Vera Rubin vs Blackwell: what NVIDIA says has changed

Vera Rubin is NVIDIA's data center platform after Blackwell, and the vera rubin vs blackwell comparison comes down to NVIDIA's claims of about 5x the inference speed per GPU and up to 10x lower cost per token. On October 27, 2026, NVIDIA and Oracle said Vera Rubin NVL72 and Vera CPUs are now generally available on Oracle Cloud Infrastructure. Nearly every performance number below comes from NVIDIA. Independent test results for the full rack were not in the sources we read.

This is the chip company's platform, not the Vera C. Rubin telescope in Chile.

What Vera Rubin is

Vera Rubin is a set of chips designed to work as one system. NVIDIA's technical blog lists six: the Vera CPU, the Rubin GPU, the NVLink 6 switch, the ConnectX-9 network card, the BlueField-4 data processing unit and the Spectrum-6 Ethernet switch. A later note on that page adds a seventh, the Groq 3 LPX.

The product most people mean is the Vera Rubin NVL72, a liquid-cooled rack. According to NVIDIA's product page, it holds 72 Rubin GPUs and 36 Vera CPUs, wired together so they act as one large accelerator. The Vera CPU has 88 custom cores. The GPUs use HBM4, a newer, faster kind of memory stacked next to the chip.

Vera Rubin vs Blackwell on NVIDIA's numbers

The table compares one Rubin GPU with one Blackwell GPU, using the table in NVIDIA's technical blog.

NVFP4 is a 4-bit number format that lets the chip do more math per second. The inference figure uses NVIDIA's Transformer Engine, which depends on the software to reach it. NVIDIA's own pages disagree in places. The product page lists 19.2 TB/s of memory bandwidth and 3 TB/s of NVLink per Rubin GPU, against 22 TB/s and 3.6 TB/s in the blog. NVIDIA does not explain the gap in the pages we read.

At rack level, NVIDIA lists 3,600 PFLOPS of NVFP4 inference (a sparse figure), 20.7 TB of HBM4 and 260 TB/s of NVLink 6 bandwidth for the NVL72.

The cost and throughput claims

The bigger numbers are for whole racks, and each depends on a workload and a baseline.

• Against GB200 NVL72, NVIDIA says Vera Rubin NVL72 delivers up to 10x more tokens per megawatt and up to 10x lower cost per million tokens. Both are for the Kimi-K2-Thinking model with 32K input and 8K output tokens, and NVIDIA marks them "subject to change."
• Against GB300 NVL72, an NVIDIA blog post from August 24, updated September 15, says up to 30x more throughput per megawatt and up to 45x lower cost per million tokens. NVIDIA calls the results early and ties the 30x figure to DeepSeek V4 Pro on SemiAnalysis's AgentX test of recorded coding sessions.
• For training, the product page says a projected 10-trillion-parameter mixture-of-experts model could be trained in a month on one-fourth the GPUs of GB200 NVL72.

A token is a small chunk of text a model reads or writes, so cost per token is a rough measure of what it costs to serve a model. "Up to" means the best case NVIDIA picked. Treat these as a ceiling, not a typical result.

Is Vera Rubin shipping

According to NVIDIA, yes, with a ramp. In its May 31 GTC Taipei release, NVIDIA said Vera Rubin was ramping into full production and that "production shipments of Vera Rubin are set to begin starting this fall." The product page now says "Available Now" and sends buyers to NVIDIA sales. The Oracle release adds the first named cloud launch: general availability on OCI as of October 27.

NVIDIA has not published, in the pages we read, how many racks have shipped or when a given buyer will get them.

Who has announced deployments

Oracle is the one cloud that has said Vera Rubin NVL72 is generally available. It and NVIDIA also announced plans to expand OCI Superclusters with Vera Rubin NVL72 to multi-gigawatt scale. That expansion is a plan, not finished capacity. The release says it builds on OCI's existing GB200 and GB300 NVL72 systems. A planned HGX Rubin NVL8 server, with eight Rubin GPUs and two Intel Xeon 6 CPUs, is not yet available.

OpenAI appears in the release for the CPU, not the GPU. NVIDIA says OpenAI is an early-access customer for Vera CPUs and is moving them into production on OCI. The release also says OpenAI trained and served GPT-6 Astra on an OCI Supercluster with more than 100,000 Blackwell GPUs, which was the previous generation.

The May release names Dell Technologies, HPE, Lenovo and Supermicro among the makers of production systems, and says CoreWeave, Lambda and Oracle are early adopters of its co-packaged optics networking. It names no AI lab as a Vera Rubin GPU customer.

What to watch

Independent benchmarks of the NVL72 against GB200 and GB300 will show how much of the 10x to 45x range holds outside NVIDIA's chosen workloads. Also watch whether other clouds announce general availability, as Oracle did, and whether NVIDIA gives shipment numbers. Oracle's disclaimers say product timing and pricing may change.