NVIDIA Vera Rubin

Vera Rubin is NVIDIA's platform after Blackwell: the Rubin GPU with 288 GB of HBM4, the Vera CPU with 88 cores, and the racks and servers built from them. The rack is the Vera Rubin NVL72, with 72 Rubin GPUs and 36 Vera CPUs, and NVIDIA's product page listed it as available now when I read it on 2026-11-12. The eight-GPU server board is the HGX Rubin NVL8. QuantaCloud does not rent Rubin by the hour. We build Rubin servers and racks to order through a capacity brief, and what you can start on today is a B300 build or an on-demand RTX PRO 6000 Blackwell.

Rubin servers and racks built to order#

Each Rubin build is ordered against your brief. Tell us whether you want eight-GPU HGX Rubin NVL8 servers or Vera Rubin NVL72 racks, how many, the network, storage, location, start date and term, and QuantaCloud replies in writing with a configuration, a lead time and commercial terms. We discuss rack-scale builds before we quote them, because one Vera Rubin NVL72 needs liquid cooling and up to 227 kW. The reserved capacity page lists what to include.

The names NVIDIA uses today#

The platform is Vera Rubin, the GPU is Rubin and the CPU is Vera. The press has used several other names, so these are the ones on NVIDIA's product pages and its October 2026 datasheet, read on 2026-11-12.

The NVL144 name came from counting dies. The Register reported that NVIDIA said at GTC 2025 it would count each GPU die, which turned 72 packages into 144, and that by CES in January 2026 it had gone back to counting packages. The rack never changed. NVL144 now means something else: in NVIDIA's March 2026 blog it is a future system with 144 Rubin Ultra GPUs, on a new rack called Kyber.

Rubin GPU specs#

The Rubin GPU has 288 GB of HBM4 and 35 PFLOPS of dense NVFP4, and it brings back the FP64 that Blackwell Ultra cut. These are NVIDIA's figures from the Vera Rubin datasheet (October 2026) and the Vera Rubin NVL72, HGX and DGX Rubin NVL8 pages, read on 2026-11-12. NVIDIA quotes NVFP4 inference with sparsity and the other Tensor Core figures dense.

Three figures differ between NVIDIA's own documents. The chip's peak memory bandwidth is up to 22 TB/s in NVIDIA's July 2026 technical blog and in the per-GPU table on its HGX page. The Vera Rubin NVL72 page gives 19.2 TB/s per GPU and 1,400 TB/s per rack, and the 130 TB/s of an HGX Rubin NVL8 is about 16 TB/s per GPU (our calculation: 130 / 8). NVLink was 3.6 TB/s per GPU and 260 TB/s per rack in the January press release, and the October datasheet says 3 TB/s and 216 TB/s. Eight GPUs at 288 GB make 2.3 TB, where NVIDIA lists 2 TB for the NVL8. I plan on the lower figure in each case.

The 288 GB itself needs the same caution. NVIDIA's MIG guide calls the B300 a 288 GB product as well, yet its largest MIG profile is 269 GB and the Blackwell Ultra datasheet says 270. I have found no nvidia-smi output from a Rubin GPU, so I do not know how much of the 288 GB software will see.

The 50 PFLOPS inference figure is the one to read carefully. It is a sparse figure that depends on the adaptive compression in Rubin's Transformer Engine, and the dense figure for the same format is 35.

The Vera CPU#

The Vera CPU is NVIDIA's own Arm processor, and it replaces Grace as the host in Rubin systems. It has 88 Olympus cores designed by NVIDIA, 176 threads, up to 1.5 TB of LPDDR5X at up to 1.2 TB/s and an NVLink-C2C link of 1.8 TB/s to the GPUs, against 72 Neoverse V2 cores and up to 480 GB on a Grace. A Vera Rubin NVL72 carries 36 of them, one for every two GPUs. NVIDIA also sells Vera without GPUs: a liquid-cooled Vera CPU rack holds 256 CPUs for the sandboxes and tool calls that agents run beside the GPUs.

For a buyer the practical point is the instruction set. On a Vera host every container image and Python wheel has to exist for arm64, as on a Grace system. An HGX Rubin NVL8 can keep an x86 host instead, and NVIDIA's own DGX Rubin NVL8 uses two Intel Xeon 6776P processors, the same CPUs as the DGX B300.

