NVIDIA GTC 2026: Mystery Chip, Vera Rubin, Feynman — Everything You Need to Know

[user nl]

Interesting preview / overview of things to come today and this week at GTC 2026:

The keynote by Jensen Huang can be viewed at many links, e.g. here is youtube's link:

 

[Xebec]

This is still going on. TIL IBM invented SQL in the (early?) 1970s. Nvidia announced DLSS5, which appears to mostly improve human faces in games.

Later on, I also saw Jensen had a Semianalysis slide up, and even mentioned Dylan Patel by name.

 

[Xebec]

Nvidia's Vera Rubin appears to be focused on improving higher end / "higher value" AI inference workloads. Where Blackwell improved "35X over Hopper" for "Free and Medium tiers", Vera Rubin only improved 2-3X at the lower tiers (still good!), but brings a "35X improvement" at the high end.

A later slide showed the effect of adding Groq-3 chips (heavy on SRAM, optimized more for latency than bandwidth) to Vera Rubin arrays, and that pushed the speed further 'to the right', enabling even higher end tiers (more guarenteed tokens/second for customers).

 

[user nl]

I watched it and an amazing show! He was really pushing OpenClaw, comparing it to Linux. With the extra security and OpenShell improved to NemoClaw.

The song/recap at the end was also really nicely done. See from here:

 

[user nl]

This is still going on. TIL IBM invented SQL in the (early?) 1970s. Nvidia announced DLSS5, which appears to mostly improve human faces in games.

Later on, I also saw Jensen had a Semianalysis slide up, and even mentioned Dylan Patel by name.

View attachment 4337

Photo also used in the NYT report:
https://www.nytimes.com/2026/03/16/technology/nvidia-gtc-ai-chips-huang.html

Mr. Huang unveiled a product incorporating technology from a start-up called Groq. The product will pair Nvidia’s chips, which excel at receiving an A.I. request, with Groq’s chips, which have components that can put a charge into how Nvidia’s chips operate.

Over the past year, A.I. companies have shifted their work. The A.I. systems they built using Nvidia’s chips have improved at creating software code, doing research and making images and videos. These capabilities, the result of a process known as inference, have put more value on chips that can generate data as inexpensively and quickly as possible.
......................
Nvidia’s deal with Groq also helps it with manufacturing problems that are constraining how fast its sales can grow, said Umesh Padval, a managing partner at the investment firm Seligman Ventures.
Groq’s chips are made by Samsung Electronics, not Taiwan Semiconductor Manufacturing Company, which makes most of Nvidia’s chips and is struggling to meet the company’s demand, Mr. Padval said. And unlike Nvidia’s chips, Groq’s don’t require high-bandwidth memory chips. Those chip manufacturers have also been swamped with orders.

“It’s a brilliant supply-chain move,” Mr. Padval said.