GB200 vs GB300 vs Vera Rubin NVL72 Explained

GB200 NVL72, GB300 NVL72 and Vera Rubin NVL72 are three generations of the same Nvidia rack design: 72 GPUs and 36 CPUs linked by NVLink so the whole rack behaves like one large accelerator. GB200 uses Blackwell GPUs, GB300 uses Blackwell Ultra, and Vera Rubin uses new Rubin GPUs and Vera CPUs with HBM4 memory. The rack count stayed at 72; the chips inside changed. This guide covers the naming, the Nvidia-published specs and what is still on the roadmap as of October 2026.

What the names mean

GB stands for Grace Blackwell: an Arm-based Grace CPU paired with Blackwell GPUs. GB200 is the original, GB300 the Blackwell Ultra refresh.
VR stands for Vera Rubin: the Vera CPU paired with Rubin GPUs. Both are named after people, and the Vera CPU replaces Grace.
NVL72 means an NVLink domain of 72 GPUs. NVLink is Nvidia's chip-to-chip interconnect, and in these racks it lets every GPU talk to every other GPU through switch trays rather than over the data-centre network.

The numbers in a name describe the NVLink domain, not the product tier. A single GB200 superchip, for example, is one Grace CPU and two Blackwell GPUs; the NVL72 rack combines 36 of them.

Why Vera Rubin went from NVL144 to NVL72

When Nvidia first previewed the Vera Rubin rack in 2025, it called it NVL144. That figure counted GPU dies, because each Rubin package contains two compute dies. Nvidia has since switched to counting packages, which matches how the Blackwell racks are named, and its current product pages and press materials say Vera Rubin NVL72. The hardware is the same idea: 72 Rubin GPU packages in one NVLink domain. If you see NVL144 in older coverage, it refers to this rack under its earlier name, not a larger system.

GB200 vs GB300 vs Vera Rubin: specs side by side

All figures below come from Nvidia's own product pages as of October 2026. Nvidia quotes the same rack size for each generation, so the comparison is mostly about memory, interconnect and low-precision compute.

GB200 NVL72 (Blackwell)

72 Blackwell GPUs and 36 Grace CPUs, with 2,592 Arm Neoverse V2 cores in total.
13.4 TB of HBM3E GPU memory at 576 TB/s across the rack.
17 TB of LPDDR5X CPU memory.
Fifth-generation NVLink: 1.8 TB/s per GPU and 130 TB/s for the rack.
1,440 PFLOPS of FP4 Tensor Core compute with sparsity (720 PFLOPS dense), per Nvidia's GB200 NVL72 page.

GB300 NVL72 (Blackwell Ultra)

72 Blackwell Ultra GPUs and 36 Grace CPUs, the same CPU count and core total as GB200.
20 TB of GPU memory at up to 576 TB/s, about 1.5 times the GB200 rack's 13.4 TB.
17 TB of LPDDR5X CPU memory and 130 TB/s of NVLink bandwidth, unchanged from GB200.
Nvidia lists FP4 compute at 1,440 PFLOPS with sparsity (1,080 PFLOPS dense) on its GB300 NVL72 page.

In short, GB300 keeps the GB200 rack layout and NVLink generation and mainly adds GPU memory. That makes it a drop-in step for operators who already run the earlier rack.

Vera Rubin NVL72 (Rubin)

72 Rubin GPUs and 36 Vera CPUs. Each Vera CPU has 88 custom Olympus cores (176 threads).
288 GB of HBM4 per GPU, or 20.7 TB across the rack.
Up to 54 TB of LPDDR5X per rack, up to 1.5 TB per Vera CPU.
Sixth-generation NVLink: 216 TB/s for the rack, against 130 TB/s on the Blackwell racks.
Per GPU, 50 PFLOPS of NVFP4 for inference and 35 PFLOPS for training; for the rack, 3,600 and 2,520 PFLOPS. See Nvidia's Vera Rubin NVL72 page.
Compute trays are fully liquid-cooled, and ConnectX-9 networking provides 1.6 Tb/s of scale-out bandwidth per GPU, according to Nvidia's Rubin platform technical post.

