Micron Bets on Physical AI to Launch a Decade-Long Memory Cycle as Korean Makers' HBM4 Progress Faces Commodity DRAM Price Hike Disruption
Taylor Wilson
Micron projects physical AI will trigger a decades-long memory demand cycle around 2030, but Korean memory makers face a near-term disruption — commodity DRAM margins overtook HBM in Q2 2026 — and the outcome of HBM4 pricing talks will determine whether the dual-track thesis holds.
What does "physical AI" actually mean for memory?
Micron defines "physical AI" as AI systems operating directly in the real world — not just in software — pushing memory demand from the cloud to edge devices: smartphones, cars, industrial equipment, and humanoid robots.
The anchor number: L2+ autonomous vehicles need more than five times the memory of a standard car. This means → every smart vehicle on the road adds a "mobile data center" worth of memory demand.
Future Markets projects the global physical AI market will grow from roughly $383 billion in 2026 to $3.26 trillion by 2040. In plain terms = the memory industry's growth engine is shifting from "centralized cloud procurement" to "distributed deployment across millions of edge devices."
Why did commodity DRAM margins suddenly overtake HBM?
Sigmaintell data shows Q2 2026 contract prices for 4GB LPDDR4X rose 75% quarter-on-quarter; 12GB LPDDR5X rose 89%. Commodity DDR profitability now broadly exceeds HBM across major memory makers.
The driver: cloud providers keep restocking, server-grade commodity DRAM channel inventory has dropped to just two to three weeks, and Korea's KB Securities estimates client memory demand is only about 50% fulfilled — with the shortage expected to deepen into 2027.
This reflects a counterintuitive reality: HBM is AI's marquee product, but in the short term the less glamorous commodity DRAM is generating returns faster.
SK Hynix is slowing HBM4 conversion — what does that signal?
SK Hynix is slowing the pace of converting production lines to HBM4 — the sixth-generation high-bandwidth memory — and shifting some capacity to commodity DRAM.
Industry sources stress this is a short-term structural adjustment, not a strategic pivot. HBM volumes are typically locked in advance with customers; contract commitments leave limited room to reallocate capacity.
In plain terms = Korean makers are not abandoning HBM4. They are filling the gap before large-scale HBM4 production ramps in H2 2026 by making more of the product that is more profitable right now.
Why does HBM4 pricing pull Nvidia into the picture?
Nvidia's next-generation AI accelerator platform Vera Rubin will use HBM4 heavily; memory already accounts for roughly 25% of the platform's total cost. This means → every round of HBM4 price increases pushes Vera Rubin's production cost higher.
Rising commodity DRAM margins strengthen memory makers' bargaining power in HBM4 negotiations. In plain terms = makers can point to commodity DRAM as a profitable alternative, raising their pricing floor.
HBM4 itself is harder to make: it integrates logic chips, stacks more layers, and has more I/O — yield and thermal challenges exceed earlier generations — which also constrains available capacity for commodity DRAM.
What does the second half of 2026 look like?
The industry expects memory makers' revenue and profitability to improve further once high-priced HBM4 begins shipping in H2.
A growing share of long-term commodity DRAM contracts should smooth price swings, letting makers benefit from commodity memory price gains + HBM4 product upgrades simultaneously.
This reflects the memory sector entering a rare "dual-track upside" window. Where HBM4 pricing ultimately lands is the pivotal test of whether this logic delivers.
Content is for reference only, not financial advice.