Racks and servers, and what the room needs#

The Vera Rubin NVL72 asks more of the room than any Blackwell rack. NVIDIA's datasheet gives one rack three power figures. The first is 227 kW at MaxP, the highest GPU power setting and the one NVIDIA says to size rack power and rack-side cooling for. The second is 187 kW at MaxQ, the setting with the best performance per watt. The third is 156 kW under MaxLPS, NVIDIA's scheme for allocating power dynamically across a whole site. A GB300 NVL72 is planned for up to 142 kW. The Rubin rack is 100% liquid-cooled with a warm-water inlet of 45°C, holds 18 compute trays and 9 NVLink switch trays in 48 rack units, and weighs roughly 4,000 lb. Each compute tray carries two superchips, eight ConnectX-9 SuperNICs and one BlueField-4 DPU, with no cables, hoses or fans inside the tray. NVIDIA's datasheet gives no power limit for a single Rubin GPU. Its projections for the four-GPU NVL4 assume 1,800 W per GPU, against limits of 1,100 W for a B300 and 1,400 W for a GB300.

The eight-GPU form is easier to place, but it is no longer air-cooled. NVIDIA's DGX Rubin NVL8 is a 2 RU liquid-cooled system that draws about 24 kW, where the DGX B300 is a 10 RU air-cooled system at about 14 kW. NVIDIA says one of its MGX racks can house up to 256 Rubin GPUs as HGX Rubin NVL8 systems, which is 32 servers (our calculation: 256 / 8). We build NVL8 servers as dedicated GPU servers or join several into a GPU cluster.

What NVIDIA has said about shipping#

NVIDIA lists the Vera Rubin NVL72 as available now. These are its own statements, in order.

None of this is a delivery date for an order placed today, and I do not publish one. Lead time on a Rubin build depends on supply when the order is placed, and QuantaCloud confirms it in writing before you commit.

Vera Rubin next to B300 and GB300#

Rubin changes bandwidth and compute far more than it changes memory.

On these tables memory per GPU rises by only 9 to 18 GB, and the usable gap may be smaller still, so Rubin is not the answer to a model that is too big for a B300. Bandwidth rises about 2.4 to 2.5 times, dense FP4 about 2.3 to 2.5 times and dense FP8 about 3.5 to 3.9 times (our calculation from the rows above). The Vera Rubin vs B300 comparison goes through the decision.

Software floor: CUDA 13.5 and an R620 driver#

Rubin is compute capability 10.7, which CUDA calls sm_107, and it needs a newer stack than any Blackwell GPU. CUDA 13.4, from September 2026, supported Vera Rubin only as a developer preview that NVIDIA said was not intended for benchmarking, performance analysis or production. CUDA 13.5, from November 2026, is the first release NVIDIA describes as general availability for the architecture, and its notes say newly enabled platforms need an R620 or later driver. The first R620 data-center driver, 620.70.06, is dated 2026-11-04. Neither its list of supported systems nor NVIDIA's compute-capability table named Rubin when I read them on 2026-11-12. On Linux, Blackwell and later GPUs run only with the open kernel modules.

Frameworks are still catching up. PyTorch 2.14.0 added what its release notes call awareness of sm_107 for extension builds and Inductor, and the 2.15.0 notes do not mention Rubin. vLLM added a Rubin build path on CUDA 13.4 in 0.30.0, nightly Rubin images in 0.31.0 and a kernel library with initial Rubin support in 0.32.0, released 2026-10-18. Rubin sits in the same major version as the B200 and B300, and NVIDIA's rule is that a binary runs on a GPU with the same major version and an equal or higher minor, so plain sm_100 code should carry over. Kernels built for an architecture-specific target such as sm_103a do not. I would confirm both on the delivered system:

nvidia-smi --query-gpu=name,compute_cap,driver_version --format=csv
python3 -c "import torch; print(torch.__version__, torch.version.cuda, torch.cuda.get_arch_list())"

The first should report 10.7 and a 620 driver or newer. The driver and CUDA version guide explains each field.

