Meta Starts Building Servers from Old DDR4 Memory Amid DRAM Shortage

Meta (the company's social networks are recognized as extremist and banned in Russia) has started assembling new servers from used DDR4 RAM modules extracted during the decommissioning of old machines — this way the company is partially solving the problem of server DRAM shortage, writes CNews.

How memory reuse works

Technically, the task is not trivial: old DDR4 modules are physically and electrically incompatible directly with new server platforms designed for modern memory types like DDR5. To link obsolete and cutting-edge hardware in one system, Meta uses a custom adapter built on a specialized microchip (ASIC) called Vistara — it acts as a "translator" between memory protocols of different generations, allowing a new server to query and use old modules as if they were its standard RAM.

• Memory source — DDR4 modules extracted during decommissioning of old Meta servers
• Compatibility is ensured by a custom adapter based on the Vistara ASIC
• Goal — partially compensate for server DRAM shortage without buying new modules

Why there is not enough memory on the market

The global shortage of RAM in 2026 is largely associated with the artificial intelligence boom. Memory chip manufacturers — Samsung, SK Hynix, Micron — are redirecting production capacity to HBM (high-speed memory installed next to AI accelerators like Nvidia H100 and Blackwell), which is sold at higher prices and is in higher demand than traditional DDR memory for regular servers and personal computers. As a result, regular DRAM is physically produced less, and its prices are rising for everyone — from hyperscalers with data centers to personal computer assemblers and smartphone manufacturers.

For companies like Meta, which simultaneously build giant AI data centers and continue to update regular server infrastructure for social networks and recommendation systems, this creates double pressure: new AI clusters require HBM, and regular servers require DDR4 or DDR5, which are becoming physically scarcer on the open market. Reusing memory from decommissioned equipment is a practical engineering way to circumvent this shortage at least partially, without waiting for new supplies from manufacturers and without overpaying for them against the backdrop of rising prices.

ASIC (application-specific integrated circuit) is a microchip designed for one specific task, as opposed to universal processors. Such a narrowly specialized approach usually provides a gain in energy efficiency and cost compared to using more expensive and universal controllers, which is especially important when scaling the solution to thousands of servers inside hyperscaler data centers.

Such an approach also has a positive side effect: reusing still-functioning memory modules instead of their immediate disposal reduces the volume of electronic waste from data centers — a topic that large technology companies are increasingly raising in sustainable development reports. If the solution proves successful, similar practices of component reuse may be adopted by other hyperscalers facing the same DRAM shortage against the backdrop of the AI boom.

What this means

Memory shortage, driven by AI infrastructure, is forcing even the largest internet companies to seek non-standard engineering solutions like reusing old hardware — this is an indirect but noticeable side effect of the global race for computing power for artificial intelligence.