Overview
- CEO Jensen Huang said purchase orders for Blackwell and Vera Rubin are expected to total about $1 trillion through 2027, up from roughly $500 billion previously projected through 2026.
- Nvidia unveiled a Groq‑based inference product, the Groq 3 LPU/LPX, which Huang said can accelerate decode workloads by up to 35 times, with systems expected to ship in the second half of 2026 and Samsung as manufacturer.
- The Vera Rubin platform is slated to roll out later this year, with Nvidia claiming about 10x better performance per watt than Grace Blackwell in a 1.3 million‑component system and a role focused on the prefill stage of inference.
- Huang previewed the Feynman roadmap targeted for 2028 and introduced a new Vera CPU, saying standalone CPU sales already represent a multibillion‑dollar business for Nvidia.
- Shares rose roughly 1%–2% during the keynote as Huang declared that the inference inflection has arrived and highlighted Nvidia’s CUDA software moat, even as rivals and hyperscalers advance custom inference chips.