customer projects can be successfully implemented.
What’s next worth watching isn’t whether any company mentions “AI power” in its announcement, but whether it has clearly entered the 800VDC product development, customer validation, and order fulfillment stages. Whether ODMs disclose enhanced full-rack testing capabilities, the reliability of liquid cooling systems under prolonged full load, and whether data center operators are willing to retrofit their power distribution and safety standards for high-voltage DC architectures will all influence the pace of this transaction.
If the Rubin-related cabinets successfully scale up, and 800VDC component orders progress from samples and validation to bulk procurement, the market will continue to increase the weighting of power supply, liquid cooling, connectors, and full-cabinet delivery capabilities. Conversely, if power consumption configurations fall below expectations, customers adopt more conservative hybrid architectures, or testing of power and liquid cooling reliability delays deliveries, 800VDC transactions will revert from directional assumptions back to validation of orders and timing. GPUs remain central, but after Rubin, the ability to consistently deliver entire cabinets of high-power systems is beginning to emerge as an asset pricing variable.