Meta is running compute express link (CXL)-based memory expansion across millions of servers, reusing DDR4 modules from decommissioned machines rather than retiring them. The company disclosed this approach in a paper published last month and stated that this deployment can cut server count by 25% while lowering operational costs. DDR memory typically lasts 10–12 years, but server refresh cycles run every 4–5 years, leaving reclaimed DDR4 modules with roughly half their usable life remaining when machines are replaced.
Meta's advantage lies in vertical integration: the company designed the CXL controller chip internally, wrote its own firmware, built its own fleet management software, and installed the boards in in-house-designed servers—avoiding the multi-vendor coordination that other organizations must navigate. This level of control allowed Meta to bake telemetry and reliability features directly into the hardware. Most enterprises attempting memory reuse must select a third-party CXL controller, validate it with a separate board vendor, and then deploy into off-the-shelf servers from companies like Lenovo, each with its own BIOS and CXL feature support.
Rising DDR5 costs have sharpened the economic case for DDR4 reuse. Until last year, the cost of CXL expansion hardware made it cheaper to simply add more DDR5; that calculation has flipped, pushing more hyperscalers into pilot programs. Marvell and other chipmakers offer CXL controllers like Structera X that support hardware-based compression, potentially doubling effective memory capacity. For data center operators, Meta's public disclosure signals that memory reuse is moving from pilot to production, with implications for capex efficiency and DRAM sourcing across the cloud infrastructure industry.