How Meta Is Reusing Retired DDR4 to Expand AI Server Memory

Meta’s answer to tight and expensive DRAM supply is not a simple adapter that inserts an old DDR4 DIMM into a new DDR5 motherboard. Instead, it separates the useful memory modules from the retired servers they came from.

Its Vistara design places reclaimed DDR4 behind a custom Compute Express Link (CXL) memory-expander ASIC, making the modules available to a modern DDR5-based host as a separate, slower capacity tier. For workloads constrained more by memory capacity than by memory latency, Meta says the result can reduce the number of servers required by as much as 25%. 1

The key distinction: CXL expansion, not a DDR4-to-DDR5 socket adapter

DDR4 and DDR5 are different memory generations; the practical reuse route described by Meta is to put the older RDIMMs in a dedicated memory appliance and connect that appliance to the host through CXL.

Vistara is a CXL 2.0/1.1 Type-3 memory expander that uses a PCIe Gen5 x16 link to bridge DDR4 memory to the host processor. Each ASIC integrates two DDR4 channels and can support up to 256 GB when paired with 64 GB DIMMs. 29

That architecture allows a newer server to retain its directly attached DDR5 while gaining additional capacity from recovered DDR4. Rather than buying enough new DRAM to cover every server’s peak capacity requirement, an operator can use the CXL-attached tier where extra capacity is valuable.

Why the software policy matters as much as the hardware

CXL-attached DDR4 is not a substitute for local DDR5 on latency- or bandwidth-sensitive paths. In reporting on Meta’s production implementation, the expanded tier delivered roughly one-tenth of the bandwidth of local DDR5 and about 60% higher latency. 4

The system therefore treats the CXL capacity as a separate NUMA-like memory tier. Hot or latency-sensitive data can remain in local DDR5, while less frequently accessed pages can move to the lower-cost DDR4 tier. 413

This is why the design is best understood as tiered memory:

• Local DDR5: fast memory for active, latency-sensitive data.
• CXL-attached reclaimed DDR4: additional capacity for data that can tolerate lower bandwidth and higher latency.
• OS and workload policy: determines which data belongs in each tier.

The business benefit comes from fitting a suitable workload into fewer machines—not from improving the inherent speed of DDR4.

Meta’s reported efficiency gains

Meta has said its CXL memory expansion is deployed across millions of servers and can reduce the required server count by up to 25% for some disaggregated inference workloads. 1

That is a fleet-efficiency result. If a workload needs fewer servers because each machine can hold more useful memory, the operator may also avoid some of the associated CPUs, networking hardware, rack space, and power demand. The result should not be generalized to every AI workload: a workload dominated by memory bandwidth or latency may see little benefit, or require more careful placement policy.

Meta has also reported a 29% reduction in average query processing time for a specific distributed-cache workload after adding capacity through the tier, illustrating that more memory can help when it improves cache effectiveness. 15

What Google and other hyperscalers are doing

Public reporting indicates that Google is also recovering DDR4 from retired servers as DRAM supply tightens. However, the available public reporting provides substantially less technical detail about Google’s implementation than Meta’s Vistara disclosures. 14

The broader takeaway is that large operators with extensive retired-server fleets have an asset that smaller buyers often do not: a large installed base of compatible, already-owned DIMMs. CXL-based expansion and pooling give those operators a way to make that inventory useful in newer systems instead of treating it as e-waste.

Why reclaimed memory is attractive in this market

The economics are being shaped by an AI-led memory squeeze. Conventional DRAM contract prices reportedly rose 90–95% quarter over quarter in the first quarter of 2026. 1 J.P. Morgan Global Research estimates that DRAM prices will be more than 400% higher at the end of 2026 than at the start of 2024. 20

Memory is also becoming a larger portion of cloud infrastructure spending. A TrendForce-based projection reported that DRAM and NAND combined could account for 68% of major cloud providers’ hardware capital expenditure in 2027, up from 47% in 2026. 18

In that environment, a reclaimed DIMM is not merely a sustainability win. It can be a strategically useful unit of capacity that does not need to be purchased in the constrained spot or contract market.

Why more fabs may not end the shortage quickly

AI infrastructure drives demand for both conventional server DRAM and high-bandwidth memory (HBM). Suppliers have been directing capacity toward HBM for AI systems, tightening the supply of other memory products. 1920

Even with major investment, new semiconductor capacity takes time to build and qualify. Deloitte notes that additional supply from current expansion plans is not expected to come online until 2029 or 2030.

Supply concentration compounds the risk. Samsung and SK Hynix together control about two-thirds of the DRAM market, according to Reuters, limiting the industry’s redundancy when demand rises sharply or production is disrupted.

SK Hynix CEO Kwak Noh-jung has forecast the industry’s worst supply shortage in 2027 and said customer demand could exceed the company’s production capacity beyond 2030. That is a company executive’s forecast, not a certainty, but it helps explain why hyperscalers are pursuing reuse, memory tiering, and disaggregation now rather than waiting for new fab output.

The practical lesson for infrastructure teams

Meta’s model is compelling only when three conditions align: the operator has a meaningful stock of retired memory, the platform supports CXL expansion, and the workloads can tolerate a clearly slower second memory tier.

For those cases, old DDR4 can become useful capacity again. For latency-critical workloads, it remains the wrong place for hot data. The central innovation is not an adapter—it is a system design that combines a CXL controller, recovered DIMMs, and data-placement software to turn legacy memory into a managed resource.