What Is DC

On August 12, NVIDIA is collaborating with Google, Microsoft, and over 80 device and infrastructure suppliers to promote the 800V DC power architecture. This initiative aims to address the power efficiency and conversion loss issues arising from the increasing power density of AI server racks, with plans to gradually introduce this architecture to existing AI factories starting in the second half of 2026. NVIDIA stated that the current bottleneck in AI infrastructure is no longer just the total power wattage, but the efficiency of power delivery from the grid to the GPU. The 800V DC architecture can reduce the power conversion steps between the grid and the accelerators. The company plans to launch 800V DC power racks compatible with the MGX architecture in the second half of 2026, allowing existing data centers to upgrade gradually without significant modifications to their buildings and power systems. For new or high-density AI factories, NVIDIA plans to introduce rack-row power centers in 2027, supplying power to multiple rows of racks via overhead 800V DC busbars, with each row capable of supporting up to 2MW of power. In the long term, there are plans to deploy DC Power Blocks at the facility level to convert grid power directly into 800V DC. NVIDIA, Google, and Microsoft are also promoting related standardization through the Open Compute Project. As AI servers evolve from hundreds of kilowatts to megawatts, specifications for equipment such as high-voltage DC/DC converters, solid-state transformers, busbars, power semiconductors, and power modules will also be upgraded.

On August 12, NVIDIA announced that it is collaborating with Google, Microsoft, and over 80 equipment and infrastructure suppliers to promote the 800V DC power architecture. This initiative aims to address the efficiency and conversion loss issues arising from the increasing power density of AI server racks, with plans to gradually introduce this architecture to existing AI factories starting in the second half of 2026. NVIDIA stated that the current bottleneck in AI infrastructure is not just the total power wattage, but the efficiency of power delivery from the grid to the GPUs. The 800V DC architecture can reduce the power conversion stages between the grid and the accelerators. The company plans to launch 800V DC power racks compatible with the MGX architecture in the second half of 2026, allowing existing data centers to upgrade gradually without major renovations to buildings and power systems. For new or high-density AI factories, NVIDIA plans to introduce rack-mounted power centers in 2027, supplying power to multiple racks via overhead 800V DC busbars, with each rack supporting up to 2MW of power. Long-term plans include deploying DC Power Blocks at the facility level to directly convert grid power into 800V DC. NVIDIA, Google, and Microsoft are also promoting related standardization through the Open Compute Project. As AI servers evolve from hundreds of kilowatts to megawatts, specifications for high-voltage DC/DC converters, solid-state transformers, busbars, power semiconductors, and power modules will also be upgraded.

On June 17, Serenity shared Mizuho Securities' latest research report, commenting that the recent concentrated sell-off of CPO concept stocks is entirely a misguided market reaction. The Mizuho report points out that the market's concerns regarding delays in the mass production of Co-Packaged Optics (CPO) and the lag in 800V DC deployment are unfounded, as both technological pathways are progressing as scheduled, with mass production expected to commence between 2028 and 2029. Driven by demand for Nvidia's CPO switch (3.2T Spectrum-X) anticipated to ramp up in 2027, Mizuho has also raised its shipment forecasts for Optical Engines (OE). In terms of technology, Mizuho believes that InP DFB lasers will remain the long-term mainstream direction for CPO, specifically naming $SIVE as a related beneficiary. VCSEL and microLED have not yet been sufficiently validated in short-distance interconnections within racks and in chip-to-chip scenarios exceeding 1.6T, making them unlikely to become mainstream alternatives in the short term. Regarding the deployment of 800V DC data centers, incremental shipments are expected to begin in 2027, with penetration rates further increasing in 2028. Nvidia's Senior Vice President of Networking had previously publicly denied any delays in the Spectrum-X Ethernet CPO switch, confirming that mass production and customer deliveries will proceed as scheduled in the second half of 2026.