Nvidia claims 10x boost as new AI server speeds up next-gen global models

Nvidia says its new AI server delivers 10x faster performance for advanced models, even as rivals like AMD and AWS push competing tech to win the growing market for AI deployment

Don't want to miss the best from Business Standard?

Nvidia on Wednesday released new performance data showing that its latest artificial intelligence (AI) server can deliver up to 10 times faster results when running next-generation models, including two widely used Chinese models, news agency Reuters reported.

 

The update comes at a time when the global AI industry is shifting focus. While Nvidia continues to dominate the market for training AI systems, competition is much tougher in the area of deploying those models for millions of users. Rivals such as Advanced Micro Devices (AMD) and Cerebras have been trying to close the gap in this growing segment.

 

Rise of mixture-of-experts models

 

Nvidia’s new results centre on mixture-of-experts (MoE) models, an increasingly popular AI technique. MoE systems break down user questions into smaller parts and route them to different “experts” within the model, improving efficiency.

 

The approach became widely known after China’s DeepSeek released a high-performing open-source model in early 2025, which required less training on Nvidia chips than many competitors.

 

Since then, several major players, including OpenAI, France-based Mistral and China’s Moonshot AI, have adopted the MoE style. Moonshot even launched its well-rated Kimi K2 Thinking model in July, further boosting interest in this technique. 

 

Nvidia says its servers still matter

 

As MoE gains traction, Nvidia has been working to show that its hardware continues to be essential, not just for training, but also for serving these models to users.

 

According to the company, its newest AI server contains 72 top-end Nvidia chips connected through extremely fast data links. This setup, Nvidia said, improved the performance of Moonshot’s Kimi K2 Thinking model by 10 times compared with the previous generation of Nvidia servers. The company reported similar gains when running DeepSeek’s models.

 

Nvidia attributed these improvements to two key strengths:

• The large number of chips it can assemble into a single system

• The high-speed connections between those chips

 

Both areas remain strong advantages for Nvidia over its competitors, the company said.

 

While Nvidia pushes ahead, competitors are also advancing. AMD is developing a multi-chip AI server, similar in concept to Nvidia's, which it plans to launch next year.

 

Amazon backs Nvidia tech for its future AI chips

 

Amazon’s cloud division AWS announced on Tuesday that it will use Nvidia’s “NVLink Fusion” technology in an upcoming AI chip called Trainium4, Reuters reported.

 

NVLink is one of Nvidia’s most valuable technologies because it creates ultra-fast connections between different types of chips, enabling efficient processing for large AI workloads.

 

The announcement was made during AWS’s annual cloud conference in Las Vegas.

 

Nvidia has been working to get more chipmakers to adopt NVLink. With Intel, Qualcomm, and now AWS joining in, the technology is expanding across the industry.

 

AWS said NVLink Fusion will help it build much larger AI systems that can communicate faster and operate in sync, an essential requirement for training massive AI models that rely on thousands of connected machines.

 

As part of the partnership, AWS customers will also gain access to what the company calls “AI Factories",  dedicated AI infrastructure inside their data centres designed to deliver higher speed, security, and readiness for large-scale AI projects.

 

(With agency inputs)