I watched Nvidia’sGTC live event in March. Like at other large technology company events,investors, developers, economists, and other ecosystem stakeholders likeLeading Edge Design Group (LEDG), look to this event forproduct announcements and projections for the coming year. I was expecting some exciting announcements, but I was floored by what I heard.

If you want to watch the entire two-hour keynote, you can do here. Not that there is not a full two-hours of industry content and has some Nvidia bragging elements.

I will get to what shocked me in this event, but first I'll share a few other takeaways from Nvidia CEO Jensen Huang's address.

Key Takeaways

• Huang talked about how AI fundamentally changed how computing is done. Instead of "retrieving" information, AI has given computing the ability to "generate" answers. Hence, the term "generative AI". Huang further defined a breakthrough in the past couple years as what Nvidia calls "Agentic AI". This means that AI has "agency", it can understand circumstances and reason ho to solve a problem.
  
  A large focus of his address was on the "next wave" of AI, which Nvidia calls "Physical AI". This is what is enabling things like robotics and autonomous vehicles. Here is an image for reference.

‍

• Huang's overview of scaling was worth noting. He shared that the amount of computational power that we need for Agentic AI, to support "reasoning", was easily 100X more than they thought they would need at this time last year.
  
  A practical example of this he cited was ChatGPT. Two years ago, Huang described ChatGPT's responses as "one shot", meaning the system would look through its trained data and provide response. It got a lot wrong. "
  
  Now, the AI applications can check their own work. Once an answer is generated, it cycles it back through multiple verfication processes. All of this increases the quality of the output but also drastically increases the amount of compute required if you want the model to be responsive.
  
  Later in the address, there is a short video example of this comparing traditional LLMs with reasoning models and the amount of compute ("tokens") required. (This video is around the 1 hour - 28 minute mark)
  ‍
• Huang has said before that he believes he will reach a trillion dollars in annual CAPEX spend on data center infrastructure. He shared his slide from analysts predicting what the remainder of the decade is forecasted to bring for data center CAPEX spend.|

• Jensen repeatedly refers to data centers as "AI factories". He uses this to demonstrate how "factories" can produce more AI tokens per second in less space as Nvidia's technology evolves. Here is an example of the number of racks in a 100mW facility with the Nvidia Grace Hopper system vs. the new Blackwell system:

‍
Toward the end, Huang unveiled the "Rubin Ultra" next-generation system, which Nvidia expects to release in the second half of 2027. He talked about the components of this system (image below) and then shared what infrastructure is required to support it. The Rubin Ultra will be 600kw per rack (note a typo) 🤯🤯🤯

About this enormous transition, Huang said "This transition is going to take years of planning, this isn't like buying a laptop, which is why I am telling you now. We must plan with the land and the power for data centers with engineering teams two or three years out, which is why I (am showing) the roadmap.

‍