Good morning. It’s the weekend. Pour yourself a cup of coffee. This is a long hypothesis.
I just got back from a week at GTC and OFC back to back, and I need you to use a little bit of your imagination with me. Not for science fiction — for what’s actually being built right now, and what the world is going to look like in the next few years when it ships. We are going to dive deeper than just the GTC Update.
On Monday, Jensen Huang stood on the SAP Center stage and showed the world seven new chips, five rack-scale systems, and a $1 trillion revenue pipeline. I covered the hardware in detail in The Fourth Piece Ships.
But the hardware isn’t the story. The hardware is the answer to a question most investors haven’t fully internalized yet.
The question: what happens when the world needs orders of magnitude more tokens than it produces today?
Not incrementally more. Orders of magnitude more. Continuously. From millions of agents running simultaneously, reasoning models thinking for minutes before responding, robots navigating physical environments in real time, and million-token context windows accumulating state across hours-long sessions. Jensen’s own slide showed inference compute scaling 10,000x from ChatGPT to Claude Code in just two years — and agentic workloads multiply that further.
That world isn’t theoretical. It arrived this year. And GTC 2026 was NVIDIA’s blueprint for the infrastructure required to serve it.
I spent the week on the ground at GTC interviewing the people building this infrastructure — Link to interviews in hyperlinks, including Dion Harris, Senior Director of HPC and AI Hyperscale Infrastructure, Stuart Pitts, Senior Manager of Product Marketing for Accelerated Computing (who joined NVIDIA from Groq), and Adel El Hallak, VP of AI and Product Management. Their perspectives, combined with the keynote data and the briefing materials, paint a picture that’s more radical than even the headlines suggest.
Leave a Reply