NVIDIA CEO Jensen Huang stated during the company's GPU Technology Conference in Berlin, Germany, on Wednesday local time that a memory supply glut is unlikely by 2029, adding that "one of the real benefits of memory is that you can always use more of it."
Recently, an expert writing in South Korean media suggested that the memory super-cycle may have ended and that the industry could enter a downcycle by 2029. As a result, global memory-related stocks came under selling pressure.
Memory is a core component of NVIDIA's chips, and an oversupply could depress prices and squeeze profit margins at manufacturers such as SK Hynix, Samsung, and Micron.
Huang said NVIDIA (NASDAQ: NVDA) will continue to use large quantities of memory chips including LPDDR and HBM. When asked whether the company would consider partnering with China's ChangXin Memory Technologies, Huang responded that he is open to the idea. "We work with Chinese companies — almost the entire world works with Chinese companies at some point in the supply chain. China is excellent at underlying infrastructure technology, and no country is stronger than China in manufacturing," he said.
On October 21, NVIDIA kicked off its 2026 GTC conference in Berlin, with Huang's keynote leading the event. Over the course of an hour and a half, Huang covered topics ranging from NVIDIA's AI infrastructure strategy to intelligent agents, and from open models to humanoid robots and autonomous driving. Huang unveiled the new-generation Vera Rubin system, which he said is 100% liquid-cooled and can connect 72 GPUs in a single rack to form a supercomputer. He noted that the previous-generation Blackwell architecture processes large language models 30 times faster than Hopper, and Vera Rubin is expected to deliver another 30-fold improvement within two years, surpassing Moore's Law expectations.
Building on the new architecture, Huang also announced the DSX POD AI factory, describing it as a complete Vera Rubin AI factory reference design that can connect directly to power, cooling, networking, and other supporting infrastructure at a prepared site, reducing deployment time from six months to 14 days. Seventeen technology companies, including Lenovo Group, Dell, and HPE, are ecosystem partners for the solution.
Notably, at the same time as Huang's keynote, Google Cloud and NVIDIA announced that A5X instances powered by NVIDIA's Vera Rubin NVL72 system will officially open next month. Google Cloud also plans to deploy up to 960,000 Rubin GPUs across multiple sites. Ineffable Intelligence, a London-based AI company, has already become an early user of the A5X platform for large-scale reinforcement learning research.
Comments