Nvidia's Kyber architecture targets 1 MW racks running 144 GPUs. The math at 48V DC -- today's touch-safe standard -- requires 32 power cables feeding that rack. That is physically impractical. The solution is 800V DC: fewer, smaller conductors, centralized conversion, higher power-per-wire. The problem is that the 800V infrastructure does not exist at the datacenter scale Kyber requires, and Schneider Electric's field data says it will not until at least 2029 for centralized solutions, with small-scale sidecar deployments starting by late 2026 covering 600 kW to 1 MW.
The mechanism is infrastructure sequencing. Sidecar 800V converters inside or adjacent to the rack come first, handling one to three Kyber-class racks each. Centralized DC distribution -- 2 to 5 MW converters feeding multiple rack rows -- follows in 2028-2029 as standards mature and interoperability between vendors emerges. Schneider's forecast is that 10% of new AI nodes will require 800V by 2030, not the broader fleet. New datacenters will have pockets of 800V alongside conventional medium and low voltage zones.
For silicon teams, this is a planning constraint with a specific shape. Chips targeting the first wave of Kyber-class deployments (late 2026 to 2027) will land in sidecar-fed racks with limited scale. The designs that assume broad 800V availability for 2027 tape-outs are ahead of the infrastructure curve by 18 to 24 months. Teams building thermal and power delivery models for Blackwell successors should plan around the sidecar scenario for initial production -- centralized 800V is the 2029 story, not the 2027 one.