Nvidia is hoarding memory supply before the next AI crunch hits.
Taiwan’s Nanya Technology has been selected to supply LPDDR5X for Nvidia’s Vera Rubin platform, as the company broadens its supplier base for server main memory.
Vera Rubin uses two memory types, with Vera CPUs on LPDDR5X DRAM and Rubin GPUs on HBM4 DRAM.
HBM4 is the high-bandwidth stuff with a short guest list of makers like Samsung, SK Hynix and Micron, while LPDDR5X is about power efficiency and density.
Micron and SK Hynix have already developed SOCAMM2 memory for Vera Rubin platforms using LPDDR5X, but Nvidia has been seeking additional partners.
Nanya Technology has become the first Taiwanese manufacturer to enter Nvidia’s AI server main memory system, breaking the previous dominance of Korean and American companies and marking a new milestone for Taiwan’s memory industry.
It is a decent scalp for Nanya because many Taiwanese memory outfits tied to the PC market have struggled to meet the tighter specs demanded by AI platforms.
TSMC has been guiding local companies on manufacturing and process optimisation, and Nanya looks like the first to cash in on that help.
Vera Rubin servers are being sold as a clear step up from Grace Blackwell, with each Vera Rubin Superchip packing 1.5 TB of memory and delivering 1.2 TB/s.
That is pitched as three times the capacity and more than 50 per cent higher bandwidth versus the previous generation.
Nvidia is pushing rack-scale plans too, with 256 Vera chips per rack, up to 400 TB of memory and up to 315 TB/s of bandwidth.
With agentic AI shifting more work onto CPUs, and memory demand refusing to behave even with KV cache compression hitting 90 per cent in newer models, Nvidia is stacking suppliers so it does not get caught short.
