What NVIDIA chose as its benchmark, and why it matters
Two weeks ago a self-described twenty-five-year market veteran asked on LinkedIn whether investing in AI was a good bet.
My answer: it depends. Based on American models alone, no. Including Chinese models in the analysis, yes.
His reply: “Only a fool would think Chinese models are better than American ones.”
Then look at the slide
At the launch of Nemotron 3 Ultra — presented by NVIDIA itself, the world leader in AI chips — which models were chosen as the benchmark of record?
The Chinese ones. Kimi. Alibaba.
Not out of ideology. Out of respect for efficiency.
The data is not ambiguous. Chinese open-weight models deliver equivalent quality at a small fraction of the cost, and in coding and mathematics benchmarks they lead. NVIDIA acknowledges this publicly, on stage, in the chart it chose to show.
Ignoring that is not patriotism. It is poor decision-making.
Why this is the same argument as everything else here
A benchmark is a statement about what you consider the frontier. Choosing one is an architecture decision, and it carries the same property as every other architecture decision I write about: it looks like a technical detail and it is actually a commitment about cost.
The same logic sits behind GOE. It is not about where something comes from — it is about what actually works and creates real value. The human workforce is the most undervalued asset in the market, and real efficiency starts by recognising that.
I run fixed-price architecture reviews and infrastructure cost audits for founders. See the engagements, or read the background.