GPU cloud · NVIDIA Vera Rubin · Coming soon
NVIDIA Vera Rubin is coming to FlexiCloud
NVIDIA’s generation after Blackwell: the Rubin GPU with 288 GB of HBM4, paired with the new Vera CPU. Register now and we will come to you first with pricing and availability.
Not yet bookable on FlexiCloud. Pricing to be announced. Specifications are NVIDIA’s published figures and may be revised.
Vera Rubin at a glance
- HBM4 per GPU, per NVIDIA
- 288 GB
- Next-gen interconnect
- NVLink 6
- Expected on FlexiCloud (our estimate)
- 2027
- Rack price vs GB300 NVL72 (Morgan Stanley estimate)
- About 2x
What changes with Rubin
The generation after Blackwell
Based on NVIDIA’s published specifications for Vera Rubin, which may still be revised. If you need compute sooner, the Blackwell B200 and B300 are the current generation.
HBM4 memory
The Rubin GPU moves to HBM4: 288 GB per GPU and up to 19.2 TB/s of memory bandwidth on NVIDIA’s figures, more than twice the B300’s 8 TB/s of HBM3e. More room and more bandwidth for the biggest models and longest context.
A big FP4 inference leap
NVIDIA rates each Rubin GPU at 50 petaflops of NVFP4 inference with sparsity (about 35 dense), a large step up from Blackwell. The target is lower cost per token at frontier scale. NVIDIA’s roadmap places Rubin Ultra in 2027; analysts report a slip to 2028 for its largest rack.
Vera CPU and NVLink 6
Rubin pairs with the new Vera CPU and sixth-generation NVLink. The rack-scale Vera Rubin NVL72 wires 72 Rubin GPUs and 36 Vera CPUs together as one accelerator, and the 8-GPU HGX Rubin NVL8 brings the same GPU to a standard server node.
Built on a 3nm-class process (reported)
Rubin is reported to be built on a 3nm-class process, a step on from Blackwell. At rack scale it is liquid-cooled, like the top Blackwell systems.
Same regions, same model
We intend to offer Rubin the way we offer every GPU: dedicated, single-tenant nodes with root access, across six regions (North America, the UK, Europe, India, Asia Pacific and the Gulf), with the hardware run by verified partner operators and the stack run by you.
Early list, no commitment
Rubin capacity is scarce, and the first racks are going to large cloud and AI-infrastructure partners. Registering interest puts you on our early list, so you hear first when pricing and allocation are set.
Be first to hear
Register early interest in Vera Rubin
Leave your name and an email or phone number. We will ask about GPU count, workload and region when we reply, and come to you first with Rubin pricing and availability. If your timeline is tight, we can get you onto Blackwell in the meantime.
Where it stands
Honest status, so you can plan
Rubin is new. Here is what is known, when to expect it, and what registering gets you, with nothing oversold.
What is confirmed
In full production
NVIDIA says Vera Rubin is in full production, and its August 2026 results reported racks already running at cloud and AI-infrastructure partners. The figures on this page come from NVIDIA’s published specifications and may be revised.
- Rubin GPU with 288 GB of HBM4
- Paired with the new Vera CPU
- NVLink 6, the NVL72 rack and the 8-GPU HGX Rubin NVL8
Timeline
We expect wider availability in 2027
The first racks are running at large cloud and AI-infrastructure partners now. We expect wider availability through our partner network in 2027. That is our estimate, not an NVIDIA date: exact timing depends on NVIDIA’s allocation, and we confirm it in writing once it is firm.
- Rubin: shipping now, to large partners first
- Wider availability: 2027 (our estimate)
- Rubin Ultra: 2027 on NVIDIA’s roadmap; its largest rack reportedly slips to 2028
How to register
Register now, no commitment
Registering interest costs you nothing and commits you to nothing. It puts you on our early list, so you get the first quote, and an allocation slot where we can secure one, before general availability.
