NVIDIA GTC to reshape AI hardware

NVIDIA’s Vera Rubin NVL72 is a rack-scale design that unifies 72 Rubin GPUs with 36 Vera CPUs and advertises up to 3.6 exaFLOPS of NVFP4 inference and a 260 TB/s NVLink fabric. videocardz.com Microsoft Azure said it became the first cloud provider to “bring up” and begin validating a Vera Rubin NVL72 rack in mid‑March 2026, with CEO Satya Nadella posting the milestone on X. computing.net NVIDIA’s Rubin materials and third‑party tests claim up to a ~10x reduction in inference token cost versus prior-gen GB200 systems and roughly 4x fewer GPUs for MoE training in comparable workloads. videocardz.com NVIDIA’s December 2025 strategic deal for Groq’s inference technology—reported at roughly $20 billion—adds dedicated LPU ideas and talent that NVIDIA plans to fold into its inference roadmap for lower‑latency, high‑throughput serving. cnbc.com Adobe and major NLE vendors have already pushed GPU hooks: Adobe’s Premiere Pro community and TechPowerUp noted Blackwell‑era hardware acceleration for 10‑bit 4:2:2 H.264/HEVC formats, and Puget Systems’ benchmarks show measurable multi‑GPU scaling for DaVinci Resolve AI effects and exports. community.adobe.com OEMs and system integrators (Compal, Supermicro, AMAX, Dell) are exhibiting Rubin reference racks and three‑rack architectures at GTC and CES, signaling immediate channel availability for on‑prem deployment rather than a distant hyperscaler‑only roadmap. tmcnet.com Taken together, the NVL72 performance claims, Groq‑style inference acceleration, and early hyperscaler validation create a technical pathway for post houses to run heavier on‑site model serving, multi‑GPU Resolve/DaVinci grading farms, and denser batch renders with cost‑per‑token and throughput improvements documented by NVIDIA and independent reviewers. videocardz.com