# AMD’s Gary Smerdon Says Mext Could Cut Server Costs by Up to 83%

By Simon Yoon

Canonical URL: https://www.tokenpost.com/news/technology/54795
Published: 2026-11-12T02:12:49.000Z
Updated: 2026-11-12T02:12:49.000Z
Section: Technology

> AMD says an internal rollout added roughly 58 terabytes across 86 to 145 hosts for about $601,000, compared with a $3.3 million DRAM expansion estimate.

Gary Smerdon, AMD’s corporate vice president of business development, says Mext Predictive Memory could lower server costs by as much as 83% by transferring some memory demand from DRAM to NVMe flash.

Smerdon’s estimate centers largely on replacing registered memory modules, known as RDIMMs, with flash. That broad cost figure is separate from AMD’s estimate for an internal deployment, which compares the cost of adding capacity through Mext with an equivalent DRAM expansion.

AMD says its initial rollout covered 86 to 145 hosts and added about 58 terabytes of usable capacity for roughly $601,000. The company estimated that a DRAM expansion providing the same capacity would have cost $3.3 million. AMD calculates an 82% saving for that comparison.

Mext’s software uses AI models to predict what data applications will need and move it from NVMe flash into DRAM before a request arrives. DRAM serves as the faster working memory, while flash adds capacity. In conventional systems, Linux uses swap and Windows uses a page file to move data to slower storage after memory pressure develops. An application can be delayed while requested data returns to memory.

AMD says that on certain workloads, systems with more than half of their physical DRAM removed retained 95% to 99% of performance. The company also estimates that mixed DRAM-and-Mext configurations delivered between 1.9 and 3.5 times as much performance per dollar.

Memory capacity has become a pressure point for AI and high-performance computing workloads, which can require large pools of fast memory. DRAM prices have also been climbing alongside demand.

Mext was founded in 2023 and launched Predictive Memory in April 2026. AMD acquired the company on June 15, about two months after the software launch.

AMD plans to expand the technology across thousands of servers and is targeting more than $60 million in cumulative savings by the end of 2027. The company’s figure is a target tied to the planned expansion, not savings already achieved.