Nvidia (NVDA.O) on Wednesday released data showing its latest artificial intelligence server can boost the performance of new AI models – including two popular Chinese models – by up to 10 times.
The report comes as the AI industry shifts focus from training models, where Nvidia dominates, to deploying them for millions of users, a space with increasing competition from rivals such as Advanced Micro Devices (AMD.O) and Cerebras.
Nvidia’s data centers on mixture-of-expert AI models, a method that improves efficiency by assigning different parts of a task to specialized “experts” within the model. The technique gained popularity this year after China’s DeepSeek released a high-performing open-source model that required less training on Nvidia chips than competitors in early 2025. Since then, the approach has been adopted by OpenAI, France’s Mistral, and China’s Moonshoot AI, which in July released a highly ranked open-source model.
Nvidia emphasized that, while such models may require less training, its servers are still essential for serving them to users. The company said its latest AI server, which packs 72 of its top chips into a single system with high-speed interconnects, improved the performance of Moonshoot’s Kimi K2 Thinking model by 10 times compared to its previous-generation servers – a gain similar to what it achieved with DeepSeek’s models.
The performance improvement stems mainly from the large number of chips and the fast interconnect links, areas where Nvidia retains an advantage over competitors. AMD is also developing a multi-chip server aimed at the AI market, expected to launch next year.
