--- title: "GB200 NVL72: what it means · The AI Compute Race · DRYBURGH" source: https://dryburgh.com/vol1/glossary/gb200-nvl72 description: "GB200 NVL72: NVIDIA's rack-scale computer, seventy-two accelerators and thirty-six CPUs wired over copper to work as one machine." --- # GB200 NVL72 What does “GB200 NVL72” mean? GB200 NVL72NVIDIA’s rack-scale computer, seventy-two accelerators and thirty-six CPUs wired over copper to work as one machine. It is the cabinet that defined the buildout of 2025. *Chapter 4.* From *The AI Compute Race*. [Read the book](https://dryburgh.com/vol1#buy) ## Where the book uses it - [Chapter 4 — One Rack](https://dryburgh.com/vol1/notes/chapter-4)How a datacenter is actually built, from a converted factory in Memphis to a greenfield campus in Texas, and what the clock from first ground to first power really measures. ## Sources that use it - Harsh Singh Banwait, “NVIDIA Vera Rubin NVL72 on CoreWeave: 10x More Tokens Per Megawatt Than Blackwell,” CoreWeave blog, 21 July 2026 (CoreWeave’s own measurement, on DeepSeek R1: “At a matched interactivity target, Vera Rubin NVL72 generated 10x tokens-per-second per megawatt compared to NVIDIA GB200 NVL72 on the same DeepSeek R1 workload”; the conditions, “with all major, state-of-the-art inference optimizations turned on,” among them “large-scale expert parallelism, NVFP4 precision, multi-token prediction, and disaggregated prefill and decode”; the page’s other wording of the result, “up to 10x higher tokens per megawatt”; what it means for customers, in the company’s words: “run far more reasoning-model traffic within the same power budget or run the same traffic on far less power”; CoreWeave sells the capacity it measured, and NVIDIA’s blog of the same day relays the figure as its own platform’s; the comparison on both systems, “This evaluation compares the performance on each platform with all major, state-of-the-art inference optimizations turned on”; read in the page as served in October 2026, which carries its publication date). [In the Sources](https://dryburgh.com/vol1/sources/company#coreweave-vera-rubin-jul26) - NVIDIA, “Hardware — NVIDIA DGX GB Rack Scale Systems User Guide” (GB200 NVL72: 72 GPUs, 36 CPUs, 18 compute trays, 9 NVLink switch trays; approximately 120 kW), accessed 29 August 2026. [In the Sources](https://dryburgh.com/vol1/sources/company#nvda-gb200-guide) - NVIDIA, “NVIDIA GB200 NVL72” (product page: 72 Blackwell GPUs and 36 Grace CPUs, liquid-cooled). [In the Sources](https://dryburgh.com/vol1/sources/company#nvidia-gb200) - Dylan Patel et al., “GB200 Hardware Architecture — Component Supply Chain & BOM,” *SemiAnalysis*, 17 July 2024 (“123.6kW per NVL72 is the total power draw”; H100 air-cooled racks about 40 kW; general-purpose CPU racks up to 12 kW; of the GB200 NVL72 at about 120 kW a rack: “most datacenter infrastructure cannot support this high of a rack density even with direct-to-chip liquid cooling”). [In the Sources](https://dryburgh.com/vol1/sources/research#semi-gb200) - Epoch AI, “AI Data Centers Documentation – Methodology,” data documentation page, read in the Library copy of 2 November 2025 (on counting units: “There is also a confusing relationship between Grace-Blackwell ‘superchips’ like the GB200, which comprise two B200 GPUs, and the GB200 NVL72 server, which is the equivalent of 36 GB200s, but contains 72 B200 GPUs”). [In the Sources](https://dryburgh.com/vol1/sources/research#epoch-datacenters-methodology) And 3 more: [search the Sources](https://dryburgh.com/vol1/sources). ## Related terms - [One machine](https://dryburgh.com/vol1/glossary/one-machine) - [Blackwell](https://dryburgh.com/vol1/glossary/blackwell) - [GB300](https://dryburgh.com/vol1/glossary/gb300) - [Scale-up](https://dryburgh.com/vol1/glossary/scale-up) - [122 days](https://dryburgh.com/vol1/glossary/122-days) - [Abilene](https://dryburgh.com/vol1/glossary/abilene) - [Accelerator](https://dryburgh.com/vol1/glossary/accelerator) - [Colossus](https://dryburgh.com/vol1/glossary/colossus) - [CPU](https://dryburgh.com/vol1/glossary/cpu) - [GPU](https://dryburgh.com/vol1/glossary/gpu) - [Greenfield](https://dryburgh.com/vol1/glossary/greenfield) - [H100](https://dryburgh.com/vol1/glossary/h100) From the glossary of The AI Compute Race. - [Campuses](https://dryburgh.com/vol1/ledger) - [Tests](https://dryburgh.com/vol1/tests) - [Scorecard](https://dryburgh.com/vol1/scorecard) - [Data](https://dryburgh.com/vol1/data/releases) - [Chapter by chapter](https://dryburgh.com/vol1/companion) - [The record at a glance](https://dryburgh.com/vol1/observatory) - [The machine](https://dryburgh.com/vol1/machine) - [One campus, four futures](https://dryburgh.com/vol1/hyperion) - [Glossary](https://dryburgh.com/vol1/glossary) - [People](https://dryburgh.com/vol1/people) - [Sources](https://dryburgh.com/vol1/sources) - [Notes](https://dryburgh.com/vol1/notes) - [Where the book was wrong](https://dryburgh.com/vol1/corrections) - [Where its sources were wrong](https://dryburgh.com/corrections?book=vol1) - [The Paper](https://dryburgh.com/vol1/paper) - [Consent](https://dryburgh.com/vol1/consent) - [Who builds and pays](https://dryburgh.com/vol1/parties) *The AI Compute Race: Five Races, One Machine* is a book by Lee S. Dryburgh, the first volume of THE COMPUTE ECONOMY, published by Dryburgh, an imprint of Contestra OÜ, Tallinn, Estonia; its record closes on 8 November 2026. [About the book](https://dryburgh.com/vol1)[Tell me when it is out](https://dryburgh.com/vol1#notify)