By Hamza Faisal

AMD has unveiled software that could change how data centres buy memory. Its new MEXT predictive memory system lets servers run with far less DRAM — in some tests, half as much — while keeping almost all of their performance. In a market where memory prices are soaring and supply is tight, that is a pitch every server buyer will want to hear.

The software was presented this week at AMD's In Depth 2026 event in Las Vegas. It comes from Mext, a startup AMD acquired in June, and it is the company's first commercial product built from that deal.

AI predicts what the server needs next

MEXT watches how applications use memory and predicts what they will need next. An ensemble of AI models, including ones derived from transformer technology, learns each workload's access patterns. Pages of memory that are not being used — so-called cold pages — are moved from expensive DRAM onto much cheaper NVMe flash storage. Pages the application is about to need are pulled back into DRAM before it asks for them, so the swap is invisible to the software running on top.

The trade-off is stark. Flash costs roughly 50 times less per bit than DRAM and draws about 30 times less power — but it is also around 500 times slower. MEXT's bet is that smart prediction can hide that speed gap almost entirely.

In AMD's testing, halving a server's DRAM while backing the rest with MEXT flash kept 95 percent of all-DRAM performance on Synopsys PrimeTime, a chip-design verification workload, and 99 percent on Apache Spark. Even with three-quarters of the memory on flash, performance held at around 80 percent, while performance per dollar spent on DRAM rose by as much as 3.5 times. Depending on the workload, AMD says DRAM requirements can fall by 20 to 75 percent.

AMD is already using it on its own servers

AMD is not just selling the idea — it is running it. The company has deployed MEXT on 145 of its own servers so far, and says it has avoided about $5 million in DRAM purchases, an 83 percent cost reduction versus expanding memory the conventional way. It plans to expand to around 2,000 servers by 2027, targeting cumulative savings of more than $60 million by the end of that year.

The timing is no accident. DRAM prices have climbed sharply as AI data centres soak up supply, and memory can account for 60 to 90 percent of the cost of a large server, according to Gary Smerdon, Mext's former chief executive who now leads business development at AMD.

Licensed, not free — and AMD-only

There are limits. MEXT is a paid licence sold through AMD's hardware partners, and it works only on AMD hardware — initially sixth-generation EPYC servers. A companion observability tool, RAM View, is free. In virtualised environments the software needs one CPU core per virtual machine, and the prediction engine itself runs on a single EPYC core with no GPU required. AMD says workloads needing more than 64GB of DRAM stand to benefit most.

AMD has not disclosed pricing or a date for wider availability, and all of the performance figures come from its own tests, so independent results will matter. But if the numbers hold up in real data centres, MEXT could take real pressure off the memory market — and give AMD a selling point its rivals cannot match, because the software only runs on its chips.