Note that Nvidia's FP4 inference and training numbers use different counting conditions (sparse versus dense), so compare like with like. Per-GPU HBM4 bandwidth is quoted differently across Nvidia pages, so this guide leaves it out.

GB200 vs GB300: what actually changed

Blackwell Ultra mostly raises memory capacity, which matters for large models and long contexts that must keep their working state in HBM. The CPU, the NVLink generation and the rack size stayed the same. Treat Nvidia's own performance multipliers as vendor claims measured under stated conditions; independent benchmarks, such as MLPerf results, are the better check.

Blackwell vs Rubin: what changed

Memory type: HBM3E gives way to HBM4. For background on that jump, see our explainer on HBM4 vs HBM3E.
Interconnect: NVLink 5 to NVLink 6, lifting rack bandwidth from 130 TB/s to 216 TB/s.
CPU: Grace (72 Arm Neoverse V2 cores) to Vera (88 custom Olympus cores).
Serviceability: Nvidia says compute-tray assembly drops from more than 1.5 hours on Blackwell to about 5 minutes.

Nvidia also claims that Vera Rubin NVL72 needs about a quarter of the GPUs of Blackwell NVL72 to train a 10-trillion-parameter mixture-of-experts model, and delivers up to 10x lower cost per million tokens for long-context inference. These are company projections, not measured third-party results.

Availability as of October 2026

Nvidia's product pages list both GB300 NVL72 and Vera Rubin NVL72 as "Available Now" through its sales channel, and the Rubin page describes the platform as ramping into full production. On October 27, 2026, Nvidia and Oracle announced plans to expand Oracle Cloud Infrastructure Superclusters with Vera Rubin NVL72 systems and to bring Vera CPUs to OCI; the announcement gives no customer availability dates. Nvidia does not publish prices; cost depends on a negotiated deal with the vendor or a cloud provider.

What comes after: Rubin CPX and Rubin Ultra

Two further products appeared on Nvidia's roadmap when it was shown in 2025. This guide has not re-checked their current timing, so confirm it with Nvidia's latest statements:

Rubin CPX is a GPU aimed at the compute-heavy context phase of long-context inference, designed to sit alongside standard Rubin GPUs in a combined rack.
Rubin Ultra NVL576 is a planned larger rack-scale system, announced in 2025 for 2027, with 576 GPU dies in one NVLink domain.

The Rubin Ultra name follows the die-counting convention that Vera Rubin abandoned for its own naming, which is a reminder to check what each number counts. For the longer history of Nvidia's chip families, see our list of Nvidia GPU architectures, and for how these chips are built, our guide to CoWoS packaging. More coverage sits in Chips and Big Tech.

Frequently asked questions

What is the difference between GB300 and GB200?

GB300 swaps Blackwell GPUs for Blackwell Ultra and raises rack GPU memory from 13.4 TB to 20 TB. The rack size, Grace CPU count and fifth-generation NVLink are the same.

What does NVL72 mean?

NVL72 means 72 GPUs connected in one NVLink domain, so they work as a single large accelerator. In the GB200, GB300 and Vera Rubin racks, those 72 GPUs sit with 36 CPUs.

Is Vera Rubin NVL144 the same as Vera Rubin NVL72?

Yes, they describe the same rack. NVL144 counted GPU dies, while NVL72 counts GPU packages, and Nvidia now uses NVL72.

How is Rubin different from Blackwell?

Rubin moves to HBM4 memory, sixth-generation NVLink and the Vera CPU. Nvidia lists 216 TB/s of NVLink bandwidth per rack against 130 TB/s for Blackwell racks.

How much does an NVL72 rack cost?

Nvidia does not publish list prices for these racks; buyers go through Nvidia sales, system makers or cloud providers. Any figure you see elsewhere is an estimate, not an Nvidia price.