NVIDIA's apparent decision to shift its upcoming Vera Rubin AI systems to smaller memory modules seems counter-intuitive, but it's actually a clever move that increases overall demand.
Recent reports suggest that NVIDIA is changing the memory mix for its NVL72 "Vera Rubin" systems, moving from high-capacity 192GB SOCAMM2 modules to smaller 96GB units. While this looks like a downgrade, it's causing total demand for LPDDR5X memory to jump by an estimated 10-20%. This surprising outcome highlights the complex dynamics of the AI hardware supply chain.
So, how does using smaller modules lead to more demand? It comes down to simple math. First, each Vera Rubin system rack is designed to have a massive 54 terabytes (TB) of memory. To reach this target, you would need 288 of the 192GB modules per rack. However, if you use the 96GB modules, you need twice as many—a total of 576 modules—to achieve the same total memory capacity. This doubling of the module count is the key to understanding the increase in demand.
This strategic shift is driven by the current global memory market. Secondly, there's a significant shortage of high-end DRAM like LPDDR5X. Memory manufacturers are reallocating their production lines to make HBM (High Bandwidth Memory), which is crucial for AI accelerators, creating a supply squeeze for other types of memory. Thirdly, by using smaller, more readily available 96GB modules, NVIDIA can ramp up production of its Vera Rubin systems more smoothly and quickly. It's a pragmatic choice to get more systems out the door now.
The ripple effect is a significant tailwind for the supply chain. Not only do the DRAM makers sell more total memory, but suppliers of other components—like the substrates the chips sit on, specialized materials, and testing services (OSAT)—see a huge jump in unit orders because twice as many physical modules are being built. This move also leaves an upgrade path open for customers to install higher-capacity modules later when supply improves.
- SOCAMM2: A new, compact memory module standard designed to deliver high-speed LPDDR5X memory performance for systems like laptops and AI servers.
- LPDDR5X: A type of high-speed, low-power DRAM (Dynamic Random-Access Memory) commonly used in mobile devices and increasingly in AI hardware for its efficiency.
- HBM (High Bandwidth Memory): A high-performance RAM interface for 3D-stacked memory, used in high-performance graphics accelerators and network devices.