HBM4: The Bottleneck Has a Roadmap

SK hynix declared HBM4 development complete in late 2025 and moved to mass-production ramp in 2026, right on schedule for the next accelerator generation. The headline spec — a 2,048-bit interface, double HBM3E's width — matters less than a structural change buried in the stack: the base die is moving to a logic process.

Why the base die matters

An HBM stack is DRAM dies piled on top of a base die that handles the interface to the GPU. Through HBM3E, memory makers fabbed that base die themselves on DRAM-adjacent processes. With HBM4, the leading roadmaps put customized base dies on foundry logic nodes — which drags TSMC directly into the memory supply chain and lets accelerator designers ask for custom features in the memory stack itself.

Two consequences:

1. The memory-logic boundary blurs. When your memory vendor is co-designing a logic die with your foundry, "memory supplier" starts to look more like "packaging partner." The companies that master this interface — and the advanced packaging (CoWoS and successors) that assembles it — capture margin that used to be sliced between separate industries.
2. Customization fragments the commodity. Commodity DRAM is fungible; a custom-base-die HBM4 stack tuned for one customer's accelerator is not. Long-term agreements get longer. Switching costs rise. The cartel-without-a-cartel gets stickier.

The capacity race

All three producers are converting and building capacity at once: SK hynix at Cheongju and its Yongin megafab program, Samsung trying to reclaim qualification ground it lost in the HBM3E cycle, Micron scaling from a low base with U.S. fab ambitions attached. TrendForce-class estimates have HBM bit supply growing at rates that would terrify any commodity analyst — unless demand from accelerator roadmaps (each generation carrying more stacks per package) grows faster still.

The pattern to watch is qualification announcements. In this market, the press release that moves billions is not "we built capacity" but "our 12-high stack passed qualification at the lead customer." Yield on thermal-compression-bonded 12- and 16-high stacks remains the hard part, and the vendor who ships qualified volume first effectively sets the price for the generation.

Memory used to be where semiconductor profits went to die. In the HBM4 era, it is where accelerator roadmaps go to live or stall.