Member-only story
A Chinese food-delivery company just trained a 1.6-trillion-parameter language model on roughly 50,000 domestic AI chips, start to finish, without a single NVIDIA GPU touching the pre-training run. That is the sentence everyone quoted when Meituan released LongCat-2.0 on 30 June 2026. It is true, and it is the least interesting thing about the model.
Here is the interesting thing. LongCat-2.0 has 1.6 trillion parameters and uses about 48 billion of them per token. Three per cent. The other 97% sit idle for any given word. That number is not a footnote to the “no NVIDIA” story. It is the “no NVIDIA” story. The whole reason a second-tier accelerator cluster could carry a frontier-scale model is that Meituan built a model whose per-step workload, and more importantly, whose per-step network traffic, is small enough for silicon that loses to Hopper on almost every axis that normally matters.
This piece is about that trade. What breaks when you take NVIDIA out of the training loop, why the usual answer is “the interconnect, not the maths,” and how a Mixture-of-Experts design with two specific tricks (zero-computation experts and a cross-layer shortcut) turns…