NVIDIA's upcoming Feynman GPU generation will require $191,000 in power semiconductor components per rack, a staggering 17-fold increase from Blackwell's $11,234. This isn't just a technical curiosity; it signals a fundamental reshaping of how AI data centers are built and powered. As compute demands explode, the infrastructure supporting those chips has become the real bottleneck.

Why Power Semiconductors Are Becoming the Hidden Cost of AI?

When most people think about AI hardware, they imagine the GPU itself, the processor doing the heavy lifting. But inside a modern data center rack, the GPU is only part of the story. The power conversion systems, voltage regulators, and distribution components that feed electricity to those chips are equally critical, and they're becoming increasingly expensive.

The progression tells the story. Blackwell's B200 chips required roughly $11,234 in power semiconductor content per rack. The GB200 variant added about $4,000 to that cost, and GB300 added another $3,500. But when NVIDIA moves to its Rubin generation, expected later in 2026, the power semiconductor costs jump to over $33,000 per rack, a threefold increase. Rubin Ultra racks will feature $95,000 in power system costs. Then comes Feynman in 2028, doubling Rubin Ultra's power requirements to $191,000.

The breakdown reveals where this money goes. Power Conversion Systems (PCS) and Voltage Regulation Modules (VRM) account for 27% and 26% of the total power semiconductor content, respectively. The Power Supply Units (PSU) that deliver electricity to the entire rack make up 19%. Lateral VRMs contribute 15%, while Intermediate Bus Converters and Battery Backup Units split the remaining single-digit percentages.

What's Driving This Exponential Growth in Power Costs?

The root cause is simple physics. As GPUs become more powerful, they consume more electricity. A single Feynman-era rack will demand megawatt-level power densities that today's infrastructure simply cannot handle efficiently. The current standard of 48V to 54V direct current (DC) power distribution, which has worked for decades, hits hard limits when you're trying to power hundreds of GPUs in a single cabinet.

Consider the practical constraints. A modern NVIDIA GB200 NVL72 or GB300 NVL72 rack requires up to eight separate power shelves just to distribute electricity to the compute equipment. If data centers tried to scale this approach to gigawatt-scale facilities, the copper cabling alone would become prohibitive. A single 1-megawatt rack using 54V DC distribution requires up to 200 kilograms of copper busbar. Scale that to a 1-gigawatt data center, and you're looking at 200,000 kilograms of copper just for the power distribution infrastructure.

Beyond the material costs, the repeated conversions of power from alternating current (AC) to direct current (DC) at different voltage levels waste energy and create failure points throughout the system. Each conversion step introduces losses, reducing overall efficiency and increasing operational costs.

How 800V Power Architectures Are Solving the Problem?

NVIDIA's answer is a radical departure from decades of industry practice: moving to 800V DC power distribution. Instead of stepping down voltage multiple times as power flows through the rack, 800V systems deliver power at a much higher voltage, then step down directly to the voltage levels GPUs need. This single architectural change solves multiple problems simultaneously.

• Space Efficiency: Higher voltage means lower current flowing through cables and components, allowing for thinner, lighter cabling and smaller power components. This frees up precious rack space for additional computing hardware instead of power infrastructure.
• Copper Reduction: Lower current requirements dramatically reduce the amount of copper needed for power distribution. Instead of 200 kilograms per megawatt rack, 800V systems require a fraction of that material.
• Energy Efficiency: Fewer conversion steps mean fewer losses. Power flows more directly from the source to the GPU, minimizing waste and reducing the total electricity needed to deliver the same compute performance.
• Scalability: 800V architectures are designed for gigawatt-scale deployments. They enable the kind of power densities that future AI factories will require without hitting physical or economic limits.

The technology enabling this shift relies on advanced semiconductor materials. Gallium Nitride (GaN) and Silicon Carbide (SiC) semiconductors can handle high-voltage switching far more efficiently than traditional silicon. These materials allow power conversion systems to operate at 800V while maintaining safety and reliability.

When Will 800V Power Systems Actually Arrive in Data Centers?

NVIDIA is not waiting for Feynman in 2028 to introduce this technology. The company will debut 800V DC architectures in its Kyber racks, expected in 2027. These racks will house the Rubin Ultra GPU family in a dense configuration, packing 576 Rubin Ultra chips into a single cabinet with an all-liquid-cooled 600-kilowatt power delivery system.

At GTC 2025, NVIDIA demonstrated an 800V sidecar power system that could supply 576 Rubin Ultra GPUs in a single Kyber rack, proving the concept works at scale. This is not theoretical; it's production-ready technology arriving within the next 18 months.

The shift to 800V will ripple through the entire semiconductor supply chain. Power semiconductor manufacturers, voltage regulator makers, and power supply vendors will need to scale production dramatically to meet demand. Companies specializing in GaN and SiC components will see their addressable market expand significantly as 800V becomes the industry standard for AI data centers.

Why This Matters Beyond the Data Center?

The power semiconductor cost explosion reveals something important about the future of AI infrastructure. Raw compute performance, measured in floating-point operations per second (FLOPS), is only half the equation. The ability to deliver power to that compute, and to do so efficiently, is equally critical. As AI models grow larger and inference demands increase, the infrastructure supporting those models becomes the limiting factor.

This has real economic implications. A company deploying AI at scale must now budget not just for the GPUs themselves, but for the power distribution infrastructure that makes those GPUs usable. The $191,000 in power semiconductor costs for a Feynman rack represents a significant portion of the total rack cost, and it's growing faster than compute costs themselves.

For enterprises considering on-premise AI deployments, understanding these power infrastructure requirements is essential. A local AI system requires not just the GPU hardware, but also the power delivery systems, cooling infrastructure, and electrical capacity to support it. The shift to 800V architectures will eventually make this more feasible, but the transition period will require careful planning and investment.

The broader lesson is that AI infrastructure is becoming increasingly specialized. The days of treating AI hardware as a simple upgrade to existing data center equipment are ending. Future AI factories will be purpose-built systems, optimized for power delivery, cooling, and compute density in ways that generic data center infrastructure cannot match.