Nvidia's Vera Rubin platform: 336B transistors, 5x Blackwell inference

Nvidia announced Vera Rubin at GTC 2026 on March 16. Built around the Rubin R100 GPU and a custom 88-core Vera CPU, the platform is the next architectural step after Blackwell [CNBC].

── What shipped ──

Rubin R100 GPU specs:

• TSMC 3nm process
• Dual-die design with two reticle-sized compute chiplets
• 336 billion transistors combined (1.6x Blackwell's 208B)
• 288GB HBM4 memory delivering 22 TB/s bandwidth (nearly 3x Blackwell's 8 TB/s on HBM3e)

Vera CPU: 88-core custom Arm-based design tightly coupled to Rubin GPUs.

Performance claims: 5x the inference performance of Blackwell at 10x lower cost per token [Nvidia newsroom].

── Why it matters ──

The performance-per-token claim is the most consequential. AI inference cost is the rate-limiting input for every commercial AI product — from API pricing to product margins. A genuine 10x cost-per-token improvement at the silicon layer would let frontier providers cut prices materially while preserving margin, or hold prices and absorb the saving.

The catch is that Nvidia's per-token claims are typically measured on best-case workloads (large batch sizes, optimised quantisation, specific model shapes). Real-world cost reductions for diverse production workloads will be smaller — historically 30–60% of the headline number. Even at the low end, that is a step change.

The HBM4 transition is the underrated piece. Inference on long-context models is increasingly memory-bandwidth-bound, not FLOP-bound. The 22 TB/s figure addresses the actual bottleneck for current-generation frontier models.

For AI infrastructure buyers (clouds, neoclouds, sovereign-AI programs), Rubin compresses the upgrade cycle. Blackwell deployments are still scaling out across hyperscalers; Rubin lands during that rollout. Expect 2026 to be a transition year with mixed Blackwell/Rubin clusters.

── Editor's take ──

Rubin is on time, on spec, and on the trajectory Nvidia laid out at last year's GTC. There is no architectural surprise here — the surprise would have been any slip. The competitive question is whether AMD's MI400 series and the various sovereign-chip efforts (Huawei Ascend, Tenstorrent, Trainium 2) close any of the gap. Independent benchmarks land in Q3.