NVIDIA’s Open Software Trap: The Real Cost of the New Inference Stack
Episode description
SUMMARY: We dig into the NVIDIA GTC keynote and highlight three things - accelerated computing for everything, the complexity of the new inference stack, and NVIDIA’s “open” software stack including NemoClaw.
SHOW: 1012
SHOW TRANSCRIPT: The Reasoning Show #1012 Transcript
SHOW VIDEO: https://youtu.be/aXOr91q76yM
SHOW SPONSORS:
- VENTION - Ready for expert developers who actually deliver?
Visit ventionteams.com
SHOW NOTES:
- NVIDIA GTC 2026 (Keynote)
- NVIDIA NemoClaw - OpenClaw + OpenShell + NVIDIA Agent Toolkit
- NVIDIA adds Groq LPU to their rack systems
- NVIDIA to invest $26B in Open Weight Models
- Interview with Jensen about Accelerated Computing (Stratechery)
Topic 1 - Jensen’s trying to paint the bigger picture of accelerated computing everywhere (robotics, autonomous driving, gen-ai, physical ai - but also just everyday enterprise apps). Everything is about keeping the stock price up, and margins high. The stock price provides the warchest to fight off all foes.
Topic 2 - The inference architecture is a complex mix of GPUs, CPUs, ASICs/LPUs, high-speed networking and seems very different from the training architecture. How big is the burden on data center providers? What are the inference alternatives emerging?
Topic 3 - Jensen talked a lot about OpenClaw and eventually about NVIDIA’s NemoClaw. How does his interest in Agentic AI tie into his interest in building NVIDIA’s own frontier model
FEEDBACK?
- Email: show @ reasoning dot show
- Bluesky: @reasoningshow.bsky.social
- Twitter/X: @ReasoningShow
- Instagram: @reasoningshow
- TikTok: @reasoningshow