What Rubin costs#

Rubin has no public list price, and the two dated estimates I have are for the rack. A Wolfe Research note covered by Investing.com on 2026-01-30 saw Rubin racks at $5 million to $6 million. On 2026-03-24 Tom's Hardware quoted a source with knowledge of the matter at $5 million to $7 million for a Vera Rubin NVL72, including about $1 million of flash storage, against $6 million to $6.5 million for an inference-optimized GB300 NVL72. The same source warned that many quotes online are manufacturer prices without a proper warranty, and the article notes that NVIDIA has never confirmed a list price for its NVL72 racks. I have found no dated price for an HGX Rubin NVL8 server from a source I would name. QuantaCloud quotes Rubin builds in writing.

What you can start on today#

You can start on Blackwell today and move the work later. NVIDIA's HGX page marks HGX B300 servers as shipping now, with nearly the same memory per GPU as Rubin and compute capability 10.3, and NVIDIA's own DGX B300 is air-cooled. NVIDIA's datasheet also calls Rubin's Transformer Engine fully compatible with Blackwell, which is what makes a later move plausible. For prototyping by the hour, the RTX PRO 6000 Blackwell has 96 GB and runs FP4 checkpoints. It is compute capability 12.0, so kernels you compile on it will not carry over to Rubin or to a B300.

Prices checked 14 Nov 2026, 09:33 UTC

On-demand instances are Ubuntu VMs in our US regions, and stopping one deletes its disk, so copy results off first. The deploy guide covers the console steps.

Vera Rubin questions#

Is Vera Rubin a GPU or a CPU?

It is both, as a platform. Rubin is the GPU and Vera is the CPU. One Vera CPU and two Rubin GPUs form a Vera Rubin Superchip, and 36 superchips form a Vera Rubin NVL72 rack. NVIDIA named the platform after the astronomer Vera Florence Cooper Rubin.

How much memory does a Rubin GPU have?

A Rubin GPU has 288 GB of HBM4 on NVIDIA's tables. That is 576 GB per superchip and 20.7 TB per NVL72 rack, and NVIDIA lists 2 TB for the eight GPUs of an HGX Rubin NVL8. How much of it software sees has not been published.

What is the Vera Rubin release date?

NVIDIA launched the platform on 2026-01-05, said on 2026-05-31 that production shipments would begin in the fall, and its product page listed the Vera Rubin NVL72 as available now when I read it on 2026-11-12. Its HGX page has no shipping statement for HGX Rubin NVL8 servers yet.

What is VR200?

VR200 is shorthand for the Vera Rubin superchip generation, by analogy with GB200. The Register wrote in January 2026 that the superchip should bear that code name, and Tom's Hardware writes "NVL72 VR200" for the rack. NVIDIA's product pages and datasheet do not use it.

What happened to Rubin CPX?

Rubin CPX is a GPU with 128 GB of GDDR7 for long-context processing that NVIDIA announced on 2025-09-09 for the end of 2026. It is not on NVIDIA's Vera Rubin pages or in the October 2026 datasheet, where the inference accelerator beside the NVL72 is the Groq 3 LPX rack. Tom's Hardware reported on 2026-03-17 that it was missing from NVIDIA's GTC 2026 roadmap slides. I would not plan a build around it.

What is Rubin Ultra?

Rubin Ultra is the GPU NVIDIA has announced to follow Rubin. Its March 2026 blog describes Vera Rubin Ultra NVL576, eight racks of 72 Rubin Ultra GPUs in one NVLink domain, and the Kyber rack with 144 GPUs. The blog gives no specifications or dates, and NVIDIA's product pages do not list either system.

Can I rent Vera Rubin by the hour?

Not on QuantaCloud. Rubin is reserved capacity that we build to order. The GPUs page shows what is on demand right now.

The rule I follow: order Rubin when memory bandwidth or dense compute limits the job, the site has liquid cooling with the power to match, and you can run a toolchain that reached general availability in November 2026. Build on B300 when you need air cooling or a settled software stack, or when memory is what limits you, since Rubin adds little of it. Either way, send the GPU count, form factor, network, storage, location, start date and term, and we will come back with a configuration, lead time and terms in writing.