HBM4’s Second Supply Chain: Logic Base Dies and Package Readiness

Research question: How is HBM changing semiconductor dependency compared with the conventional DRAM supply chain?

The HBM relationship is more concentrated than ordinary DRAM

The clearest contrast is downstream.

Altsets shows SK hynix supplying both NVIDIA and Apple, but the economic shape of those relationships is very different. NVIDIA represents 27.88% of SK hynix revenue, while SK hynix represents 27.33% of NVIDIA COGS. By comparison, Apple represents 8.75% of SK hynix revenue, while SK hynix represents only 1.76% of Apple COGS.

These are company-level relationships, not product-level HBM disclosures, so they should not be read as proof that all NVIDIA exposure is HBM or that all Apple exposure is conventional DRAM. But the contrast is still useful. The Apple relationship looks like the traditional memory model: a large device maker can be important to the memory supplier without the supplier being equally important to the customer's total cost base. The NVIDIA relationship is much closer to bilateral dependence.

Micron shows the same pattern. NVIDIA represents 17.62% of Micron revenue, while Micron represents 14.00% of NVIDIA COGS. That is materially more balanced than the classic commodity-memory relationship investors are used to seeing.

This is the first major supply chain change created by HBM. The memory vendor is no longer just selling standardized bits into a broad electronics market. It is supplying a high-value subsystem that must be qualified against a specific accelerator platform, with capacity, performance and roadmap decisions coordinated much earlier with a smaller number of customers. SK hynix said in its latest quarter that it has long-term agreements with around 10 key customers, reinforcing how AI memory is moving toward deeper customer alignment rather than purely spot-driven commodity supply. [2]

HBM adds a second upstream chain

The second change is upstream. The conventional DRAM supply chain does not disappear. Micron still depends on the same fabrication ecosystem that makes leading-edge memory possible.

Altsets estimates that ASML represents 11.91% of Micron COGS, while Micron represents 7.64% of ASML revenue. Lam Research is nearly balanced at 5.52% of Micron COGS and 5.61% of Lam Research revenue. KLA represents 2.84% of Micron COGS, Applied Materials 3.84%, and Shin-Etsu Chemical 1.34%.

Those relationships matter because HBM still begins with advanced DRAM dies. More HBM does not reduce the importance of lithography, deposition, etch, inspection or silicon materials. It layers additional manufacturing dependencies on top of them.

The most important new layer is the logic and packaging interface. SK hynix said this year that HBM4 uses TSMC advanced logic for its base die and that the two companies are expanding collaboration around memory and logic integration. [3] That is a meaningful structural change from ordinary DRAM. A foundry that historically sat mostly in the processor supply chain is now moving directly into the architecture of a memory product.

That gives TSMC a new type of economic relevance to memory makers. The value is not simply more wafer demand. HBM4 links the DRAM stack, logic base die and accelerator package more tightly, which means the performance of one supplier increasingly depends on process decisions made by another. The supply chain becomes less linear and more co-designed.

For investors, this is why advanced packaging, base-die manufacturing and memory-logic integration deserve to be treated as a separate semiconductor exposure rather than as footnotes to DRAM.

The investment exposure is shifting from bits to interfaces

The conventional memory cycle still depends on bit supply, utilization and pricing. HBM adds a completion constraint across interfaces: qualified DRAM stacks, logic base dies, bonding and accelerator integration all have to be ready together.

A hypothetical build can have capacity for 100 DRAM stacks, 80 compatible base dies and 70 completed package assemblies. Under a one-of-each simplified bill of materials, output is limited to 70 completed units. Adding 20 DRAM stacks alone produces no additional finished units. The counts are illustrative; actual architectures and stack requirements differ.

That is the useful distinction between an HBM interface thesis and a memory-volume thesis. The company relationships identify candidate suppliers at each interface, while product qualification and capacity evidence identify the limiting stage. TSMC's disclosed logic-base-die role gives a concrete reason to include foundry readiness in the memory model rather than treating it only as a processor input.

For systematic research, separate memory pricing and wafer output features from base-die qualification and package-completion features. A model that uses only more DRAM capacity can miss an HBM delivery constraint. A historical test should ask whether interface milestones explain delivery or forecast revisions beyond ordinary memory-cycle variables.

The key point is that HBM has not replaced the DRAM supply chain. It has created a second semiconductor supply chain on top of it.

The move from ordinary DRAM toward HBM is changing semiconductor dependency in two directions at once. Downstream, memory suppliers are becoming more economically tied to a small group of AI accelerator customers. Upstream, memory manufacturing is expanding beyond the classic wafer-fab stack into logic base dies and advanced packaging.

Altsets data makes the difference visible. SK hynix's relationship with Apple remains heavily weighted toward the supplier side, while its relationship with NVIDIA is close to economically bilateral. Micron shows a similar NVIDIA concentration. At the same time, ASML, Lam Research and other conventional fab suppliers remain important, while TSMC is entering the memory architecture itself.

That is the second supply chain investors need to follow: not just who makes DRAM, but who controls the interfaces that turn DRAM into HBM.

Sources

1. "Micron Technology, Inc. Reports Record Results for the Third Quarter of Fiscal 2026," Micron Technology, June 24, 2026. https://investors.micron.com/news/press-release/2026/Micron-Technology-Inc--Reports-Record-Results-for-the-Third-Quarter-of-Fiscal-2026/default.aspx
2. "SK hynix Announces 2Q26 Financial Results," SK hynix, July 29, 2026. https://news.skhynix.com/en/q2-2026-business-results/
3. "Overcoming Technological Limits Requires the Integration of Memory and Logic: SK hynix Presents Vision at TSMC Symposium," SK hynix, April 23, 2026. https://news.skhynix.com/en/tsmc-technology-symposium-2026/

HBM therefore creates a second research layer below the memory manufacturers: the HBM supplier reference maps equipment and materials, while the second-tier suppliers beneath Nvidia shows how exposure can become more concentrated farther upstream.

Go Deeper