AMD Just Acquired MEXT to Crack the Memory Optimization Problem. Should Micron and Sandisk Investors Be Nervous?

Micah Zimmerman, The Motley Fool

Sun, June 28, 2026 astatine 1:20 PM CDT 5 min read

About 2 weeks ago, Advanced Micro Devices announced the acquisition of MEXT, a start-up that has built artificial quality (AI)-driven bundle designed to marque NAND flash behave similar dynamic random-access representation (DRAM).

The exertion uses predictive algorithms to place often accessed information and determination it betwixt flash retention and high-speed representation successful existent time, reducing the magnitude of costly DRAM a information halfway needs to tally AI workloads astatine scale. According to MEXT's ain property release, the bundle tin chopped representation costs by astir fractional portion expanding usable representation capableness by 2 to 4 times.

Missed Nvidia successful 2009? This Rare Signal Is Flashing Again. In 2009, a "Double Down" awesome flashed for a little-known chipmaker called Nvidia. For the archetypal clip successful years, that aforesaid "Total Conviction" awesome is flashing for a institution 1/100th the size of Nvidia. Continue »

For investors successful Micron Technology (NASDAQ: MU) and Sandisk (NASDAQ: SNDK), the knee-jerk work is obvious: If AMD tin thatch flash to behave similar DRAM, request for high-bandwidth representation contracts declines. The knee-jerk work is terribly wrong.

What MEXT really does (and doesn't do)

MEXT's exertion operates successful the bundle tier betwixt existing retention and compute. It doesn't regenerate DRAM oregon HBM. Instead, it reduces the magnitude of high-speed representation definite workloads necessitate by optimizing what lives successful it astatine immoderate fixed moment. That's a meaningful ratio summation for endeavor customers moving general-purpose AI workloads, wherever representation is simply a outgo constraint.

What it cannot interaction is the physics of grooming ample AI models and moving inference astatine the show levels that hyperscalers require. An Nvidia Blackwell graphics processing portion (GPU) demands HBM4 not due to the fact that nary 1 has tried to enactment astir it, but due to the fact that the bandwidth requirements of grooming trillion-parameter models are architectural constraints, not bundle problems. No predictive tiering algorithm changes what the silicon needs.

MEXT is simply a instrumentality for enterprises trying to agelong existing infrastructure. It is not a substitute for the representation products that Micron and Sandisk merchantability to monolithic tech companies.

Micron's presumption is structurally insulated

Micron Technology's full 2026 HBM4 accumulation is sold retired nether binding multi-year contracts. At COMPUTEX 2026 successful May, the institution laid retired an end-to-end AI representation portfolio spanning information halfway to intelligent edge, each successful high-volume production. Fiscal first-quarter 2026 gross deed $13.64 billion, up 57% twelvemonth implicit year, with gross margins astir 56%, driven by HBM pricing powerfulness that comes from contracted scarcity.