Nvidia introduced the new computing platform Vera Rubin

Core Features of the Nvidia Vera Rubin Architecture

At the GTC 2026 event, Nvidia introduced Vera Rubin, a next-generation computing platform consisting of seven different chips. Intended to replace Blackwell systems, the platform delivers up to 10 times more inference efficiency per watt while reducing processing cost per token by one tenth.

According to Nvidia’s data, the new architecture can perform the same task for training large mixture-of-experts (MoE) models with only one quarter of the GPUs required by Blackwell systems. The infrastructure is supported not only by major cloud providers such as Amazon Web Services, Google Cloud, Microsoft Azure, and Oracle Cloud, but also by AI developers such as OpenAI, Anthropic, and Meta.

Seven-Chip Unified Supercomputer Structure

Optimized especially for agentic (autonomous agent) AI systems and reinforcement learning, Vera Rubin consists of five different rack-scale systems that operate like a unified supercomputer. The system’s hardware infrastructure includes the following core components:

NVL72 Rack: Brings together 72 Rubin GPUs and 36 Vera CPUs in a single unit, communicating with each other over NVLink 6 switches.
Vera CPU Rack: Houses 256 liquid-cooled processors supporting more than 22,500 concurrent environments. The processors run with LPDDR5X memory offering 1.2 terabytes per second of bandwidth and 88 custom Olympus cores.
Groq 3 LPX Rack: Integrates 256 dedicated inference processors with 128 GB of on-chip SRAM to meet the low-latency demands of trillion-parameter models.
Networking and Storage (BlueField-4 STX / Spectrum-6 SPX): While high-speed storage is provided with BlueField-4 DPU for the massive key-value caches produced by agentic systems, Spectrum-6 delivers five times higher optical power efficiency than conventional hardware.

The Vera Rubin platform is planned to be made generally available in the second half of 2026 through leading cloud service providers and hardware manufacturers.