〈CES 2026〉NVIDIA's Rubin era is upon us!

The platform integrates six brand-new chips and has been fully put into production. It is expected to be officially launched onto the market in the second half of 2026. What is particularly noteworthy is that the cost per token of the Rubin platform is 10 times lower than that of the previous generation Blackwell platform, accurately fulfilling Huang Renxun's previous technical promise of "the more you buy, the more you save". It is expected to reshape the cost and performance landscape of the AI computing power market.

The Rubin platform released this time features "six-chip collaboration" as its core architectural highlight, with six brand-new chips each performing its own function while deeply interacting with each other, forming a computing power ecosystem that covers the entire process of computing, networking, and data processing. Specifically, the platform includes two core computing units: NVIDIA Vera CPU and NVIDIA Rubin GPU, paired with four auxiliary chips: NVIDIA NVLink 6 switch, NVIDIA ConnectX-9 SuperNIC, NVIDIA BlueField-4 DPU, and NVIDIA Spectrum-6 Ethernet switch. Through extreme collaborative design, the six chips can achieve efficient interaction, which not only significantly shortens the training time of AI models but also significantly reduces the cost per word in the inference stage, providing comprehensive computing power support for ultra-large-scale AI tasks.

Renxun Huang emphasized at the press conference that the launch of the Rubin platform is timely. Currently, the global demand for AI computing is experiencing explosive growth, and cost control and efficiency improvement have become core demands in the industry. The 10-fold reduction in word cost achieved by the Rubin platform will effectively break through the cost bottleneck of large-scale AI applications, helping enterprises significantly reduce computing power investment, whether it is for large model training or inference deployment. This core advantage perfectly aligns with NVIDIA's previous value proposition of "the more you buy, the more you save", further consolidating its leading position in the high-end AI chip field.

From a technical architecture perspective, the six chips on the Rubin platform form a complete computing power loop. Among them, the Vera CPU and Rubin GPU constitute the core computing engine, responsible for undertaking the main computational tasks; the NVLink 6 switch ensures high-speed data transmission between chips, while the ConnectX-9 SuperNIC and Spectrum-6 Ethernet switches build an efficient network interconnection system. The BlueField-4 DPU focuses on data processing and security protection. This full-stack collaborative design approach enables the platform to systematically address the bottlenecks of computing power, network, and storage in AI tasks, providing strong support for cutting-edge application scenarios such as delegated AI and physical AI.