Key takeaways
The company published data on Wednesday demonstrating that its latest server configuration, which integrates 72 high-performance chips into a single computer with accelerated inter-chip connections, substantially improved the performance of Moonshot's Kimi K2 Thinking model compared to previous-generation Nvidia servers.
The company reported similar performance gains with DeepSeek's AI models.
Shifting competitive landscape in AI hardware
The benchmark release comes as the artificial intelligence industry pivots from model training to deployment and inference, putting trained models to work for millions of users.
This shift represents a strategic challenge for Nvidia, which has dominated the AI training market but now faces stronger competition from rivals, including Advanced Micro Devices and Cerebras, in the inference space.
The performance improvements center on mixture-of-experts AI models, an efficiency technique that divides computational tasks among specialized subnetworks within the model.
This architectural approach reduces computing requirements by activating only relevant portions of the model when processing queries.
Mixture-of-experts architecture exploded in adoption following DeepSeek's surprise release of a high-performing open-source model in early 2025.
The Chinese startup demonstrated that competitive AI models could be built using significantly less training compute on Nvidia chips than Western alternatives.
The approach has since been adopted by major AI developers, including OpenAI, France's Mistral, and China's Moonshot AI, which released its highly-ranked open-source Kimi K2 model in July 2025.
Rising competition from AMD and other chipmakers
While Nvidia maintains dominance in AI model training, competitors are advancing rapidly in the inference market.
AMD is developing multi-chip server systems expected to launch in 2026, with the company projecting its data center business will grow at a compound annual rate exceeding 60% through 2030.
Oracle Cloud Infrastructure announced in October that it will deploy 50,000 AMD graphics processors starting in the second half of 2026.
Earlier that month, OpenAI entered a major deal with AMD for processors requiring 6 gigawatts of power over multiple years, with an initial 1-gigawatt deployment beginning in 2026.
Qualcomm has also entered the competitive arena, announcing in October its AI200 chip for 2026 release and AI250 for 2027, both designed for full rack-scale systems comparable to Nvidia's offerings.
Chinese AI models demonstrate cost-efficient innovation
Moonshot AI's Kimi K2 Thinking model, released in November 2025, cost approximately $4.6 million to train, according to sources familiar with the project.
The model features a 1-trillion-parameter mixture-of-experts architecture with 32 billion active parameters and can autonomously execute 200 to 300 sequential tool calls for complex tasks.
The Beijing-based startup, backed by Alibaba, built its model using the H800 chip, a downgraded version of Nvidia's H100 GPU that was marketed exclusively in China until October 2023.
Despite U.S. export restrictions limiting Chinese companies' access to advanced AI chips, Moonshot and DeepSeek have demonstrated that algorithmic efficiency and architectural innovations can partially compensate for hardware constraints.
DeepSeek's breakthrough with its V3 model, which the company claims cost approximately $5.6 million to train, contrasted sharply with the billions spent by OpenAI on comparable models.
The Chinese firm's success sent shockwaves through global markets in January 2025, temporarily reducing Nvidia's market value by $600 billion in what observers called a "Sputnik moment" for American AI development.
Industry analysts note that while export controls may slow Chinese AI development, they have simultaneously incentivized creative approaches to maximize computational efficiency.
The continued emergence of competitive Chinese AI models at significantly lower costs raises questions about the sustainability of current Western AI business models and the effectiveness of technology export restrictions as a strategic tool.
Read more:
Thai Police Apologize For AI-Manipulated Flood Relief Image
Anthropic Prepares For Potential 2026 IPO As AI Startup Competition Intensifies
Proteus Space Launches First AI-Designed Spacecraft, Sets Multiple Records