Meta Built a Custom Chip to Reuse Old RAM — and It Just Cut Server Counts by 25%

Meta's Vistara CXL chip lets the company reuse memory from decommissioned servers in new machines, cutting costs and carbon emissions across its hyperscale fleet.

Meta has found a way to reuse old server RAM in new machines — and the savings are substantial.

About 40% of Meta's millions of servers are memory-capacity bound, limiting performance and scalability. At the same time, the company has a surplus of older DIMMs from decommissioned servers, because RAM chips can last about twice as long as the rest of the machine.

To bridge this gap, Meta developed a custom Computer Express Link (CXL) chip called Vistara, along with supporting software, that decouples older memory from server memory channels and makes it usable in new machines alongside their native memory. Using the older RAM with the CXL interface does not significantly affect performance — which would have been a problem if the old DIMMs were plugged directly into newer servers.

The results, published in a paper at ISCA 2026, are striking: up to a 25% reduction in server count for disaggregated ML inference, a 29% reduction in average latency for distributed caches, and a 50% drop in out-of-memory events for certain workloads. The paper also refutes common misconceptions about CXL, demonstrating that with careful engineering, tail latencies are comparable to local DRAM even under high contention.

There is also a sustainability angle. Meta's lifecycle analysis found that DRAM DIMMs account for 69% of server component carbon emissions — more than CPUs, SSDs, NICs, and boards combined. By extending the service life of existing DIMMs, Vistara directly reduces the need for new DRAM production.

The timing could not be better. Memory prices are projected to double by the end of 2026, with shortages potentially lasting into 2027. Apple reportedly explored using cheap Chinese chips as an alternative, but Meta's approach of reusing what it already owns may prove the more efficient path forward.

Sources: Network World, Meta Research — ISCA 2026 Paper