NVIDIA Vera Rubin
Vera Rubin is the platform after Blackwell Ultra: a Rubin GPU paired with a Vera CPU. Published NVL72 figures put 288 GB of HBM4 on each GPU at up to 22 TB/s, with NVFP4 inference around 50 PFLOPS and NVLink 6 at 3.6 TB/s. Oracle Cloud made the NVL72 generally available on October 27, 2026. The marketplace row is an estimate, because public rental indexes still had no on-demand rate.
Implante NVIDIA Vera Rubin do seu jeito: pods Docker ou VPS KVM root
A NovaPod deixa você escolher como usar a GPU. Um contêiner já em cache, uma máquina virtual root com IP público ou um cluster NVLink de vários nós: você paga só os segundos em que roda, em cripto.
Contêineres de IA leves e rápidos
Computação de GPU sem o peso do hipervisor. Entre direto no PyTorch, vLLM, ComfyUI ou JupyterLab.
- Medição com precisão de 1 segundo
- Mais de 37 modelos de framework de ML já em cache
- Sem taxa ociosa quando a instância está parada
Máquina virtual root completa
Sistema dedicado (Ubuntu 24.04 LTS / Windows GPU Server) com SSH root e sudo desde o primeiro minuto.
- IPv4 público dedicado (todas as portas abertas)
- Pause quando quiser; pague só US$ 0,00014/GB/h pelo disco
- Rode Docker-in-Docker, k3s e drivers próprios
Supercomputação e treino distribuído
Junte até 8× NVIDIA Vera Rubin em um nó com InfiniBand Quantum-2 RDMA para escalar a vários nós.
- Troca de gradiente all-reduce sem perda
- Pronto para PyTorch FSDP e DeepSpeed Zero-3
- GPUDirect Storage nativo e malha RDMA
Especificações de hardware
- Platform
- Rubin GPU + Vera CPU
- Also called
- R100 / R200 in supply-chain notes
- VRAM
- 288 GB HBM4
- Memory bandwidth
- Up to 22 TB/s
- FP4 inference
- About 50 PFLOPS NVFP4
- Dense FP8 / FP6
- About 17.5 PFLOPS
- Interconnect
- NVLink 6 at 3.6 TB/s
- Networking
- ConnectX-9, about 1.6 Tb/s
- System
- Vera Rubin NVL72, 72 GPUs
- Status
- First cloud cohort targeted for H2 2026
- Wider supply
- 2027, if the public timeline holds
- Rental ask
- None published yet
What Vera Rubin is
Vera Rubin is NVIDIA's name for the platform that pairs a Rubin GPU with a Vera CPU. Supply-chain notes also call the GPU R100 or R200. It is the generation after Blackwell Ultra, not a refresh of H200. NVIDIA showed the architecture at GTC and has described first cloud samples in the second half of 2026, with wider availability in 2027.
The numbers that matter are still marked preliminary. Each GPU is specified at 288 GB of HBM4, up to 22 TB/s of memory bandwidth, about 50 PFLOPS of NVFP4 inference, about 17.5 PFLOPS of dense FP8 or FP6, NVLink 6 at 3.6 TB/s, and ConnectX-9 networking around 1.6 Tb/s. Transistor count and a final board power are not published, so this page does not invent them.
Why the bandwidth is the story
B300 already has 288 GB. Rubin does not win by adding capacity. It wins if the published 22 TB/s holds, which is about 2.75 times B300's 8 TB/s. Large-language-model decode is usually limited by how fast weights and cache can be read, so that gap is the reason teams with flexible timelines are watching the part.
A single GPU at 50 PFLOPS FP4 and 288 GB is aimed at serving a 200B-class model in FP4 with cache left over, and at cutting the GPU count on 400B-class mixture-of-experts jobs that today need several B300s at FP8. Those are planning numbers, not a benchmark we ran.
The NVL72 rack
The product NVIDIA sells as a system is the Vera Rubin NVL72: 72 Rubin GPUs in one NVLink domain. Public rack math puts aggregate HBM in the tens of terabytes and NVLink 6 fabric well above the GB300 rack. Models that spill across racks on Blackwell are the ones that might fit in one Rubin rack.
Per-GPU figures on this page are the rack specification divided by 72. NVIDIA has said that table can change. Do not treat a blog multiplier as a purchase order.
