NVIDIA Vera Rubin: What It Is and Why It Matters

NVIDIA’s Vera Rubin platform has moved from keynote slides to real shipments. NVIDIA says production shipments began in August, Google Cloud is preparing to make Vera Rubin instances generally available, and at GTC Berlin last week NVIDIA showed a shipping container that holds seven of its racks. Here is what Vera Rubin is, how it compares with Blackwell and why it matters for the AI chip race.

What is NVIDIA Vera Rubin?

Vera Rubin is NVIDIA’s AI computing platform after Blackwell. It is named after the astronomer Vera Rubin, whose work provided key evidence for dark matter. The name combines two chips:

• Vera, NVIDIA’s new Arm-compatible CPU
• Rubin, the GPU that does the AI computation

NVIDIA now treats the whole rack as the product. Its technical blog describes Vera Rubin as six new chips designed together: the Vera CPU, the Rubin GPU, the NVLink 6 switch, the ConnectX-9 SuperNIC, the BlueField-4 DPU and the Spectrum-6 Ethernet switch. For background on how this fits the wider contest with Google’s TPUs, see our explainer on Google’s Ironwood and Axion chips.

Vera Rubin specs that matter

The headline figures from NVIDIA’s Vera Rubin NVL72 product page are below. The inference numbers are sparse figures and the training numbers are dense, as NVIDIA states.

The Vera CPU has 88 custom “Olympus” cores and 176 threads, using what NVIDIA calls Spatial Multithreading. It supports up to 1.5 TB of LPDDR5X memory and has 1.8 TB/s NVLink-C2C links to the GPUs, double the 900 GB/s of the Grace CPU it replaces.

Vera Rubin vs Blackwell

NVIDIA’s technical blog compares the two generations directly:

• Transistors: 336 billion on Rubin, against 208 billion on Blackwell
• NVFP4 inference: 50 PFLOPS on Rubin against 10 PFLOPS on Blackwell. The Rubin figure is for its Transformer Engine
• Memory: HBM4 doubles the interface width of HBM3e, and NVIDIA says Rubin “nearly triples” memory bandwidth compared with Blackwell
• Agent throughput: in its May 2026 production announcement, NVIDIA claimed 10x the agent throughput at scale of Grace Blackwell

These are vendor claims. Independent benchmarks will matter more once systems are running in customer data centres.

Memory is the main constraint. Rubin’s HBM4 competes for the same DRAM production capacity as phones and laptops, which is one reason we looked at the 2026 RAM shortage and consumer prices.

Where Vera Rubin stands right now

Shipping since August

NVIDIA’s Q2 FY27 investor presentation says the company “commenced production shipments of Vera Rubin in early August”. In the same quarter NVIDIA reported revenue of $96.2 billion, up from $46.7 billion a year earlier.

Google Cloud A5X

On 21 October, NVIDIA and Google Cloud announced that A5X instances built on Vera Rubin NVL72 will become generally available next month, and that availability will expand to Europe. Google Cloud plans to scale its Vera Rubin infrastructure to as many as 960,000 Rubin GPUs across multiple sites.

The “AI factory in a container”

At GTC Berlin, NVIDIA unveiled the DSX Pod AI Factory, which puts seven Vera Rubin racks inside a container that can be moved by truck. CNBC reported that it weighs about 28,000 kg and arrives with the cabling and liquid cooling already installed. NVIDIA says it cuts the time for a new AI data centre to start serving tokens from six months to 14 days. NVIDIA gave no price; analysts cited by CNBC estimate that a single Vera Rubin rack can cost more than $7 million. Deliveries are expected to start in early 2027.

Why Vera Rubin matters beyond NVIDIA

Revenue per gigawatt

Power, not chips, is increasingly what limits AI data centres. On NVIDIA’s August earnings call, Jensen Huang put NVIDIA’s revenue opportunity at about $18 billion per gigawatt for Hopper, $25 billion for Grace Blackwell and about $40 billion for Vera Rubin, as The Motley Fool reported. That is a sales opportunity, not booked orders, but it explains why every new generation brings a fresh wave of capital spending.

Who pays for it

Racks costing several million dollars each, multiplied across hundreds of thousands of GPUs, need very large amounts of financing. We covered that in who is paying for the AI boom. It also explains why markets react so strongly to every NVIDIA update, as we noted in why investors are nervous despite the AI boom.

What it means for India

Most Indian companies will use Vera Rubin through the cloud rather than buy racks. As Google Cloud and other providers roll out Rubin capacity in more regions, Indian AI startups and IT services firms should get access to more capable hardware for training and running agentic AI.

What to watch next

Three things will show whether Vera Rubin lives up to its claims. The first is independent performance results once customers run production workloads. The second is whether the supply of HBM4 memory keeps pace with GPU orders. The third is how quickly cloud regions outside the US, including in Asia, get Rubin capacity. We will update this explainer as those details come out.

Frequently asked questions

When was NVIDIA Vera Rubin released?

NVIDIA says it began production shipments of Vera Rubin in early August 2026. Cloud availability is rolling out from late 2026, with Google Cloud’s A5X instances due to reach general availability in November.

How much memory does a Rubin GPU have?

Up to 288 GB of HBM4 per GPU. A full NVL72 rack has 20.7 TB.

Is Vera Rubin faster than Blackwell?

On NVIDIA’s figures, yes. It quotes 50 PFLOPS of NVFP4 inference per Rubin GPU against 10 PFLOPS for Blackwell, and nearly triple the memory bandwidth. Real-world gains will depend on the workload.

How much does a Vera Rubin rack cost?

NVIDIA has not published prices. Analysts quoted by CNBC estimate more than $7 million per rack.

What comes after Vera Rubin?

NVIDIA’s public roadmap lists Rubin Ultra and then the Feynman generation, but it has not released detailed specs or firm dates for them.