FULL STORY

NVIDIA Vera Rubin: From Platform Reveal to Real-World Sighting

NVIDIA unveiled the Vera Rubin platform and detailed its hardware, culminating in the real-world sighting of a massive NVL72 rack in Finland.

2026-07-21 ~ 2026-07-24 · 4 episodes · 19 posts

Episode 1 · NVIDIA Delivers 102.4 Tbps Spectrum-6 Switches for AI Factories (2026-07-21, 3 posts)

NVIDIA has begun delivering its Spectrum-6 Ethernet switches with 102.4 Tbps bandwidth to global AI factories. Designed for the Vera Rubin platform, the switches aim to support the networking demands of next-generation gigascale AI data centers.

Episode 2 · NVIDIA Launches Vera Rubin with 10x Energy Efficiency (2026-07-21, 6 posts)

NVIDIA has officially launched the Vera Rubin platform, designed for the Agentic era to significantly reduce inference costs for large-scale AI models by maximizing compute efficiency. The Vera Rubin NVL72 is currently being rolled out with global partners including CoreWeave, Google Cloud, and Microsoft.

Confirmed

NVIDIA claims that Vera Rubin delivers 10x better performance per watt. According to initial real-world performance data obtained by CoreWeave, the Vera Rubin NVL72 achieves 10x more token generation per megawatt compared to existing Blackwell architectures (like the GB200) when running the DeepSeek-R1 model. Furthermore, NVIDIA executive Kyle Kranen disclosed the design goals for the next-generation Rubin architecture, aiming to maximize "agentic tokens/watt." The core hardware specifications include a compute target of 50 Petaflops and HBM4 memory with a bandwidth of 22TB/s.

Why it matters

As AI models scale and agentic applications become widespread, the compute and energy costs during inference have become a primary bottleneck. By delivering a 10x improvement in performance per watt, Vera Rubin is poised to drastically reduce the operational costs of large-scale AI and accelerate the adoption of highly energy-efficient infrastructure.

Episode 3 · NVIDIA Reveals More Vera CPU Details Ahead of AMD Event (2026-07-22, 8 posts)

NVIDIA lifted the veil on more of its next-generation server CPU, Vera, just ahead of AMD’s Advancing AI event. The disclosed specs point to a design centered on agentic AI and inference system economics rather than standalone CPU benchmark positioning: 88 in-house Olympus cores, 176 threads, up to 1.5TB of LPDDR5X, and 1.2TB/s of memory bandwidth. The announcement matters because it shows NVIDIA continuing to build out its own data-center CPU strategy and tying CPU design closely to GPU-led AI systems.

Confirmed

Across the posts, NVIDIA is described as having released fuller technical details for Vera ahead of AMD’s event. According to @ryanshrout’s post, Vera uses 88 Olympus cores—described as NVIDIA’s first in-house data-center CPU core—with 176 threads via spatial multithreading. A single chip is said to support up to 1.5TB of LPDDR5X memory and 1.2TB/s of bandwidth, and NVIDIA claims the architecture can scale to 22,000 cores at rack level. @BenBajarin’s notes say the briefing emphasized how Vera fits agentic inference workflows and their economics, rather than focusing on raw CPU benchmark comparisons. A reposted summary also said NVIDIA highlighted early customer interest, including OpenAI.

Why it matters

@BenBajarin argues Vera shows NVIDIA continuing to bet on a monolithic CPU design instead of a chiplet approach, on the view that this better suits future agentic workloads. @ryanshrout frames the disclosure as a direct challenge to long-standing x86 data-center design assumptions, and as part of a broader shift in discussion from simply adding more cores toward renewed attention on single-core performance and system-level coordination.

Episode 4 · First NVIDIA Vera Rubin NVL72 Racks Spotted in Finland (2026-07-23, 2 posts)

Nebius showcased the first complete NVIDIA Vera Rubin NVL72 rack at its Finnish data center. Photos reveal the massive scale of the system, packing 72 GPUs into a single cabinet with dense cabling, liquid cooling, and Spectrum-6 switches.