With the RAM shortage driving up prices, it’s no surprise that even some of the tech industry’s biggest players are turning to storage to save money. While there are some encouraging signs, like increased RAM production and Google’s moves to help lower prices, the near-term chip shortage is having a significant impact on many companies’ bottom lines. Meta, in response, took the dramatic step of recycling old RAM into some of its AI servers.
The company revealed at ISCA 2026 that it was salvaging DDR4 from decommissioned servers and integrating it directly into what were once DDR5-only servers, according to a report from TechSpot. It uses technical techniques to ensure that mixed memory can work well together, which has always been a major problem when trying to mix memory from multiple generations. Meta also works in a way that does not cause a noticeable decrease in performance.
A bridge between generations
Hoyka Studio/Shutterstock
Meta’s new servers, which it calls “MemServers,” are made of about 75% DDR5 and 25% recycled DDR4. Meta’s ability to prioritize frequently called data and keep it stored in DDR5 is critical to ensuring that older, slower memory does not cause an undue loss in performance. DDR4 is reserved for less used data.
Old and new RAM modules are able to communicate and coordinate with each other using a technology called Compute Express Link (CXL). CXL is built on the same physical infrastructure as the existing PCIe architecture, allowing Meta to plug custom CXL memory expansion cards into PCIe Gen5 slots without requiring dedicated DDR4 memory slots on the motherboard. The cards carry Meta’s Vistara ASIC, which presents itself as a CXL Type-3 memory device on this PCIe link. According to Meta (via TechSpot): From an OS perspective, the DDR4 connected via Vistara appears as a “CPU-less NUMA node,” meaning it is separate from the local DDR5 DRAM that sits on the CPU’s memory channels.
