# Rack Power Density

Canonical URL: https://menfem.com/kb/datacenters/concepts/rack-power-density
Knowledge base topic: [Datacenters & Digital Infrastructure](https://menfem.com/kb/datacenters)
Frontier status: active
Tags: rack-power-density, positioning, market-structure

---

Rack power density is the amount of electrical power (and therefore heat) concentrated into a single server rack. The shift to Nvidia's **GB300 NVL72** rack-scale systems is the density step-change driving the modern AI datacenter: the NVL72 is a single rack integrating 72 Blackwell GPUs as one unit, and these racks are now the atomic build-unit of the largest AI factories. The NVIDIA / DOE source quantifies the deployment scale — **Lambda's AI Factory in Kansas City is 100+ MW with 10,000+ GB300 NVL72 GPUs**, and **Global AI's New York site is 128 GB300 NVL72 racks (9,000+ GPUs)** — but it describes the systems in GPU-count and total-megawatt terms, not in per-rack kilowatts.

The reason density matters for the buildout is thermal. As power per rack climbs into a regime where air cannot remove the heat, **liquid cooling becomes mandatory rather than optional** — which is the mechanism that re-rates the power-and-cooling supply chain (Vertiv, Eaton, GE Vernova, Schneider Electric, and the rest of the NVIDIA-named vendor list). Higher density also concentrates the power problem: the same gigawatts of campus demand arrive in fewer, hotter racks, tightening the link between this concept and [Power: The Binding Constraint](./power-the-binding-constraint.md). The frontier debate this feeds — "does liquid cooling become mandatory and therefore Vertiv-et-al a secular winner, or stay niche?" — turns on exactly how high NVL72-class density goes.

> **RESOLVED 2026-07-23 — the kW/rack figure is now sourced, closing the topic's oldest gap.** A Lenovo vendor product guide for the GB300 NVL72 (doc updated 2026-07-15) states **135 kW rack TDP, up to ~155 kW peak** (50V DC rail; 72 B300 GPUs + 36 Grace CPUs), with heat captured **~90% liquid / ~10% air**. This is the first primary vendor datasheet in the topic and fills the gap that had been open across the prior 13 sources. Provenance: `technical-report`-tier, but a **vendor document** — authoritative for the spec, not independent. Cross-vendor corroboration from the same search (not ingested): HPE QuickSpecs cited 132–140 kW with EDPp provisioning up to ~192 kW; the Lenovo 135 kW TDP / 155 kW peak sits inside that band. At 135 kW nominal a single NVL72 rack draws roughly **18–19× a legacy ~7 kW enterprise rack** — the physical reason the 90/10 liquid split is forced, not optional. (Note: the older "~120–140 kW" number previously cited secondhand is now corroborated by a primary vendor spec; the "peak up to 155 kW" is a ceiling, not a sustained draw.)

## Key Claims

- **GB300 NVL72 is the rack-scale build-unit** of the largest 2025–26 AI factories. *Evidence: moderate (technical-report — NVIDIA)* ([NVIDIA — Building America's AI Infrastructure](../../raw/nvidia-doe-ai-infrastructure.md))
- **Lambda AI Factory (Kansas City): 100+ MW, 10,000+ GB300 NVL72 GPUs.** *Evidence: moderate (technical-report — NVIDIA)* ([NVIDIA — Building America's AI Infrastructure](../../raw/nvidia-doe-ai-infrastructure.md))
- **Global AI (New York): 128 GB300 NVL72 racks (9,000+ GPUs).** *Evidence: moderate (technical-report — NVIDIA)* ([NVIDIA — Building America's AI Infrastructure](../../raw/nvidia-doe-ai-infrastructure.md))
- **DOE supercomputers anchor on Blackwell at scale:** Solstice 100,000 Blackwell GPUs (DOE's largest), Equinox 10,000 Blackwell (2026), combined 2,200 exaflops AI; xAI Colossus 2 (Memphis) 500,000+ GPUs. *Evidence: moderate (technical-report — NVIDIA)* ([NVIDIA — Building America's AI Infrastructure](../../raw/nvidia-doe-ai-infrastructure.md))
- **~~Per-rack kW figure is NOT in any ingested source~~ — RESOLVED 2026-07-23.** GB300 NVL72 is **135 kW rack TDP, ~155 kW peak** per a Lenovo vendor product guide (as-of doc date 2026-07-15). *Evidence: moderate (technical-report — vendor datasheet, not independent)* ([Lenovo GB300 NVL72](../../raw/gb300-nvl72-rack-power-135kw-lenovo.md))

### Added 2026-07-23 (closes the topic's oldest gap)

- **GB300 NVL72 rack power: 135 kW TDP, up to ~155 kW peak** (50V DC rail; power shelves 6–8 × 33 kW = 198–264 kW installed headroom above nominal). *Evidence: moderate (technical-report — Lenovo vendor guide, as-of 2026-07-15)* ([Lenovo GB300 NVL72](../../raw/gb300-nvl72-rack-power-135kw-lenovo.md))
- **Compute per rack: 72 × B300 (Blackwell Ultra) GPUs, 186 GB HBM3e each + 36 × Grace CPUs.** *Evidence: moderate (technical-report — vendor guide)* ([Lenovo GB300 NVL72](../../raw/gb300-nvl72-rack-power-135kw-lenovo.md))
- **Heat capture ~90% liquid / ~10% air** at the rack — CPUs, GPUs, HBM, NVSwitch and NICs are liquid-cooled; only OSFP modules, drives and PDUs stay air-cooled. This is the sourced mechanism behind "liquid cooling becomes mandatory." *Evidence: moderate (technical-report — vendor guide)* ([Lenovo GB300 NVL72](../../raw/gb300-nvl72-rack-power-135kw-lenovo.md))
- **A GB300 NVL72 rack draws ~18–19× a legacy ~7 kW enterprise rack**, and facility provisioning runs above nameplate (HPE cross-ref ~192 kW busway for EDPp headroom). *Evidence: moderate (derived from vendor spec + cross-vendor cross-reference)* ([Lenovo GB300 NVL72](../../raw/gb300-nvl72-rack-power-135kw-lenovo.md))
- **~7.4 racks per MW** at 135 kW nominal — the unit that ties rack density to the capex/MW and grid-queue figures. *Evidence: moderate (arithmetic on the vendor spec)* ([Lenovo GB300 NVL72](../../raw/gb300-nvl72-rack-power-135kw-lenovo.md))