What to run until it exists
There is no hourly ask for Vera Rubin on this marketplace. If the job has to start this quarter, use B300 for 288 GB HBM3e, B200 for 192 GB and FP4, or H200 when 141 GB of Hopper memory is the actual constraint.
Registering interest is not a reservation. When a host lists a Rubin offer, it will show up as a live ask next to the other cards, billed per second like everything else.
Comparação entre gerações
Per-GPU figures published for Vera Rubin are preliminary and come from the NVL72 rack spec divided by 72.
| Spec | Vera Rubin | B300 | B200 | H100 |
|---|---|---|---|---|
| Architecture | Rubin | Blackwell Ultra | Blackwell | Hopper |
| VRAM | 288 GB HBM4 | 288 GB HBM3e | 192 GB HBM3e | 80 GB HBM3 |
| Bandwidth | Up to 22 TB/s | 8 TB/s | 8 TB/s | 3.35 TB/s |
| FP4 | 50 PFLOPS | 15 PFLOPS | 9 PFLOPS | n/a |
| Interconnect | NVLink 6 · 3.6 TB/s | NVLink 5 · 1.8 TB/s | NVLink 5 · 1.8 TB/s | NVLink 4 · 900 GB/s |
| Cloud | First cohort H2 2026 | Shipping | Shipping | Shipping |
Armazenamento rápido, portas nativas e trilhos cripto
Cada instância NVIDIA Vera Rubin na NovaPod se conecta a uma infraestrutura de alto débito feita para tirar gargalos de E/S e a espera inicial.
Armazenamento de rede compartilhado a 25 Gb/s
Monte checkpoints e conjuntos de treino em uma ou várias instâncias ao mesmo tempo. Os dados duram além de um pod e continuam acessíveis quando as instâncias param.
JupyterLab, VS Code & SSH
Conecte do terminal local com chaves SSH comuns ou abra IDEs na web em abas do navegador, sem configurar proxy reverso.
Trilhos multi-chain sem custódia
Deposite USDT, USDC, SOL, ETH ou BNB no seu endereço pessoal. A confirmação on-chain leva segundos e não há verificação de identidade obrigatória.
Cargas indicadas para NVIDIA Vera Rubin
Trillion-parameter FP4 inference
288 GB of HBM4 is the same capacity as B300. The reason to wait is bandwidth, if the 22 TB/s figure holds, for decode that is already memory-bound.
Frontier pre-training
Higher FP4 throughput and bandwidth can shrink the node count versus an all-B300 run. Treat every multiplier as preliminary until NVIDIA freezes the table.
One NVL72 domain
The product is a rack of 72 GPUs on NVLink 6, not a PCIe card you drop into an existing server.
Quando escolher NVIDIA Vera Rubin
01
Wait only if bandwidth is the ceiling
If a late-2026 or 2027 start is acceptable and B300's 8 TB/s is the limit you already hit, Rubin is the next step. Capacity does not go up from 288 GB.
02
Run B300 if the job starts this year
B300 already has 288 GB of HBM3e and a live path on the marketplace when a host lists it.
03
Run B200 if 192 GB is enough
Most 70B–200B jobs do not need the Rubin bandwidth premium, and they can start now.
04
Do not treat nicknames as a second product
R100 and R200 refer to the Rubin GPU. Vera Rubin is the platform: that GPU plus a Vera CPU, sold as an NVL72 rack.
Where it sits
Hopper H100 and H200, then Blackwell B200, then Blackwell Ultra B300, then Rubin. Everything before Rubin can be rented now. Vera Rubin is the one you register for.
What the bandwidth is for
Decode speed on large language models tracks memory bandwidth. 22 TB/s is roughly 2.75× B300's 8 TB/s, which is why the part is discussed for memory-bound serving rather than as a drop-in for small models.
Perguntas frequentes
Is there a rental price?
When can I deploy one?
Is Vera Rubin the same as R100?
How does it compare with B300?
How does it compare with B200 and H100?
What is the NVL72?
Should I reserve capacity now?
Outras opções de hardware
Pronto para implantar NVIDIA Vera Rubin?
Inicie uma instância no mercado ao vivo em menos de 15 segundos. Abasteça o saldo com criptomoeda e pague por segundo, sem compromisso.
Iniciar no mercado →