Gadgets · by Thandi Mokoena · 2026-10-10

Nvidia's data center dominance vs AMD's MI-series — is the GPU gap actually narrowing for AI workloads?

AMD's Instinct MI-series keeps getting better software support (ROCm maturity, PyTorch upstream support improving), and some hyperscalers are reportedly diversifying away from pure Nvidia dependence for cost and supply reasons. Nvidia still wins decisively on CUDA ecosystem lock-in and raw interconnect (NVLink) for massive training clusters, but for inference workloads specifically the cost-per-token story on AMD hardware is apparently getting competitive. Anyone running production inference on MI300X/MI325X who can share real numbers vs H100/H200 clusters?

0 sparks · 4 comments

Comments

Luciana Vargas · 0 likes

We moved a chunk of our inference fleet to MI300X this year mainly for availability reasons (lead times on H100s were brutal). ROCm still has rough edges for some custom kernels but vLLM support is solid now.

Maya Putri · 0 likes

Good to hear vLLM support matured. That was the main blocker for us testing AMD last year.

Ethan Brooks · 0 likes

For training at scale, nobody's touching Nvidia's NVLink/NVSwitch topology yet. The software moat compounds the hardware moat there.

Avery Johnson · 0 likes

The real story is cost per token at inference, not raw FLOPs. If AMD can undercut on \$/token for serving, enterprises will route around CUDA lock-in pretty fast regardless of brand loyalty.