The Definitive Guide toAI Data Centers
Ask the GuideAboutAccount
Guide › Numbers register › Claim

per-rack power roadmap by system: four-system DGX H100 40.8 kW (2023); HPE GB200 NVL72 132 kW nominal; Lenovo GB300 NVL72 135 kW TDP and up to 155 kW peak (NVIDIA reference architecture: up to 142 kW for a full rack); Pegatron VR200 NVL72 188 kW Max Q / 228 kW Max P, with NVIDIA's DSX reference design sizing the Vera Rubin NVL72 cabinet at 330 kW TDP; Rubin Ultra / Kyber ~600 kW facility planning point

~40 → 600 kWforecastcontested

Value kindforecast — Forecasts are time-bound outlooks, not specifications; test decisions across plausible scenarios.
CaveatNVIDIA targets H2 2027 for Rubin Ultra / Kyber; SemiAnalysis reported in July 2026 that it slips to 2028. The earlier points come from different OEM and reference-design profiles, so they do not trace one platform's progression. The 330 kW figure is the reference design's sizing basis, not a measured operating load.
As of2026-08
SourceNVIDIA DGX H100 data-center planning guide (4 x 10.2 kW); HPE NVIDIA GB200 NVL72 QuickSpecs; Lenovo Press LP2357 (updated 2026-08-30); NVIDIA Enterprise Reference Architecture, March 2026; Pegatron RA4803-72N3 datasheet (2026); NVIDIA DSX Facilities Infrastructure Reference Design v2.0 (2026-08-19); Kyber from NVIDIA and SemiAnalysis reporting · Per point: DGX H100 rack-density table; HPE 'Power Delivery and Requirements' (132 kW nominal); Lenovo 'Each rack consumes 135 kW TDP'; NVIDIA RA 'Full rack requiring up to 142 kW'; Pegatron 'Max Q = 188kW Max P = 228kW'; DSX 'Vera Rubin NVL72 at 330 kW' (power and cooling sizing basis).
Reviewchecking…review by 2026-12-15 · standard cadence
Recorded changeslast 2026-10-13 · 7 revisions tracked
Claim idper-rack-power-roadmap-h100-40-kw-2023-gb200

Where the guide uses it

← Full numbers register — every date-stamped figure in the guide, with revision history.