DDR5 prices have surged 5-10 times, and obsolete DDR4 memory is being repurposed: CXL interface proves remarkably effective.

Memory prices have skyrocketed over the past year, with DDR5 memory now generally more than 5 times higher than this time last year, and some models seeing price increases of up to 10 times—even AI giants find it absurdly expensive.

If AI servers were to be fully equipped with high-spec memory like DDR5/LPDDR5X, not only would supply fall short, but costs would also spiral out of control. That's why repurposing DDR4 has become the trend—server manufacturers have been retiring large quantities of DDR4 memory, and new technology is now giving it a second life in AI servers.

We recently reported that Meta has developed Vistara technology, which redeploys previously retired DDR4 memory onto current AI server platforms.

The principle behind their approach is using CXL controller chips to connect DDR4 to platforms that originally required native DDR5 bus support. This means millions of decommissioned DDR4 modules no longer have to be sold as scrap—they can be reused.

This approach offers many benefits. Servers typically have a lifespan of just 4-5 years, while DDR4 memory can last 10-12 years. After redeployment, it can serve for many more years, significantly cutting costs, reducing demand for DDR5 memory, and helping curb the runaway price increases of DDR5.

There are downsides too: DDR4's bandwidth and latency can't match DDR5, so performance takes a hit. But not all AI applications demand peak performance—many workloads prioritize large memory capacity. By expanding DDR4 memory via CXL interfaces, stacking up to 2-2.5 times the memory capacity proves highly valuable.