Nvidia announces new AI server with 10x higher performance.

Nvidia announces new AI server with 10x performance, integrating 72 chips and MoE architecture, leading the global artificial intelligence field.

On December 3rd, Nvidia released the latest data showing that its new artificial intelligence (AI) servers are capable of improving the performance of advanced AI models, including popular ones from China, by up to 10 times compared to the previous generation of servers.

This data comes as the AI ​​industry shifts its focus from model training, an area where Nvidia currently leads, to deploying models to millions of users, an area seeing increased competition from rivals such as Advanced Micro Devices (AMD) and Cerebras.

Nvidia says these improvements stem primarily from its ability to integrate a large number of chips into a single server and the high-speed connections between them, an area where the company remains a leader and maintains a clear advantage over competitors. Nvidia's latest AI server is equipped with 72 of the company's top-of-the-line chips.

The data released by Nvidia primarily focuses on AI models using the Mixture-of-Experts (MoE) architecture, a method that optimizes AI model performance by dividing tasks into separate parts and assigning them to different "experts" within the processing model.

MoE architecture gained popularity in 2025, especially after DeepSeek, a Chinese AI company, introduced a high-performance open-source model that required less training time on Nvidia chips compared to competitors.

Since then, major companies such as OpenAI (the maker of ChatGPT), Mistral of France, and Moonshot AI of China have begun applying MoE methods to their models. Moonshot AI released a highly-rated open-source model last July that uses this technique.

While Nvidia continues to maintain an advantage in AI model deployment, competitors like AMD are also working hard to develop competing products.

AMD is expected to launch a similar AI server next year, integrating several powerful chips with the goal of directly competing with Nvidia's servers in the inference (model processing and deployment) field.

(VNA/Vietnam+)

Source: https://www.vietnamplus.vn/nvidia-cong-bo-may-chu-ai-moi-co-hieu-suat-cao-gap-10-lan-post1080980.vnp