AI Chips · Product page

NVIDIA GB300 NVL72: The Complete Guide

Same 72-GPU, 36-CPU rack as GB200. NVIDIA calls it a generational leap — most of the gain is 50% more memory per GPU. Simple to expert, every number links to its source.

Stuck at any point? Ask an AI about this page →
Ask an AI about this chip:CAGPX
NVIDIA GB300 NVL72 10 sections · 2 levels 10 linked sources Checked September 2026
How to read this page. Each section starts Simple, then goes to a Deep dive: stop wherever you have what you need. The small numbers are sources: click one to open the original document. Where the maker has not published a figure -- a price, a die size, a factory address -- this page says so rather than estimate it.

1.At a glance

SimpleStart here

The GB300 pairs NVIDIA’s Grace CPU with two Blackwell Ultra GPUs on one module; sold rack-scale as GB300 NVL72 — 72 GPUs and 36 CPUs in one liquid-cooled cabinet. Announced 18 March 2025 at GTC, the same day as B300 1.

Deep diveThe technical detail

Microsoft Azure announced the first production-scale GB300 NVL72 cluster on 9 October 2025, built with OpenAI, at more than 4,600 Blackwell Ultra GPUs 2. CoreWeave announced its own first deployment, built with Dell, on 3 July 2025 3.

2.Launch and history

SimpleStart here

GB300 NVL72 is the rack-scale twin of B300: the same Blackwell Ultra memory refresh, but sold as a complete 72-GPU cabinet with NVIDIA’s own Grace CPUs built in, rather than a board that a customer installs into their own server.

Deep diveThe technical detail

Announced 18 March 2025 alongside B300 1; DGX GB300 availability was described as expected “later this year” (2025) in the same-day SuperPOD release 4. Real deployments followed quickly: CoreWeave with Dell in July 2025 3, and Microsoft Azure’s NDv6 GB300 cluster with OpenAI in October 2025 2.

3.What’s inside it

SimpleStart here

Same shape as GB200 NVL72 — 72 GPUs, 36 CPUs, connected by NVLink — but with half again as much HBM3E memory in the rack, which NVIDIA puts at “1.5x larger HBM3E memory”, and faster networking 5.

Deep diveThe technical detail

GB300 NVL72 carries 72 Blackwell Ultra GPUs and 36 Grace CPUs, the same counts as GB200 NVL72 5 — NVIDIA’s page states GB300’s own configuration rather than drawing the comparison, so the like-for-like reading is this site’s. What changes: total rack GPU memory is 20TB, which NVIDIA describes as “1.5x larger HBM3E memory” than its predecessor without naming that predecessor’s figure; CPU memory is 17TB of LPDDR5X at 14 TB/s, for 37TB of combined fast memory per rack 5. Networking moves up a generation to ConnectX-8 SuperNICs (800 Gb/s per GPU) and BlueField-3 DPUs 16.

4.Full spec table

SimpleStart here

72 Blackwell Ultra GPUs, 36 Grace CPUs, about 20TB of GPU memory and 37TB of combined fast memory per rack, connected by fifth-generation NVLink at up to 130 TB/s 5.

Deep diveThe technical detail
SpecGB300 NVL72 (per rack)
GPUs72x Blackwell Ultra
CPUs36x Grace (Arm Neoverse V2)
GPU memory~20TB HBM3e, up to 576 TB/s aggregate bandwidth
CPU memory17TB LPDDR5X, 14 TB/s
Total fast memory37TB
FP4 Tensor (dense)1,080 PFLOPS
FP8/FP6 Tensor720 PFLOPS
NVLink (5th gen)130 TB/s aggregate per rack
Networking72x ConnectX-8 (800 Gb/s each), 18x BlueField-3, 9x NVLink switches

Figures per NVIDIA’s own GB300 NVL72 and DGX GB300 pages 56; corroborated by CoreWeave’s own deployment post, which independently reports “up to 21TB of high-bandwidth GPU memory per rack”, which it puts at 1.5x the GPU memory of GB200 NVL72 3. Where NVIDIA publishes both a sparse and a dense FP4 figure for this rack, this table uses the dense one as the more conservative number.

5.Where it’s made

SimpleStart here

The Blackwell Ultra GPUs inside GB300 are made by TSMC on the same custom 4NP process used for B300 7. NVIDIA does not state a fab or process for the Grace CPU in its GB300 materials.

Deep diveThe technical detail

NVIDIA’s Blackwell architecture material names TSMC and the custom 4NP node for the GPU die shared by B300 and GB300 7. Neither NVIDIA nor TSMC discloses which fab site produces the wafers, and GB300’s own release material does not restate Grace’s foundry for this generation, so this page does not state one.

6.Which systems use it

SimpleStart here

Sold as the GB300 NVL72 rack and DGX GB300 system. Microsoft Azure and CoreWeave have both announced production deployments; Wistron is named as a manufacturing partner 235.