- First to get pricing and availability
- Early access to launch allocation where we can secure it
- We can bridge you onto Blackwell meanwhile
Questions
What people are asking about Rubin
Vera Rubin is NVIDIA’s GPU platform after Blackwell, named after the astronomer Vera Rubin. It pairs a new GPU called Rubin, with 288 GB of HBM4 memory, with a new CPU called Vera. At rack scale it is the Vera Rubin NVL72, which connects 72 Rubin GPUs and 36 Vera CPUs over NVLink 6 as a single accelerator. NVIDIA’s earlier roadmap called the same rack NVL144, counting GPU dies rather than GPU packages. There is also an 8-GPU server form, the HGX Rubin NVL8.
NVIDIA says Vera Rubin is in full production, and its first racks are running at large cloud and AI-infrastructure partners. We expect wider availability through our partner network in 2027. That is our estimate, not an NVIDIA date, and it depends on NVIDIA’s allocation. Rubin Ultra follows on NVIDIA’s roadmap in 2027, though analysts report a slip to 2028 for its largest rack. Register your interest and we will tell you the moment we have firm FlexiCloud availability.
Our price is to be announced: we will quote once our partners confirm allocation and pricing. For scale, NVIDIA publishes no prices, but analysts estimate a Vera Rubin NVL72 rack at about US$7.8 million (Morgan Stanley, May 2026) to US$9.1 million (Bernstein, June 2026); Morgan Stanley’s figure is about twice its US$4 million estimate for a GB300 NVL72. SemiAnalysis estimated three-year rental rates for Rubin at over US$8.50 per GPU-hour in July 2026, against about US$5 for a GB300 NVL72. If you need GPUs before Rubin, our B300 nodes start from US$4.50 per GPU-hour on our longest term; a one-year term costs more, and prices are before tax.
The B200 and B300 are the current Blackwell generation, with up to 288 GB of HBM3e. Rubin is the next generation: 288 GB of HBM4 at more than twice the B300’s memory bandwidth, a large jump in FP4 inference performance on NVIDIA’s figures, and NVLink 6. If you need compute today, Blackwell is the answer; Rubin is for planning a 2027 buildout.
Rubin is the first part of the generation, and NVIDIA says production shipments began in August 2026. Rubin Ultra is the higher-end follow-on, which NVIDIA’s March 2026 roadmap places in 2027 with more memory per GPU. Analysts have since reported design changes and a slip to 2028 for its largest rack; NVIDIA says its roadmap is intact. We intend to offer both as they become available.
Not yet: nobody can book Rubin through us today. Register your interest through the form on this page instead. There is no cost and no commitment. It puts you on our early list for the first quote and early access to launch allocation where we can secure it, and we can put you on Blackwell in the meantime if your timeline is tight.
Go deeper
While you wait for Rubin
What Rubin changes and when it matters, how the current generations compare, and how to size memory, so the Blackwell decision you make today still fits when Rubin lands.
Hopper vs Blackwell: NVIDIA’s GPU generations compared
NVIDIA’s Hopper (H100, H200) and Blackwell (B200, B300) generations compared: memory, architecture, interconnect, power and price, and a practical guide to which one actually fits your AI workload.
Read itVera Rubin explained: what changes and when it matters to buyers
NVIDIA Vera Rubin is shipping. Here is what changes from Blackwell, what analysts estimate the hardware costs, how the cable-free trays are built, whether it is the "wireless" server Jensen Huang talked about, how much faster it really is, and when it matters for your plans.
Read itB300 vs B200: which Blackwell node, and why the right one lowers your TCO
The B200 is 17% cheaper per GPU-hour; the B300 carries 60% more memory. Which is lower total cost is not about the sticker rate; it is about whether your workload fits in 180 GB, because memory decides how many GPUs and nodes you need. An honest B300-vs-B200 TCO guide.
Read itWill your model fit? A practical guide to GPU memory for AI
Whether your model fits is arithmetic, not a benchmark. The bytes-per-parameter maths for inference and training, the levers that shrink it (quantization, LoRA, checkpointing), and a back-of-envelope guide to size any job, mapped to what fits on a B300.
Read it
Get on the Rubin early list
Tell us the rough shape of what you are planning and we will come to you first with pricing and availability. Need compute sooner? We can quote Blackwell B300 and B200 nodes and confirm the lead time for your region.
Bringing Vera Rubin capacity online?
We bring you buyers we have checked, and you sign with them directly. Tell us the model, the count and where it runs.