NVIDIA's research lands in other labs' models, not its own Nemotron3 Ultra

SemiAnalysis, an industry analysis outlet, argues that NVIDIA's research division has produced techniques used inside other labs' models, citing LatentMoE inside Moonshot AI's Kimi K3 and GatedDeltaNets inside Alibaba's Qwen models, but that NVIDIA's own frontier training effort, which it attributes to the company's internal culture, has not matched that research output. It says NVIDIA's Nemotron3 Ultra, at 550 billion total parameters with 55 billion active, trails several Chinese models including Qwen3.8 27B, which has roughly 20 times fewer parameters.

SemiAnalysis does not name which benchmark produced that specific comparison in the post text, and the supporting chart is an image this review could not independently verify, so the ranking claim is presented here as the outlet's own assessment rather than a cited, reproducible score.