### Added 2026-07-22

- **Cooling architecture is visibly changing generation-over-generation:** Tesla's Cortex 2 doubles Cortex 1's cooling capacity (two chiller plants, ≥12 fan units) and moves cooling equipment *inside structures* rather than siting it at ground level — an independent, non-NVIDIA datapoint that density is forcing physical redesign, not just vendor selection. *Evidence: weak (single analysis source, one site)* ([Tesla Cortex 2](../../raw/tesla-cortex-2-500mw-gigatexas.md))
- **Density is now supply-gated as well as thermally gated:** NVL72-class rack deployment depends on CoWoS advanced packaging, reported sold out through end-2026 with lead times into 2027. Density can be designed for faster than racks can be delivered. *Evidence: weak (single analysis source)* ([TSMC Q2 2026](../../raw/tsmc-q2-2026-cowos-sold-out-through-2026.md))
- ~~**kW/rack remains a gap after this compile.**~~ *(True as of 2026-07-22; **closed 2026-07-23** by the Lenovo GB300 NVL72 vendor spec — 135 kW TDP / 155 kW peak. See the 2026-07-23 block above.)*

## Benchmarks & Data

- GB300 NVL72: 72 Blackwell GPUs per rack (rack-scale system)
- Lambda AI Factory: 100+ MW, 10,000+ GB300 NVL72 GPUs
- Global AI (NY): 128 GB300 NVL72 racks, 9,000+ GPUs
- Solstice: 100,000 Blackwell GPUs (DOE's largest AI supercomputer)
- Equinox: 10,000 Blackwell GPUs (2026); Solstice+Equinox = 2,200 exaflops AI
- xAI Colossus 2 (Memphis): 500,000+ GPUs
- **kW/rack (GB300 NVL72): 135 kW TDP, ~155 kW peak** — *filled 2026-07-23* (Lenovo vendor guide); ~18–19× a legacy ~7 kW rack; ~7.4 racks/MW
- GB300 NVL72 compute: 72 × B300 GPUs (186 GB HBM3e each) + 36 × Grace CPUs; ~90% liquid / ~10% air cooling; 50V DC rail

*Added 2026-07-22:*

- Tesla Cortex 2 cooling: 2 chiller plants, ≥12 fan units (~2x Cortex 1); cooling housed inside structures
- Tesla Cortex 2 campus power: 500MW
- NVL72 rack supply gate: CoWoS sold out through end-2026, lead times into 2027

## Open Questions

- **~~What is the actual kW/rack for GB200/GB300 NVL72?~~ ANSWERED 2026-07-23: 135 kW TDP / ~155 kW peak** (Lenovo vendor guide). Remaining refinement: an NVIDIA-primary or independent-lab figure on *sustained* vs peak draw, and the GB200-vs-GB300 delta (this source is GB300-specific).
- **Is the density step-change relieved or worsened by the packaging constraint?** If racks are supply-limited, operators may build for higher density per delivered rack, pushing kW/rack up faster than the cooling supply chain expects.
- **Does liquid cooling become mandatory** (making the cooling vendors a secular winner) or stay niche? (Frontier debate #3.)
- At what power-per-rack threshold does air cooling become physically infeasible?
- How does rising density change capex-per-MW economics for plant and cooling?

## Related Concepts

- [Power: The Binding Constraint](./power-the-binding-constraint.md) — density concentrates the same gigawatts into fewer, hotter racks
- [Hyperscaler Capex Cycle](./hyperscaler-capex-cycle.md) — the ~$85B networking/cooling slice of plant capex is what density pulls upward
- [Vertically Integrated Compute](./vertically-integrated-compute.md) — Cortex 1 → Cortex 2 is a documented density-driven cooling redesign
- [Advanced Packaging Bottleneck](./advanced-packaging-bottleneck.md) — NVL72 racks cannot be deployed faster than CoWoS delivers them

## Changelog
- 2026-06-24 — Created from 1 source (nvidia-doe-ai-infrastructure)
- **2026-07-22** — Added Tesla Cortex 2 as an independent non-NVIDIA datapoint on density-driven cooling redesign, and the CoWoS supply gate on NVL72 deployment. kW/rack gap re-confirmed as still unfilled across all 10 sources.
- **2026-07-23** — **Closed the topic's oldest gap:** GB300 NVL72 rack power = **135 kW TDP / ~155 kW peak** from the first primary vendor datasheet (Lenovo product guide, 2026-07-15), with the ~90% liquid / ~10% air split, 72×B300 + 36×Grace compute, and the derived ~18–19× legacy-rack ratio and ~7.4 racks/MW. Provenance flagged as vendor-primary (authoritative for spec, not independent).

## Sources

- nvidia-doe-ai-infrastructure
- tesla-cortex-2-500mw-gigatexas
- tsmc-q2-2026-cowos-sold-out-through-2026
- gb300-nvl72-rack-power-135kw-lenovo

---

Cite as: MenFem Knowledge Base — https://menfem.com/kb/datacenters/concepts/rack-power-density