Close-up of a circuit board with memory and processor chips
Investing

AI's memory

Micron & SanDisk: The Memory Chips Powering the AI Boom — and Why They Matter to Malaysia (2026)

Rates.my Editorial7 min read

The AI race is usually told as a compute story — Nvidia, GPUs, ever-bigger data centres. But in 2026 the binding constraint has quietly moved. The GPUs are starved for one thing: memory. Two American chipmakers sit right at that chokepoint — Micron (NASDAQ: MU) and SanDisk (NASDAQ: SNDK) — and the story runs straight through Malaysia.

3
HBM makers on Earth
SK Hynix · Micron · Samsung
~US$100b
HBM market by 2028
≈3× the 2025 size
US$2b
Micron's Penang bet
2024 expansion

Why memory — not just GPUs — is the bottleneck

A GPU is only as fast as the data you can feed it. Today's AI models are enormous, and a big chunk of the time the expensive GPU is simply waiting for weights and activations to arrive from memory. Engineers call this the “memory wall.” The fix is HBM (High Bandwidth Memory) — stacks of DRAM bonded right next to the GPU so data has barely any distance to travel. Every modern Nvidia AI accelerator ships wrapped in it.

Close-up of server cooling fans inside a modern data centre
Inside a modern AI data centre — racks of accelerators, every one hungry for memory bandwidth. · Photo: Unsplash

Training and running these models also needs somewhere to keep all that data — datasets, model checkpoints, vector databases, caches. That job falls to NAND flash and the enterprise SSDs built from it. So the AI build-out actually needs two different kinds of memory: the ultra-fast HBM/DRAM that feeds the GPU, and the vast, fast NAND storage that surrounds it.

HBM vs NAND, in one line. HBM is the GPU's short-term working memory (speed); NAND flash is its long-term storage (capacity). AI needs a lot more of both — and that is the whole investment thesis here.
A 9-minute explainer on why HBM is the memory at the heart of every AI accelerator.

Micron: the only American HBM maker

Micron makes DRAM, NAND and — crucially — HBM. It is one of just threeHBM suppliers on the planet, alongside South Korea's SK Hynix and Samsung, and the only US-based one. Its HBM3E went into the Blackwell generation of Nvidia GPUs; in early 2026 Micron began high-volume production of HBM4 for Nvidia's next Vera Rubin platform, roughly doubling the data path (a 2,048-bit bus) and pushing past 2.8 terabytes per second per stack.

The tell is in the order book. Micron has said its entire 2026 HBM output is already committed under contract, and it has signed unusually long multi-year supply deals — the kind of thing that only happens when buyers are terrified of being left short.

HBM is effectively sold out across all three makers through 2026 — and supply still can't catch demand.
Industry capacity reports, 2026

SanDisk: the storage half of the story

For decades SanDisk was the flash-memory brand living inside Western Digital. In February 2025 it was spun back out as a standalone, Nasdaq-listed pure-play in NAND flash and SSDs (Western Digital kept the spinning-disk hard-drive business). That makes SanDisk a focused bet on the othermemory squeeze of 2026: a tightening NAND market, as AI clusters write staggering volumes of data and demand fast enterprise SSDs to keep the GPUs fed. SanDisk's NAND is manufactured through a long-running partnership with Japan's Kioxia, and its shares have re-rated sharply through 2026 as the shortage took hold.

Memory is brutally cyclical. DRAM and NAND are the most boom-and-bust corner of tech. Prices and profits swing hard, and today's shortage has a habit of becoming tomorrow's glut. A vertical share-price chart is a reason for caution, not FOMO — size any position accordingly.

Why this matters to Malaysia

This isn't just US news. Malaysia is woven into the memory supply chain. Micron has operated here since 2010 in Muar, Johor, and 2018–2020 in Prai and Batu Kawan, Penang, employing roughly 5,700 people locally — and in late 2024 it committed another ~US$2 billion to expand in Penang. Penang and Kulim are the country's semiconductor back-end heartland — the assembly, test and packaging work that scales directly when AI chip volumes explode.

Petronas Twin Towers against the Kuala Lumpur skyline
Malaysia is quietly inside the chip supply chain — Micron's Penang and Johor plants put local hands on the AI build-out. · Photo: Unsplash

And on Bursa, a cluster of “pick-and-shovel”names ride the same wave: OSAT and test players like Inari Amertron, Malaysian Pacific Industries (MPI), Unisem and Globetronics; and equipment and automation names like ViTrox and Greatech. They don't make HBM themselves — they package, test and tool the chips that do — so a memory supercycle tends to lift the whole Penang ecosystem.

What it means for a Malaysian investor

  • Direct exposure: you can own MU or SNDK through a broker that offers US shares, or indirectly via a semiconductor ETF, a unit trust, or your EPF external / i-Invest funds — check what they actually hold.
  • A home-market angle:the Bursa chip names give correlated exposure in ringgit, but they're smaller and more volatile than the US giants.
  • Respect the cycle:memory swings harder than almost anything else in tech. Don't chase a chart that has already gone vertical.
  • Know your real exposure: AI and semiconductors may already sit inside your ETFs and EPF funds — the same lesson from the recent AI selloff.

The bottom line

You can't run the models without somewhere to put the data — and that is exactly where Micron and SanDisk come in. For Malaysian investors, the memory supercycle is both a US-stock story and a Penang story: watch the names, respect the cycle, and remember the boring rule that beats the hype — diversify. See where the money is moving on today's trending US and Bursa stocks.

Sources & further reading

Keep reading

This article is general information, not personalised financial advice. Rates.my is not a licensed financial adviser — always verify rates with the institution and consider your own circumstances.

← All insights