NVIDIA Adds Taiwan Memory Maker to Power Next‑Gen AI Chips

NVIDIA is boosting its Agentic AI platform, Vera Rubin, by adding a new memory supplier from Taiwan. The company now uses LPDDR5X chips from Nanya Technology, a first for the firm’s high‑performance AI servers. This move gives NVIDIA more flexibility. Vera Rubin relies on two kinds of memory: the CPUs use LPDDR5X for its low power and high density, while the GPUs need HBM4 for very fast data transfer. Only a few companies can produce HBM4, so NVIDIA has turned to more partners for LPDDR5X. Previously, Micron and SK Hynix supplied these chips, but the new partnership with Nanya expands the supply chain.

Nanya’s chips meet strict AI requirements that many local Taiwanese memory makers could not reach before. TSMC helped local firms improve manufacturing, enabling Nanya to qualify for the demanding Vera platform. The new LPDDR5X memory triples Vera’s capacity and adds 50 % more bandwidth compared to older servers. Each superchip can now hold 1. 5 TB of memory and run at 1. 2 TB/s, while a full rack can reach 400 TB and 315 TB/s. With Agentic AI shifting focus from GPUs to CPUs, having multiple global suppliers is essential. NVIDIA’s choice of a Taiwanese maker shows the company’s strategy to secure reliable, high‑performance memory from diverse sources.

Actions