Dylan PatelFounder and Chief Analyst, SemiAnalysis · The semiconductor-AI stack

Beyond CUDA: NVIDIA’s push into networking

27 Aug 2025 · 1 min

Watch on YouTube

0:39

NVIDIA is pushing beyond its CUDA software ecosystem into networking hardware. The key breakthrough is shifting to copper interconnects in the H100, which cuts the energy costs of moving data during AI training and inference, lowering total cost of ownership for customers.

Chapters

What are your thoughts on NVIDIA expanding their networking and CUDA to even box out the hyperscalers more?
I feel like CUDA is not even incremental. Everyone who's listening to this should understand, hey, CUDA is this really great ecosystem that really makes a big difference. The networking part, I think, is the real incremental thing that's really changed in the last, let's say, year or two. The Copper domain scale-up is the big difference between the GP200 and the H100. And like, what's the big deal? Well, moving bits is such a big part of the energy and power of inferencing and training these models. If you have it all in copper, which is the energy-efficient way to do this, you're gonna have a much better TCO. And that's like, you know, the big Nvidia networking thing.