Flex Connections: Material Selection in Nickel vs. Copper Tabs
Flex Connections: Four Lessons in Engineering Judgment for PCB Designers
Flexible Thinking: ‘Let There Be Light’–Emitting Diodes!
TrendForce's latest research on power architectures for AI servers highlights that NVIDIA has been developing its proprietary 800V HVDC Power Rack. The platform is expected to be ready for customer shipments in 3Q26 and will be offered as an optional configuration—rather than a standard feature—for customers deploying the Vera Rubin platform.
TrendForce anticipates that, given current rack-level power needs, adoption of 800V Power Racks will begin to increase after the Rubin Ultra generation launches in the second half of 2027. Widespread deployment is likely to follow in 2028.
NVIDIA's current VR200 rack consumes approximately 225 kW, up from around 150 kW for the previous-generation GB300. Despite the increase, the higher power demand remains within the capabilities of conventional in-rack power supply units (PSUs).
The transition to the Rubin Ultra generation, however, marks a significant shift in power architecture. Rack power consumption is expected to surge to approximately 660 kW, while next-generation air-cooled systems may require 1.2 MW to 1.3 MW of power. TrendForce estimates that a single 800V Power Rack could support one to two Rubin Ultra racks, although the actual deployment ratio will depend on each customer’s redundancy requirements. Supply chain sources also indicate that NVIDIA is developing additional power delivery architecture, providing customers greater flexibility in system deployment.
Hyperscaler data center deployment will continue to depend on grid connections and the availability of critical power infrastructure
Hyperscalers and data center developers continue to invest aggressively in new AI infrastructure, with multiple gigawatt-scale campuses across North America expected to enter operation before the end of 2026. However, actual deployment schedules will depend on several supply-side constraints, including ongoing shortages of key server components such as memory and CPUs, as well as the ability of local power grids to accommodate rapidly growing electricity demand.
TrendForce observes that while most U.S. power markets possess sufficient generation capacity to support approved data center projects, the primary bottleneck lies in electricity transmission. This issue is particularly acute in regions served by PJM Interconnection, where grid interconnection requirements are more stringent. Some grid connection queues can now exceed five years for newly proposed data centers.
In response, PJM and several other regional transmission organizations are working with the U.S. federal government to accelerate grid interconnection processes for large-scale data center projects.
Long lead times for critical electrical infrastructure are also adding further uncertainty. As of mid-2025, the average lead time for large power transformers had extended to approximately 2.5 years, while transformers in the 345 kV to 765 kV voltage class require four to five years—roughly double the lead times seen in 2020. Switchgear is experiencing similar supply constraints.
Against this backdrop, TrendForce believes the industry should closely monitor whether suppliers can accelerate production and delivery of these critical electrical components and supporting power infrastructure, as any delays could postpone data center construction schedules.
