Nvidia moved to strengthen its lead in the AI Semiconductor market as it began full-scale deliveries of its next-generation artificial intelligence (AI) platform, "Vera Rubin." With systems already built for major cloud providers including Google, Microsoft (MS), and Oracle, the company's strategy is to counter rivals by emphasizing performance and power efficiency.
On the 21st (local time), Nvidia said it supplied the next-generation AI rack "Vera Rubin NVL72" to Google Cloud, Microsoft Azure, Oracle Cloud, and CoreWeave, and that it is currently running at major customers.
Vera Rubin NVL72 is an AI supercomputer that integrates seven new chips into a single system, including the next-generation GPU "Rubin" and the CPU "Vera." Nvidia also said it built a supply chain for mass production that consolidates more than 350 factories across 30 countries worldwide.
According to performance evaluation results released by Nvidia, Vera Rubin's performance per watt measured using the DeepSeek R1 model in a CoreWeave environment improved by up to 10 times compared with its predecessor, "Grace Blackwell." The company said that with liquid cooling technology at 45 degrees Celsius applied, annual water use at data centers can also be reduced by thousands of tons.
Network performance was also reinforced. By applying the "Spectrum-6" Ethernet switch with throughput of 102.4 terabits (Tb) per second—double the previous generation—the system was designed so that large-scale AI data centers can be operated like a single system.
Ian Buck, a Nvidia vice president, said at a briefing at the company's headquarters in California on the day, "Vera Rubin has entered full-scale mass production and is already running at major customers."
Nvidia is highlighting the performance of its next-generation platform because competition in the AI Semiconductor market is growing increasingly intense. AMD plans to supply its next-generation AI rack "Helios" to major customers including MS, while OpenAI and Meta are expanding development of in-house AI chips. Anthropic is also reportedly reviewing development of its own AI chips.
It also stepped up its push into the CPU market. Nvidia claimed that Vera CPU's Python execution performance is up to 1.8 times higher than AMD's next-generation server CPU "Turin." However, with AMD also set to launch its next-generation CPU "Venice," competition over the AI infrastructure market is expected to intensify further.
Meanwhile, although there were reports that the launch of the next rack architecture "Kyber NVL144," which will be equipped with the next-generation GPU "Rubin Ultra," was delayed to 2028 by more than a year, Nvidia has denied those reports.