NVIDIA Vera Rubin by the Numbers: Specs, Claims and First Results

What Vera Rubin is

Vera Rubin is NVIDIA’s data-center platform after Blackwell, named for the astronomer Vera Florence Cooper Rubin.⁴ NVIDIA launched it at CES on January 5, 2026 as “six new chips” designed as one system: the Vera CPU, the Rubin GPU, the NVLink 6 switch, the ConnectX-9 SuperNIC, the BlueField-4 data processing unit and the Spectrum-6 Ethernet switch.⁴ By October 2026 NVIDIA’s datasheet says “seven new chips” and describes five rack types that work together: the Vera Rubin NVL72, a Vera CPU rack, the Groq 3 LPX inference rack, BlueField-4 STX storage and Spectrum-6 SPX Ethernet.¹²⁰

The Rubin GPU has 336 billion transistors across two compute dies, against 208 billion for Blackwell, with 224 streaming multiprocessors, according to NVIDIA’s technical blog.⁵ Each carries 288 GB of HBM4, a high-bandwidth memory generation that NVIDIA says doubles the interface width of HBM3E and, with new memory controllers, nearly triples the GPU’s memory bandwidth over Blackwell.⁵ The Vera CPU has 88 NVIDIA-designed Olympus Arm cores and 176 threads, up to 1.5 TB of LPDDR5X at up to 1.2 TB/s, and a 1.8 TB/s NVLink-C2C link that lets the CPU and two GPUs share one coherent memory pool.¹⁵ Two Rubin GPUs and one Vera CPU make a Vera Rubin Superchip, and 36 of those fill the NVL72 rack, NVIDIA’s third generation of 72-GPU rack after the GB200 and GB300.¹

The headline per-GPU figure, 50 PFLOPS of NVFP4 (NVIDIA’s 4-bit floating-point format with a shared scale factor for every 16 values¹⁹), is a sparse figure that NVIDIA credits to a new Transformer Engine with “hardware-accelerated adaptive compression”; the dense figure is 35 PFLOPS.¹⁵ NVIDIA’s blog lists Blackwell at 10 PFLOPS of NVFP4 for both inference and training on the same table.⁵

The rack against Blackwell

Set side by side, NVIDIA’s own rack figures show where the generation moved most. By our calculation from the table, Vera Rubin NVL72 lists 2.3 times the dense NVFP4 compute of the GB300 NVL72 (2,520 against 1,080 PFLOPS) and 3.5 times the GB200 NVL72’s (720 PFLOPS); 2.4 times the GPU memory bandwidth (1,400 against 576 TB/s); and 1.7 times the NVLink bandwidth inside the rack (216 against 130 TB/s).¹⁶⁸ GPU memory capacity barely moves against the GB300: 20.7 TB of HBM4 against 20 TB of HBM3E, or 288 GB per Rubin GPU against 279 GB per Blackwell Ultra.¹⁸ CPU memory more than triples, to up to 54 TB, which NVIDIA pitches for offloading the key-value cache of long contexts.¹⁵

Bandwidth, rather than raw compute, is the figure NVIDIA leans on for inference. Its technical blog argues that reasoning, mixture-of-experts and long-context workloads “are not limited by peak floating point operations alone” but by memory and communication, since every token of an MoE model is routed across experts on different GPUs.⁵

Double precision goes the other way. The Vera Rubin rack lists 2,400 TFLOPS of FP64 against 2,880 for the GB200 NVL72, and NVIDIA’s blog puts Rubin’s FP64 vector rate at 33 TFLOPS per GPU against Blackwell’s 40. Its FP64 matrix figure, 200 TFLOPS, is reached through tensor-core emulation, which NVIDIA marks as such.¹⁵⁶ The GB300 rack lists 100 TFLOPS of FP64.⁷

What changed since January

Several Vera Rubin figures in NVIDIA’s October datasheet differ from the ones it gave at launch. The per-GPU NVLink figure fell from 3.6 TB/s to 3 TB/s and the rack’s from 260 TB/s to 216 TB/s, although NVIDIA’s HGX product page still lists 3.6 TB/s per Rubin GPU in its eight-GPU HGX Rubin NVL8 table and its July partner blog still cites the 260 TB/s fabric.¹³⁴¹⁰ Memory bandwidth per GPU is 22 TB/s in the January blog and on the HGX page, and 19.2 TB/s in the NVL72 rack.¹³⁵

NVIDIA hasn’t explained the changes in the pages we opened. The January column is NVIDIA’s CES press release and technical blog; the October column is the Vera Rubin datasheet and, where marked, the HGX product page. Where NVIDIA’s pages disagree, the main table uses the NVL72 datasheet.

NVIDIA’s claims for the AI industry

NVIDIA’s case that Vera Rubin will lift the industry is about the cost of producing tokens, and it rests on four headline claims, all its own projections. Against the GB200 NVL72, it says the Vera Rubin NVL72 delivers one-tenth the cost per million tokens and up to 10 times the tokens per megawatt, both modeled on Moonshot AI’s Kimi K2 Thinking with 32,000 input and 8,000 output tokens, and both marked “subject to change”.¹² For training, it says the rack needs one-fourth as many GPUs, projected for a 10-trillion-parameter MoE model trained on 100 trillion tokens in one month.¹ Paired with Groq 3 LPX racks, it claims up to 35 times the throughput per megawatt for trillion-parameter models.²

