As compute requirements grow in AI datacenters, so do the power requirements, which are estimated to reach 17x higher with NVIDIA's Feynman.
NVIDIA Feynman Racks Estimated To Feature 17x Higher Power Semi Costs Per Rack Versus Blackwell
NVIDIA Feynman GPUs feature several groundbreaking features and will launch in 2028, after Rubin. The company has been working hard to deliver more efficient AI solutions, but as requirements grow, power requirements have increased tremendously.
Morgan Stanley Research has published a chart that visualizes the total power semi content of three AI rack solutions from NVIDIA.
Starting with the baseline Blackwell or B200, the total power semi content is estimated around $11,234 US with GB200 adding roughly $4000 US to the cost, and GB300 adding a further $3500. The whole Blackwell generation scales up to $17,761 in just power semiconductor costs, but as NVIDIA racks evolve with future chips such as Rubin and Feynman, the power cost alone is going to see major upticks.
With Rubin, which launches later this year, the power semi cost is estimated to go past $33,000 US, a 3x increase over Blackwell GB200. The NVIDIA Rubin Ultra racks are going to feature 3x the power systems cost versus Rubin, estimated at around $95,000 US.
Feynman racks will double the power semi content of Rubin Ultra, jumping to an astonishing $191,000 US+ mark. This is a 17x increase over Blackwell and shows the scale of the power content alone for the Feynman generation of AI-focused racks.
Breaking down the figures, the PCS (Power Conversion System) and VRM (Voltage Regulation Module - VPD/SiVR) - 2nd stages take up the bulk of the semi content, amounting to 27% and 26% share, respectively.
These are followed by the PSU that delivers power to the rack, making 19% of the share. The lateral VRMs make 15% of the share, while IBC (1st Stage Intermediate Bus Converter), and BBU (Battery Backup Unit)/UPS (Uninterrupted Power Supply) will take up 4-5% of the pie. The rest of the single-digit shares are taken up by Switches, NICs, and eFuses.
NVIDIA has already announced its move to 800 VDC architectures for future AI datacenters, which will replace the legacy 48V/54V standards, eliminating bottlenecks, reducing current, copper use, and cable bulk, while offering safer and scalable infrastructure designs. 800VDC systems are compact and optimal for next-generation power distribution demands, which decrease conversion and routing volumes and also minimize distribution losses.
Bottlenecks encountered in existing designs include:
- Space constraints: Today’s NVIDIA GB200 NVL72 or NVIDIA GB300 NVL72 feature up to eight power shelves to power the MGX compute and switch shelves. Using the same 54 VDC power distribution would mean power shelves would consume up to 64 U of rack space for Kyber at MW scale, leaving no room for compute. At GTC 2025, NVIDIA exhibited an 800 V sidecar to power 576 of the Rubin Ultra GPUs in a single Kyber rack. The alternative approach is to use a dedicated rack of power supplies for every compute rack.
- Copper overload: The physics of using 54 VDC in a single 1 MW rack requires up to 200 kg of copper busbar. The rack busbars alone in a single 1 gigawatt (GW) data center could require up to 200,000 kg of copper. Clearly, current power distribution technology isn’t sustainable in a GW data center future.
- Inefficient conversions: Repeated AC/DC transformations across the power chain are not energy efficient and increase failure points.
The key advantages of 800 VDC systems include:
- High Efficiency & Lower Losses: The shift to 800V VDC decreases power conversion steps (e.g., from 800V directly down to 6V for chips), minimizing energy loss.
- Reduced Infrastructure Footprint: Lower current allows for thinner, lighter cabling and smaller power components, freeing up precious IT rack space for more computing power.
- Enabled by Advanced Power Electronics: The system heavily utilizes Gallium Nitride (GaN) and Silicon Carbide (SiC) semiconductors, which allow for efficient, high-voltage switching.
- Data Center Applications: AI factories use this architecture to deliver power to racks hosting hundreds of GPUs, supporting megawatt-level densities, according to 2026 industry standards.
- Safety & Stability: While operating at a higher voltage, 800V DC architectures include specialized components like solid-state relays, high-voltage hot-swaps, and isolated sensors to maintain safety
800VDC will first be introduced in NVIDIA's Kyber racks, which are expected in 2027, and will rock the Rubin Ultra AI GPU family in a dense rack configuration with 576 Rubin Ultra chips, and an all liquid-cooled 600kW solution.
The increased reliance on 800VDC architectures and the massive increase in power components will allow a radical response from VRM makers and power providers who will scale up their production to meet the increasing demands for next-gen data centers.
About the author: A Software Engineer by training and a PC enthusiast by passion, Hassan Mujtaba serves as Wccftech's Senior Editor for hardware section. With years of experience in the industry, he specializes in deep-dive technical analysis of next-generation CPU and GPU architectures, motherboards, and cooling solutions. His work involves not only breaking news on upcoming technologies but also extensive hands-on reviews and benchmarking.
Follow Wccftech on Google to get more of our news coverage in your feeds.
