Kimi K3 on AMD MI355X: cheaper than NVIDIA?

What's this about?

Inference provider wafer.ai published benchmarks claiming the open Kimi K3 model runs at better performance-per-dollar on AMD's MI355X than on NVIDIA's B300 — a notable claim, since NVIDIA's software moat is the usual reason nobody bothers with AMD for serving LLMs.

What exactly is "perf per dollar" in inference?

Serving an LLM is a throughput business: tokens generated per second, divided by what the GPU-hour costs you. A chip can lose on raw speed but win on economics if it's cheap enough — which is exactly wafer.ai's claim (and exactly what critics dispute, since the rental prices they assumed drive the whole result).

What's the community saying?

1. Top comment is terminology policing: it's an "open-weight" model, not "open source" — no training data or code was published.
2. Number-crunchers found B300 won every raw-throughput row; only the assumed $2.50 vs $6.00 GPU-hour pricing flips the result, and those prices were called unrealistic.
3. Several accuse wafer.ai of a hype pattern — "making themselves synonymous with slop in the inference space."
4. A capex subthread: at these rental prices the MI355X can't pay for itself in its lifetime, so someone is subsidizing.