Nvidia's 800V gain is 3 to 5 points; operations explain the rest
Nvidia's own blog puts the 800V DC gain at 3 to 5 points. An Nvidia vice president spoke of 20 points, and the difference sits in rack configuration, restarts and stranded provisioning.
A developer signs a lease for 100 megawatts of AI capacity. Nvidia vice president Rod Evans says it receives at best 93 megawatts at the meter. At Vertiv's EMEA press conference in Zagreb, Evans said, "you sign a contract for 100 megawatts, but you never actually get that."
Evans added that with a bit of luck the buyer gets 93 megawatts, and at certain times of the day even less. That loss comes before a single GPU, the chip that does the AI work, is switched on.
The loss grows after the meter. Evans said a gigawatt scale AI data center draws power from the grid, but only 60% to 70% of that draw is used for AI work.
Applying that share to the megawatts that arrive, a 100 megawatt lease yields roughly 56 to 65 megawatts of AI work. Each useful megawatt costs 54% to 79% more than the contract rate.
Capacity is sold in contract megawatts and consumed in delivered ones, yet that gap rarely appears in a term sheet. The recoverable part of the loss sits in operations, meaning rack configuration and job failures. Nvidia's own numbers put it outside the 800V DC architecture that draws the attention.
According to Evans's remarks as reported by Techzine Global, about 20% of the power is lost to the building itself. That covers cooling, power distribution and conversion. Another 10% goes to racks that are not optimally configured, and about 10% goes to failures and restarts.
The two 10% items are operational losses, neither of them a matter of voltage. Nvidia therefore offers two very different estimates of what can be won back.
Evans said 800V DC distribution should raise the efficiency of an AI facility from 60% to between 80% and 85%. That would be a gain of 33% to 42% in AI work for the same grid draw.
Nvidia's blog post of September 15 is far more modest. It projects a 3% to 5% end to end efficiency gain from 800V DC against today's 54V distribution. That gain would arrive with Vera Rubin NVL72 racks in 2027.
The two figures cover different scopes. The blog's number appears to cover the conversion chain, the electrical steps between the building and the chip. Evans's number covers the whole facility, including rack layout and restarts that a change of voltage cannot affect.
If the blog is right about conversion, most of the 20 point improvement he described has to come from somewhere other than 800V DC. Nvidia's blog points to dynamic allocation: DSX MaxLPS software allocates power across a cluster.
Combined with data center power planning, the company projects up to 40% more GPU capacity on Vera Rubin NVL72 within the same megawatt budget. In a Lambda test, 19 nodes ran within the power budget of 16 nodes at full power. Cluster wide token throughput rose 24%.
Evans described the mechanism himself. A data center is designed for peak load and outages, while its usual load is lower. Static provisioning therefore reserves headroom that the workload rarely needs.
Software that watches actual draw and reallocates it turns reserve into compute without a new feeder or transformer. The operational items are therefore the first ones to test. A lessee should price a contract on delivered rack megawatts at the meter, not the nameplate figure.
The same logic applies to anyone financing a campus. The model should assume 56 to 65 megawatts of AI work for every 100 contracted. Stranded rack capacity should be tested before paying for new grid connections.
The 20% building loss has a physical floor. It implies a power usage effectiveness (PUE) of 1.25, the ratio of total facility power to the power reaching IT equipment.
Only about 36% of the facilities required to report had done so. The Nvidia figure therefore sits above that average without being extreme. Part of the difference may be definitional, because Techzine notes that Nvidia counts every loss from grid power while Vertiv counts only the power used by supporting infrastructure.
That leaves the two 10% items as the discretionary part of the loss. They are the ones a lessee can negotiate over, and they are far larger than a 3% to 5% voltage gain.
The test matters because grid capacity is scarce. Evans said the US has 12 gigawatts of new data centres planned for 2026 and will complete about 5. He put 2027 at 36 gigawatts planned, with about 7 expected.
Roughly three quarters of announced capacity does not arrive on schedule. The megawatts already contracted are therefore worth more than the ones on a planning slide. The share of them that reaches a GPU becomes a first order variable in any return model.
The first public test will be regulatory. From 2027 the EU label for data centres of 500 kilowatts or more will show energy and water consumption. The Commission is expected to propose minimum requirements for new facilities in the second quarter of 2027.
Those rules will force operators to publish an efficiency figure. A building share of 20%, which implies a PUE of 1.25, can then be set against the Dutch average of 1.17. A lessee will have a number to hold a landlord to.