NVIDIA’s Open Software Trap: The Real Cost of the New Inference Stack Artwork

The Enterprise AI Show

The Enterprise AI Show explores the AI journey for Enterprise companies around the world.

As the AI revolution moves from experimentation to execution, The Enterprise AI Show provides the clarity needed to lead. Join Aaron Delp and Brian Gracely as they explore the intersection of generative AI, enterprise systems, and global business strategy. Each episode features clear-headed conversations with the people making actual decisions—founders, investors, and practitioners—focusing on the technical architectures and business models that drive real-world ROI.

New shows every Wednesday and Sunday.

Topics: Enterprise AI strategy · The AI Economy · LLMs in production · AI leadership · Agentic AI · Digital Sovereignty · Machine Learning · AI startups · Cloud Computing

All Episodes

The Enterprise AI Show

NVIDIA’s Open Software Trap: The Real Cost of the New Inference Stack

March 22, 2026 • Massive Studios

0:00 | 28:00

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?