64-GB DDR5 DIMM prices have risen 3.5-fold between Q3 2025 and Q1 2026, according to Counterpoint Research, and AMD is betting a small startup can ease the cost pressure by treating flash as a memory extender. Counterpoint projects the same module to reach five times its Q3 2025 price by Q3 2026, and Micron's most recent outlook sees no relief before 2028. The result is a sharp reordering of the server bill of materials. DRAM, which represented about half of a server's cost in 2023, now averages 75 percent of the build, with a range of 60 to 90 percent depending on configuration.

Where did MEXT come from?

MEXT emerged from stealth in early April, and AMD closed the acquisition within weeks. Neither side has disclosed the price. The company was founded in 2023 by Gary Smerdon, who built solid-state memory programs at LSI Logic and later ran Fusion-io, where Apple and Meta Platforms were anchor customers more than a decade ago. Smerdon went on to co-found TidalScale, which raised $70.3 million across multiple rounds before Hewlett Packard Enterprise acquired it in December 2022.

The 39-person team draws heavily from TidalScale. Co-founder David Reed held chief scientist roles there and at Lotus Development and remains a longtime MIT computer science professor. Carl Waldspurger, the principal engineer at VMware who architected the Distributed Resource Scheduler behind live migration in ESX, joined as chief scientist. The bench combines hyperscale flash shipping experience with deep operating-system memory expertise.

How does flash-based memory extension work?

Flash costs 50 times less per bit than DRAM today, up from a similar ratio in 2023 to perhaps 100 times now, and draws roughly 30 times less power per bit. The catch is speed: flash is about 500 times slower than DRAM. Pushing cold pages out is straightforward; the hard case is data that turns hot between one instruction and the next.

MEXT calls its approach Predictive Memory. Machine-learning models built on sequence-prediction techniques borrowed from large language models watch access patterns and pull pages back from flash before the operating system or an application asks for them. "We have developed sophisticated machine learning models that have much better prediction accuracy and coverage than what has been done in the past," chief scientist Waldspurger said. The software runs on Linux on X86 or Arm processors and supports bare metal, virtual machine hypervisors, and Kubernetes. The recommended configuration pairs equal amounts of DRAM and flash, though the company has validated 25 percent DRAM with 75 percent flash. Memory extension works whenever software can keep the right pages ready before they are needed.

What do the benchmarks show?

On a Dell PowerEdge platform, the extended configuration reached 3 TB and 6 TB effective capacities using 1.5 TB of DRAM, after Dell stopped shipping 3 TB and 6 TB main memory options for the R6725 on February 1. Testing on an AMD "Zen 5" Epyc 9755 with 128 cores, twelve 6.3 GHz memory channels, and a Kioxia NVMe flash drive, the team ran Redis, Memcached, and Neo4j with 32 GB of DRAM plus 32 GB of flash allocated to memory. That configuration delivered 1.7 times the price-performance of a 64-GB all-DRAM baseline at roughly 95 percent of its throughput.

A separate Houdini swap test on a 64-core Threadripper Pro 5995WX with 64 GB of memory and 1 TB of flash widened the disparity. Cutting real DRAM to 16 GB pushed the workload to 22,000 seconds. Running MEXT on the same 16 GB with flash extension dropped that to 5,000 seconds, a 4.4-fold improvement. Performance per dollar climbs as the DRAM share falls, even though raw throughput drops when flash takes more of the address space.

What comes next?

AMD has not disclosed how MEXT will be integrated or priced. MEXT had been charging $3.99 per GB per year under subscription before the deal, and AMD has not said whether the technology will remain available to non-AMD processors. The cautionary tale is Intel's 3D XPoint, launched with Micron in July 2015, killed by 2022, and sold to SK Hynix along with Intel's flash business. The three large memory makers continue to shift output toward HBM at the expense of standard DIMMs, and aggregate DRAM and flash capacity grows only 20 to 25 percent per year against demand rising far faster. Whether AMD keeps MEXT open across processor architectures or restricts it the way Intel did will determine how broadly the bottleneck eases.