More than 80 manufacturers are already building to the same standard: 800 volts direct current. The bottleneck is the cable to the chip.
Fact
The facts: on 11 August 2026 NVIDIA published an explanation of the 800 VDC architecture for AI factories. The argument: with traditional power delivery, electricity arrives as alternating current and is converted several times, each conversion losing something; at 800 volts direct current there are fewer conversions and more of the power reaches the accelerator. NVIDIA, Google and Microsoft develop it together through the Open Compute Project - a joint white paper in March 2026 and the LVDC Solid-State Transformer Specification v0.3 in July 2026; according to NVIDIA more than 80 equipment manufacturers are already building products to it. The path: an MGX-compatible 800 VDC rack for existing AC buildings in the second half of 2026, row-level power centres up to 2 megawatts in 2027, and a block that converts from the grid straight to 800 VDC - for new buildings.
This is the quiet story of the week and, to me, the more important one. Models change every three months. A building's electrical installation - once. Whoever sets the cable standard now will have it for twenty years.
The system, not the model. This time literally - the wires.
And look who is in the room: NVIDIA, Google, Microsoft. Not out of love, but because nobody wants to pay alone for transformers only they use. An open standard with 80 suppliers = a supply chain, not a whim.
For us, the small ones: we will not build a factory. But next time someone tells you AI is 'just software' - show them this article. There is copper and there are substations.
Official primary sources
NVIDIA Blog - Why Scaling AI Compute Performance Requires a New Power ArchitectureThe visual is generated code-art, no third-party images.