AMD Acquires MEXT, Turns Flash Into DRAM for AI
Key insights
- MEXT's Predictive Memory software, launched April 2026, makes flash storage behave like DRAM with microsecond speeds on a single CPU core.
- AMD claims MEXT's technology expands usable memory 2-4x and cuts infrastructure costs up to 50% with no hardware, OS, or application changes.
- AMD brings MEXT's team into its data centre portfolio to address mounting memory access constraints in AI and HPC workloads.
Why this matters
Memory capacity is the binding constraint on AI inference at scale, and MEXT's no-hardware-change approach lets operators expand it using flash storage they already own. A claimed 50% infrastructure cost reduction with zero code changes is the kind of drop-in economics that can shift procurement decisions faster than any chip upgrade cycle. AMD gains a software foothold at the memory layer inside data centres, a differentiated position at a moment when GPU compute leadership remains concentrated elsewhere.
Summary
AMD has acquired MEXT, whose Predictive Memory software makes flash storage behave like DRAM, targeting memory capacity as AI infrastructure's core bottleneck.
Launched April 2026, MEXT proactively moves memory pages to flash and restores them before workloads need them. No GPU, no hardware or OS changes, single CPU core.
Essentially: (AMD, MEXT) offer a software layer that multiplies usable memory without hardware changes.
- MEXT claims 2-4x usable memory capacity growth and up to 50% lower infrastructure costs.
- The software works on-premises and in the cloud with microsecond prediction speeds.
- AMD SVP Dan McNamara: memory gaps limit 'performance per dollar, operational efficiency, and the pace of large-scale deployments.'
AMD integrates MEXT's team into its data centre portfolio; acquisition terms were not disclosed.
Potential risks and opportunities
Risks
- If AMD makes MEXT's software exclusive, multi-vendor data centre operators face unexpected ecosystem lock-in tied to AMD hardware at renewal time.
- Flash carries higher write amplification and wear than DRAM; sustained page-swap workloads at scale could erode the claimed 50% cost savings over equipment lifetimes.
- Intel and hyperscalers running proprietary memory-tiering solutions could narrow AMD's differentiation window before MEXT's technology reaches broad commercial deployment.
Opportunities
- Enterprise data centre operators can pilot MEXT's no-hardware-change approach to expand AI inference capacity without new rack capital expenditure.
- AMD can bundle MEXT's Predictive Memory software with EPYC CPU and Instinct GPU products to compete on total infrastructure economics rather than compute silicon alone.
- Flash storage vendors (Kioxia, Western Digital, Micron) stand to benefit if AMD's endorsement accelerates adoption of flash as working memory across AI clusters.
What we don't know yet
- Acquisition price: financial terms were not disclosed and no indication was given of when or if they will be made public.
- Whether MEXT's Predictive Memory software will remain available to non-AMD hardware platforms or become exclusive to AMD's infrastructure stack.
- No independent benchmarks or named production deployments were cited to validate the claimed 2-4x capacity expansion and 50% cost reduction at real-world scale.
Originally reported by verdict.co.uk
Read the original article →Original headline: AMD Acquires MEXT to Address AI Memory Bottlenecks With Tech That Makes NAND Act Like DRAM