Nvidia unveils Groq 3 inference chip and Vera CPU racks at GTC 2026, taking aim at Intel

Nvidia opened its GTC 2026 conference with new silicon meant to tighten its grip on AI infrastructure and blunt inroads by Intel and AMD. The company introduced the Groq 3 language processing unit—its first dedicated inference chip since striking a $20 billion deal to license Groq’s technology and hire key executives—and an LPX rack that packs 128 Groq 3 processors. Paired with Nvidia’s Vera Rubin NVL72 system, Nvidia says customers can achieve up to 35x higher throughput per megawatt and 10x greater revenue potential for large-scale inference, targeting trillion-parameter models and million-token contexts.
Nvidia also separated its Vera CPU from the Vera Rubin “superchip,” rolling out a standalone, liquid‑cooled Vera CPU rack with 256 chips per system—an overt challenge to Intel’s long-held data‑center franchise. Executives pitched a hybrid approach that marries the Groq 3 LPU’s memory speed with Nvidia GPUs’ capacity to keep inference costs and power in check as AI shifts from training to serving. The moves underscore Nvidia’s strategy to defend its data‑center lead as hyperscalers press for faster, cheaper inference and rivals push specialized silicon.

Read more