Deep diveThe technical detail

Microsoft Azure’s NDv6 GB300 VM series powers the first production-scale GB300 NVL72 cluster, built for OpenAI at over 4,600 GPUs 2. CoreWeave’s deployment was built with Dell Technologies for rack integration, with Switch and Vertiv as infrastructure partners and Moonvalley as an early customer 3. NVIDIA names Wistron’s Texas facility as a manufacturing partner for Grace Blackwell Ultra systems, and says Equinix “will be first to offer the new DGX GB300 and DGX B300 systems” in its data centres across 45 markets; NVIDIA’s related Instant AI Factory service was, in the same release, only “planned to be available” later that year 4. Broader OEM partners named for the platform include Cisco, Dell, HPE, Lenovo and Supermicro; cloud partners include AWS, Google Cloud, Azure and Oracle Cloud 1.

7.Official pricing

SimpleStart here

No official GB300 NVL72 price has been published anywhere this page could verify. CoreWeave lists it as “Contact sales” rather than a published rate 8.

Deep diveThe technical detail

CoreWeave’s official pricing page lists GB300 NVL72 as “Contact sales,” with no per-hour or per-rack figure disclosed 8. No published GB300 pricing was found on AWS, Google Cloud or Azure’s own pricing pages; Azure’s NDv6 GB300 announcement discusses capability, not cost 2. As of this check, no official or primary-source GB300 price exists anywhere publicly — this page states that plainly rather than estimating one.

8.Real-world performance

SimpleStart here

NVIDIA says GB300 NVL72 delivers “1.5x more AI performance” than GB200 NVL72 1 — a company-stated comparison, not an independent benchmark.

Deep diveThe technical detail

In MLPerf Inference v5.1, published 9 September 2025, NVIDIA reports GB300 NVL72 set records on the newly added reasoning-inference benchmark, delivering 45% higher DeepSeek-R1 throughput (offline scenario) than GB200 NVL72 9. Microsoft Azure separately cites the same MLPerf v5.1 results as showing up to 5x higher per-GPU throughput on DeepSeek-R1 versus the prior Hopper generation 2.

9.What came before, what came next

SimpleStart here

Came before: GB200 NVL72, announced March 2024. Came after: Vera Rubin NVL72, officially announced 5 January 2026 at CES 10.

Deep diveThe technical detail

GB300 NVL72 succeeds GB200 NVL72, keeping the same 72-GPU rack topology while upgrading memory and networking. NVIDIA’s own newsroom names Vera Rubin NVL72 — pairing NVIDIA’s next-gen Vera CPU with its Rubin GPU — as the platform that succeeds Grace Blackwell, expected in the second half of 2026 10.

10.Hidden in plain sight

SimpleStart here

GB300 NVL72 is marketed on the same headline count as GB200 NVL72 — still “72 GPUs, 36 CPUs.” The generational leap is almost entirely a memory and networking refresh, not a new rack design.

Deep diveThe technical detail

Because the rack topology, NVLink domain size and CPU core count carry over unchanged from GB200 NVL72, and only the GPU memory (up 50%, from 192GB to 288GB per GPU) and networking generation actually changed, customers already running GB200 NVL72 infrastructure could adopt GB300 largely as a drop-in upgrade rather than a new rack architecture — which is exactly why CoreWeave and Microsoft Azure were each able to stand up “first” GB300 NVL72 deployments within months of their GB200 rollouts 32.

11.Sources

10 sources, checked September 2026. Where NVIDIA or another maker has not published a figure, this page says so rather than estimate it.

  1. NVIDIA Newsroom: NVIDIA Blackwell Ultra AI Factory Platform Paves Way for Age of AI ReasoningOfficial
  2. NVIDIA Blog: Microsoft Azure Unveils World’s First NVIDIA GB300 NVL72 Supercomputing Cluster for OpenAIOfficial
  3. CoreWeave: CoreWeave Leads the Way With First NVIDIA GB300 NVL72 DeploymentOfficial
  4. NVIDIA Newsroom: NVIDIA Blackwell Ultra DGX SuperPOD Delivers Out-of-the-Box AI Supercomputer for Enterprises to Build AI FactoriesOfficial
  5. NVIDIA: NVIDIA GB300 NVL72Official
  6. NVIDIA: NVIDIA DGX GB300Official
  7. NVIDIA: NVIDIA Blackwell ArchitectureOfficial
  8. CoreWeave: CoreWeave PricingOfficial
  9. NVIDIA Blog: NVIDIA Blackwell Ultra Sets the Bar in New MLPerf Inference BenchmarkOfficial
  10. NVIDIA Newsroom: NVIDIA Kicks Off the Next Generation of AI With Rubin — Six New Chips, One Incredible AI SupercomputerOfficial

Ask an AI about this page

Opens your assistant with this page as the source, and a question rather than a summary. It will ask what you are building before it answers.

ChatGPTClaudeGeminiPerplexityGrok

Nothing is sent from here. The link carries only this page’s title and address.