Nvidia's 800-volt data centre now starts with one rack in the building you already have
The 2025 pitch converted grid power at the perimeter, while the product that actually ships slots into existing AC plant and keeps most of the conversion stages it was meant to delete.
The pitch used to start at the fence line. Nvidia's plan for powering AI factories began where the grid meets the property, with industrial rectifiers taking high-voltage alternating current straight down to 800 volts DC and most of the conversion stages inside the building deleted.
The product that actually arrives is a rack.
On 9 September 2026 SemiAnalysis laid out the rollout in three steps, each designed to run alongside the AC plant rather than replace it. First a power rack that delivers 800 VDC into the compute racks and standardises the DC interface. Then a row-level power center that rectifies centrally and feeds a dense row over an overhead busway. Then, eventually, a hall-scale DC power block, with the level of adoption decided facility by facility.
Nvidia's own description of the first step is the giveaway. The power rack is "designed to slot into existing AC infrastructure and deliver 800 VDC to compute racks within the row — no changes to the building's electrical system required."
That is a much smaller claim than the one made in 2025.
It is also, if you run a building that already exists, a much better product.
The physics was never the part in dispute
At 54 volts DC, a one-megawatt rack needs up to 200 kg of copper busbar. Scale that to a gigawatt campus and the rack busbars alone want 200 tonnes of copper, before a single cable leaves the row.
And space runs out before the copper does. At the same voltage, the power shelves for a Kyber-class rack at megawatt scale would swallow up to 64 rack units, in a cabinet with roughly forty to give. So there is nowhere left to put the computer.
The savings figure does not survive the retreat
Nvidia's published numbers are 85 percent more power through the same conductor and 45 percent less copper, both measured against the 415-volt AC distribution in use today. Then come the softer ones, covering efficiency, maintenance and the whole cost of ownership, where the headline is up to 30 percent off the cost of running the place.
Hold that last one loosely. They were published in 2025 against a fully native facility with a single conversion step at the perimeter. A power rack sitting behind an unchanged AC plant keeps every upstream conversion stage the design was meant to remove, so it captures the in-rack share of the saving and not much of the rest.
Nobody has published a figure for the hybrid case. Until somebody does, read the headline saving as a description of a building almost nobody will operate before 2029. (Which is not the same as saying the number is wrong.)
The retreat is how electrical standards actually get adopted
So is the smaller product bad news? Probably the opposite. A new distribution voltage arrives through hybrids or it does not arrive, and the backing here is unusually solid.
Nvidia, Google and Microsoft developed the architecture together inside the Open Compute Project, published a joint white paper in March 2026 and a low-voltage DC solid-state transformer specification in July, and more than 80 equipment makers are already building to it. A specification with that many suppliers behind it is a supply chain rather than a slide.
The row is where it stops being free. The power center carries up to 2 MW per row over an overhead busway, Nvidia dates it to 2027, and it is the first step that asks an operator to change something structural. SemiAnalysis says Nvidia will "collaborate with industry partners to expand demonstration and pilot sites ahead of large-scale deployment", which is how you describe a product with no reference customer to name yet (the pilot sites are presumably named in private).
We would expect essentially every 800 VDC deployment announced through 2027 to be a power rack inside an AC building, and the hall-scale power block to still have no named operator by the end of that year. A greenfield gigawatt campus committing to the native design before then would move us.
Either way, somebody saves 200 tonnes of copper per gigawatt. Copper is not getting cheaper.
