NVIDIA Rubin

NVIDIA Rubin is an AI platform that NVIDIA announced at CES 2026 in January 2026. The company said the platform was in full production and described it as its first extreme-codesigned six-chip AI platform and the successor to the Blackwell architecture.^[1] NVIDIA’s press release announcing Rubin is dated January 5, 2026.^[2] Rubin therefore refers not to a single accelerator but to a coordinated set of chips and networking products intended to work as one system.

A six-chip platform

NVIDIA lists six components: Rubin GPUs, Vera CPUs, NVLink 6 scale-up networking, Spectrum-X Ethernet Photonics, ConnectX-9 SuperNICs and BlueField-4 DPUs.^[1] The list separates the platform into compute, host processing, scale-up interconnect, Ethernet optics, network interfaces and infrastructure processing. In a conventional server, these parts can come from different design cycles. NVIDIA’s “extreme-codesigned” description presents them as co-designed around the same platform. That distinction matters because large AI systems depend on how fast accelerators exchange data, not only on how fast one chip computes.

The GPUs are the compute center of the platform. The CPUs support the host side. NVLink 6 handles scale-up networking, the links among accelerators inside a system. Spectrum-X Ethernet Photonics addresses Ethernet networking; ConnectX-9 SuperNICs provide network interfaces; BlueField-4 DPUs handle infrastructure tasks. Together they show that Rubin is a platform offering, spanning chips that compute, connect and manage data movement.^[1]

Performance and cost claims

NVIDIA says Rubin GPUs deliver 50 petaflops of NVFP4 inference performance.^[1] Petaflops is a measure of computing throughput; NVFP4 is the format named in NVIDIA’s performance claim. The figure is an inference figure, not a training figure, which reflects where the platform is positioned. Inference is the stage at which a trained model generates answers, so throughput and cost per token directly affect how expensive it is to run an AI service.

NVIDIA also says Rubin targets AI token generation at roughly one-tenth the cost of the previous platform.^[1] A token is a unit of text that a language model reads or produces. Lower cost per token means a provider can serve the same amount of generated text for less, or serve more text for the same budget. The claim is a target and a company statement.

Availability through partners

NVIDIA said Rubin-based products will be available from partners in the second half of 2026.^[2] It named AWS, Google Cloud, Microsoft and OCI among the first cloud providers to deploy Vera Rubin-based instances in 2026. It also named NVIDIA Cloud Partners CoreWeave, Lambda, Nebius and Nscale among the first cloud providers to deploy those instances.^[2]

The partner list includes general cloud providers and NVIDIA Cloud Partners. For customers, cloud instances are often a main route to a new AI platform, so the deployment list is a concrete part of the launch. The stated schedule places Rubin’s partner availability after the January 2026 production announcement.

What Rubin changes

Rubin follows Blackwell as NVIDIA’s next architecture in its stated lineup.^[1] Its launch materials emphasize two ideas: the platform is composed of six chips designed together, and it is aimed at reducing the cost of token generation. Blackwell is the reference point for the cost claim, and Rubin’s six-part design is the main structural difference NVIDIA highlights.^[1]

The platform’s importance lies in that combination. AI systems are increasingly judged by the cost of serving models, not only by peak training performance. By tying GPUs, CPUs, scale-up networking, Ethernet photonics, SuperNICs and DPUs into one announced platform, NVIDIA is selling a complete design rather than an isolated processor.^[1] The concrete test will be the partner products and cloud instances scheduled for the second half of 2026.^[2]