NVIDIA showed a problem many AI app developers might face: balancing throughput with quality. More detailed responses and more advanced AI models require more tokens, but those tokens also have to be generated quickly, because otherwise users might switch to another app.

The company also shared a roadmap of upcoming datacenter CPUs and GPUs, which starts in the second half of 2025 with Blackwell Ultra NVL72.

Blackwell Ultra packs 2 GPUs in each chip. The GPUs have new instructions for attention, which is used in many AI applications, as well as 1.5 times more memory for KV cache (which stores the model and its context).

NVIDIA is also developing a switch called SpectrumX to manage the massive amounts of data transfers in a datacenter. Spectrum-X is built on Ethernet, with congestion control, very low latency and special software. The world's largest AI supercomputer, xAI's Colossus, uses Spectrum-X technology.

The Project DIGITS miniature AI supercomputer is now DGX Spark. It's powered by the NVIDIA GB10 Grace Blackwell Superchip, with a 20-core Grace CPU, a 1 petaflop Blackwell GPU and 128 GB of RAM.

And to go along with the new GPUs, new AI models were also announced.

NVIDIA Llama Nemotron is a family of open reasoning models built on top of Meta's Llama language models.

NVIDIA also released Isaac GR00T N1, an open foundation model for humanoid robots which can be trained to perform a variety of tasks.

To finish off GTC, NVIDIA announced Newton, an open-source physics engine developed with Google DeepMind and Disney Research.