AMD has acquired MEXT, a company whose software blurs the line between fast-but-expensive DRAM and slower-but-cheaper NAND flash storage, according to Tom's Hardware.
The core technology AMD is after is MEXT's Predictive Memory Engine. It works by monitoring which data in memory is being actively used and which is sitting idle — then quietly shifting the cold data off expensive DRAM and onto flash storage. To the applications running on top, nothing changes: flash appears to behave like DRAM.
AMD framed the deal as a move to "address growing memory constraints" in the data center — a nod to a real and worsening bottleneck. AI workloads in particular are notorious for requiring enormous amounts of fast memory, and DRAM remains costly and physically limited in how much can be packed into a server.
Memory tiering is not a new idea, but making it work transparently — without rewriting applications or manually managing what lives where — is the hard part. That's what MEXT's predictive approach is designed to solve.
This matters because the economics of AI infrastructure are increasingly shaped by memory costs, and any technology that lets data centers squeeze more effective capacity out of cheaper storage could meaningfully reduce the price of running large-scale AI systems.