Nvidia efficiency claims coincide with a sharp rise in AI rack power density

Nvidia’s claimed improvements in inference efficiency are arriving alongside a materially higher power-density requirement at rack level. According to HackerNoon, the Vera Rubin platform, now being supplied to customers, claims up to 10x more inference performance per watt than Blackwell. The source also says disaggregated inference using Nvidia Dynamo can deliver up to 35x more tokens per watt than GPU-only designs.

Those efficiency claims do not, however, imply lower infrastructure demand per rack. HackerNoon compares an approximately 40 kW Hopper rack in 2023 with a Vera Rubin NVL72 rack at around 210 kW. It further describes the projected Rubin Ultra, or Kyber, rack as a 600 kW design expected in 2027, while noting that systems above 1 MW are reportedly already in development.

For data-center design teams, the implication is a continued shift from conventional air-cooled deployment assumptions toward direct liquid cooling and CDU-led thermal architectures. Higher rack densities also raise requirements across the electrical chain, including utility supply, grid interconnection, on-site distribution and heat-rejection infrastructure.

The source cites a Dell’Oro forecast that the data-center liquid-cooling market will exceed US$7 billion by 2029. This could support structural demand for liquid-cooling equipment, CDUs, high-power electrical distribution and associated energy infrastructure, although the forecast is not independently substantiated in the provided material.

HackerNoon also argues that large technology companies’ nuclear-energy initiatives are moving beyond press releases toward construction and regulatory activity. The available extract provides no project-level capacity, investment, timetable or facility details, so this assertion cannot be assessed at an individual-project level.