Nvidia, Google, and Microsoft are pushing an 800-volt direct-current power architecture as the new standard for AI factories, with more than 80 equipment manufacturers already building hardware to the specification. The three companies co-developed the design through the Open Compute Project, publishing a joint white paper in March 2026 and the LVDC Solid-State Transformer Specification v0.3 in July 2026. Wood Mackenzie projects $9 trillion in global AI and data infrastructure investment through 2040, and the bet is that most of it won't land in facilities still routing power the traditional way.
The problem Nvidia is targeting isn't total wattage — it's conversion overhead. In conventional data centers, grid alternating current gets stepped down and rectified multiple times before reaching a GPU, and each stage bleeds efficiency. At the rack densities that Blackwell-class and successor systems demand, those losses compound into real capacity limits. Distributing power as 800 VDC collapses the conversion chain, so more of the delivered wattage actually reaches the accelerator.
The rollout has three tiers. The Nvidia MGX-compatible 800 VDC power rack arrives in the second half of 2026 and is designed as a retrofit — it slots into existing AC infrastructure and delivers 800 VDC only within the row, requiring no changes to a building's electrical system. For dedicated AI builds, a row power center using an overhead 800 VDC busway will support up to 2 megawatts per row, with availability in 2027. Greenfield facilities get a DC power block that converts medium-voltage grid power directly to 800 VDC in a single step.
Key facts
- 01Nvidia, Google, and Microsoft co-developed the 800 VDC architecture through the Open Compute Project, publishing a joint white paper in March 2026.
- 02More than 80 equipment manufacturers and infrastructure companies are building products to the LVDC Solid-State Transformer Specification v0.3, published July 2026.
- 03The MGX-compatible 800 VDC power rack ships in the second half of 2026; row power centers supporting 2 megawatts per row arrive in 2027.
- 04Wood Mackenzie projects $9 trillion in global AI and data infrastructure investment through 2040.
The staged design matters because most operational AI facilities were engineered around AC distribution, and Nvidia is telling those operators they don't need to strand the investment. Land, power interconnects, and building shells stay in place; the hybrid rack lets a colo or hyperscaler push rack density higher without rebuilding the substation-to-slab electrical path.
The 80+ vendors building to the OCP specification are the point of the announcement more than the voltage number itself. An open standard with common interfaces means power shelves, busways, and solid-state transformers from different suppliers can be mixed inside the same facility — a supply-chain condition that AC hyperscale infrastructure took a decade to reach. Nvidia is packaging the building blocks into its DSX reference designs so operators get a system-level blueprint tying power architecture, rack-scale compute, and facility infrastructure together.
Google and Microsoft's involvement gives the standard weight it wouldn't have as a Nvidia-only spec. Both companies operate at a scale where power-delivery architecture is a competitive variable, and both have historically pushed their own internal designs into OCP once they wanted a broader supply chain. Aligning the three largest AI infrastructure buyers on a single DC voltage is the fastest path to getting transformer, busway, and rack vendors to actually invest in tooling.
The open questions are execution and timing. Solid-state transformer economics at 800 VDC are still maturing, and the row power center — the tier most relevant to purpose-built AI campuses — doesn't arrive until 2027, by which point Nvidia will be shipping at least one more GPU generation. Operators committing to sites now have to bet that the retrofit rack in late 2026 is enough of a bridge, and that the vendor ecosystem hits its 2027 delivery windows. Facilities that miscalculate will find their power architecture, not their chip supply, capping compute density.
The strategic read is that Nvidia is no longer selling only silicon and systems — it is trying to set the physical substrate the entire AI industry builds on. Owning the reference architecture for how grid power reaches a GPU is a quieter form of platform control than CUDA, but a stickier one. If 800 VDC becomes the default for new AI campuses and 80+ vendors ship into that spec, the cost curve for anyone trying to build an AI factory without Nvidia's blueprint gets meaningfully worse, and the $9 trillion Wood Mackenzie is forecasting flows disproportionately toward the ecosystem Nvidia already anchors.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
