Nvidia Shows Big Speed Boost With New AI Server Technology

Nvidia has shared new performance results on Wednesday showing that its latest artificial intelligence (AI) server can deliver up to ten times faster speed while running next-generation models. This includes better performance with two popular Chinese models, according to a report by Reuters. The update comes at a time when the global AI market is rapidly evolving. Nvidia still leads in building powerful hardware for training AI systems, but competition is rising when it comes to deploying and serving these models for millions of users across the world. Companies like AMD and Cerebras are working hard to reduce Nvidia's lead in this fast-growing segment.

The latest results shared by Nvidia focus on Mixture-of-Experts (MoE) models, a modern AI technique that is gaining strong attention. In MoE systems, a user’s question is divided into smaller tasks and sent to different specialist “experts” inside the model. This process helps the model work faster and more efficiently. The approach became widely known after China’s DeepSeek released a powerful open-source model in early 2025. The model impressed many in the AI industry because it required less training on Nvidia chips than most other AI systems.

Following DeepSeek's success, several global AI leaders such as OpenAI, France’s Mistral, and China’s Moonshot AI started adopting the MoE method. In July, Moonshot AI launched its well-reviewed Kimi K2 Thinking model, which further increased interest in MoE-based technologies.

As more companies shift toward MoE models, Nvidia aims to show that its hardware remains essential not only for training large models but also for running them efficiently at scale. The company said its newest AI server contains 72 high-performance Nvidia chips connected with extremely fast data links. According to Nvidia, this setup increased the performance of Moonshot’s Kimi K2 Thinking model by nearly ten times compared to older Nvidia servers. The same level of improvement was also reported for DeepSeek’s models.

Nvidia says these high speed gains come from two major advantages:

1. The ability to combine a large number of chips into one powerful system

2. The extremely fast communication links between those chips

These strengths, Nvidia claims, still give it an edge over its rivals in the growing AI hardware market.

Meanwhile, competitors like AMD are also moving forward. AMD is working on a new multi-chip AI server that uses a design similar to Nvidia’s. AMD plans to launch this new product next year, adding more competition in the field of AI infrastructure.

In another major development, Amazon Web Services (AWS) announced it will use Nvidia’s NVLink Fusion technology in its upcoming AI chip called Trainium4. NVLink is one of Nvidia’s most important inventions because it creates extremely fast connections between processors, allowing huge AI workloads to run smoothly and quickly.

According to AWS, using NVLink Fusion will help the company build much larger and faster AI systems that can communicate effectively across thousands of connected machines. This is especially important for training massive models that require continuous, high-speed data movement.

AWS also revealed that customers will soon have access to special “AI Factories” inside its data centres. These will provide high-speed and secure AI infrastructure designed for large-scale AI projects. With more partners such as Intel, Qualcomm, and AWS adopting NVLink, Nvidia’s influence across the AI industry continues to expand.