Chip Stock Investor Podcast

Beyond the GPU: Nvidia’s Secret Weapon for AI Inference in 2026

Jan 8, 2026
Nvidia has unveiled its latest innovations at CES, spotlighting the Vera Rubin architecture and the Bluefield-4 DPU. The shift from single chips to a full-stack compute approach is paving the way for enhanced AI inference performance. With Bluefield-4 designed as an inference storage processor, it tackles the memory wall challenge by offloading KV cache to NAND. Nvidia's Dynamo architecture promises a striking fivefold increase in inference speed. Updated profit growth expectations and potential valuation risks for Nvidia stock are also discussed.
Ask episode
AI Snips
Chapters
Transcript
Episode notes
INSIGHT

Full-Stack Trumps Single-Chip Wins

  • NVIDIA is focusing on full-stack compute rather than individual chip milestones.
  • System design, software, and infrastructure now drive competitive advantage in AI.
ANECDOTE

ChipStock Investor Built For Full-Stack Research

  • Nicholas describes building the ChipStock Investor web app to explain full-stack compute and the supply chain.
  • The app links research, portfolio actions, and company-level supply-chain details for investors.
INSIGHT

Bluefield-4 Extends Inference Memory

  • Bluefield-4 is positioned as a storage processor to expand inference context memory capacity.
  • Using NAND-based KV cache lets GPUs access larger inference state without relying solely on DRAM.
Get the Snipd Podcast app to discover more snips from this episode
Get the app