Three details matter when reading those numbers. The baseline for the 10× and 4× claims is the GB200 NVL72, the older of the two Blackwell racks, not the GB300.¹ The eight-GPU version’s “10x more token factory throughput versus HGX B200” compares, in NVIDIA’s own footnote, “HGX Rubin NVL8 with Sparse NVFP4, HGX B200 with Dense NVFP4”, so it sets a sparse figure against a dense one.¹³ And the claims are framed for agents: NVIDIA says agentic systems “consume up to 15x more tokens than traditional AI applications”, which is its own figure, not a measured industry average.²

NVIDIA also pitches the platform at power-limited sites. Its datasheet sizes a 100 MW AI factory at 40,000 Rubin GPUs with 12 PB of HBM4, using its MaxLPS power profile, and says the rack runs fully liquid-cooled at a 45°C coolant inlet, which it says allows dry coolers without chillers and lowers power usage effectiveness (PUE).¹¹⁰ The MGX 1.2 rack specification NVIDIA contributed to the Open Compute Project in October supports racks of up to 250 kW.¹⁵

First results: MLPerf and the first clouds

The first Vera Rubin results reviewed by a third party came in MLPerf Inference v6.1, published September 16, 2026, where MLCommons lists NVIDIA Rubin and Vera Rubin NVL72 in its preview category rather than as available systems.¹² Two organizations submitted Vera Rubin results: NVIDIA and the cloud provider Nebius.¹³

The four MLPerf rows are results MLCommons verified (entries 6.1-0107 for Nebius’s Vera Rubin system and 6.1-0080 for its GB300 NVL72), as Nebius lists them; they are labeled measured but were run by the submitter, as every MLPerf entry is.¹³ Per-GPU throughput is the result divided by the GPU count, our calculation (Nebius states the Vera Rubin server figure as 16,427 tokens/s per GPU). The two systems ran at different sizes, 36 and 72 GPUs, and the Vera Rubin entry is in the preview category, so the per-GPU comparison is indicative only. The other rows are ratios NVIDIA published, from its own MLPerf entries (6.1-0106 and 6.1-0074) or from partners’ tests whose method isn’t published; they are vendor-stated and don’t rank against the MLPerf figures.

On the MLPerf figures alone, Nebius’s 36-GPU Vera Rubin system served about 2.0 times the DeepSeek-R1 tokens per GPU of its 72-GPU GB300 NVL72 in the server scenario and about 1.6 times in the offline scenario, by our division of the published results.¹³ NVIDIA’s own entries put the gap higher, up to 2.5 times on DeepSeek-R1 and up to 3.7 times on Qwen3-VL, using disaggregated serving that splits the prefill and decode stages of inference across GPUs.¹¹ MLCommons says its next datacenter suite, MLPerf Endpoints, will replace MLPerf Inference.¹²

The first cloud results came through NVIDIA’s blog. NVIDIA says CoreWeave was the first AI cloud to bring up and validate the Vera Rubin NVL72 and measured 10 times the DeepSeek-R1 tokens per second per megawatt of a Grace Blackwell NVL72.¹⁰ On September 30, 2026 it said CoreWeave had received its first production racks that month, and that Cognition, maker of the Devin coding agent, saw up to 4.8 times the total token throughput of a GB200 NVL72 on a software-engineering workload.¹⁴ Neither test’s full method has been published.

Who is deploying it

At launch NVIDIA named AWS, Google Cloud, Microsoft and Oracle Cloud Infrastructure, plus CoreWeave, Lambda, Nebius and Nscale, as among the first clouds to deploy Vera Rubin-based instances in 2026, and listed AI labs including Anthropic, Meta, Mistral AI, OpenAI and xAI as looking to the platform.⁴ On May 31, 2026 NVIDIA said Vera Rubin was ramping into full production, with production shipments “set to begin starting this fall”.⁹ These are the deployments with a named customer and a public status.

“Available” and “Running” quote the operator’s or NVIDIA’s own wording; “Announced” means a commitment or plan with no public statement yet that the systems are running. Gigawatt figures are the companies’ own capacity terms, not measured power draw. We found no public update on the status of OpenAI’s first Vera Rubin gigawatt as of Oct 30, 2026.

What the numbers do and don’t show

What is documented is a large step in NVIDIA’s stated peak figures, most of all in memory bandwidth and 4-bit compute, and a first independent check: an MLPerf result showing roughly twice the per-GPU DeepSeek-R1 throughput of a GB300 rack, from a preview system. What isn’t documented yet is the industry-level claim. The one-tenth cost per million tokens and 10 times the tokens per megawatt are NVIDIA projections against the GB200, and the partner figures that echo them come through NVIDIA’s blog without a published method. We found no independent measurement of Vera Rubin’s cost per token as of Oct 19, 2026.

The platform also leans on the rest of the stack. Each NVL72 rack carries 72 GPUs with 288 GB of HBM4 apiece, so its ramp adds to demand for the memory covered in The Memory Crunch, by the Numbers.¹ Its power profiles, up to 227 kW per rack, set what a campus needs, the subject of Largest AI Data Centers by Estimated Power Draw.¹ And NVIDIA now sells CPUs for agents alongside the GPUs: its Vera CPU rack holds 256 Vera CPUs and, it says, supports more than 22,500 concurrent sandbox environments for tool calls and reinforcement learning, while CoreWeave’s deployment puts 128 Vera CPUs and 11,264 cores in a rack.²⁰¹⁴ Rubin’s 288 GB of HBM4 per GPU is also #1 (tied) in AI Accelerators With the Most HBM per Chip.