The US Can Build HBM Fabs. It Still Needs Korea, Japan and Taiwan

Building the memory in America only solves one part of the problem

SK hynix's Indiana project illustrates the distinction. The company held a groundbreaking ceremony in August 2026 for an advanced packaging facility in West Lafayette and said it expects next-generation HBM mass production there in the second half of 2029. The investment exceeds $4 billion. The facility is important, but its stated function is advanced packaging, not a wholesale transplantation of SK hynix's Korean DRAM fabrication network.

Micron is moving further upstream. Its Idaho expansion includes two planned leading-edge fabs, with DRAM output from the first expected in 2027. Micron explicitly says its US manufacturing investments will support products including HBM. The company also has major planned memory capacity in New York.

This gives the United States credible paths to domestic DRAM fabrication and HBM packaging. Yet the production sequence makes clear why fabs alone are not enough.

HBM contains vertically stacked DRAM dies linked using through-silicon vias. Beginning with HBM4, the base die beneath the memory stack becomes more sophisticated because it handles increasingly important logic and connectivity functions. SK hynix has partnered with TSMC to use TSMC advanced logic technology for its HBM4 base die and to optimize the integration of SK hynix memory with TSMC's CoWoS packaging platform.

Then there is another packaging step at the accelerator level. TSMC describes CoWoS as a technology that integrates system-on-chip devices with HBM stacks into a high-performance computing package. Nvidia's latest annual filing says it purchases memory from SK hynix, Micron and Samsung, uses TSMC and Samsung for semiconductor wafer production, and relies on CoWoS technology for packaging. Nvidia also states that its supply chain remains mainly concentrated in Asia even as it expands manufacturing elsewhere.

So an HBM stack assembled in Indiana can still contain technology produced through a supply chain extending into Korea, Taiwan, Japan, the Netherlands and other semiconductor manufacturing centers. Localization is better understood as changing the geography of individual production stages, not moving the entire network inside one border.

Korea still dominates the memory layer

The first remaining concentration is the most obvious one: HBM itself.

Counterpoint Research data reported in September put SK hynix at 50% of global HBM revenue in the second quarter of 2026 and Samsung Electronics at 33%. Micron held 18%. On that measure, the two Korean suppliers accounted for roughly 83% of the HBM market. Samsung's share had risen sharply from the prior quarter, but the geographic concentration of the leading suppliers remained intact.

That concentration matters differently from physical manufacturing concentration. A US SK hynix fab would reduce the amount of SK hynix production that must physically occur in South Korea. It would not create another independent HBM supplier. The market would still depend heavily on technology, qualification decisions, capacity allocation and capital spending by the same small group of memory companies.

Altsets relationship data makes the downstream side of that concentration visible. Nvidia represents 27.88% of SK hynix revenue, while SK hynix represents 27.33% of Nvidia's cost of goods sold under the Altsets relationship estimates. Micron also supplies Nvidia: Nvidia represents 17.62% of Micron revenue, while Micron represents 14.00% of Nvidia COGS.

Those figures are company-level economic relationships, not HBM-only measurements, so they should not be interpreted as the percentage of an Nvidia GPU attributable to a particular memory supplier. They nevertheless show how unusually material the economic connection between the AI accelerator leader and major memory producers has become.

This changes how US localization should be interpreted by investors. Moving SK hynix production to Ohio or Indiana could reduce location risk without materially reducing customer concentration. The same factories would still be serving an AI memory market in which large accelerator customers can account for a meaningful portion of supplier economics.

The relationship also works in the opposite direction. HBM shortages can become accelerator constraints rather than remaining isolated memory-industry problems. That mechanism is already visible elsewhere in the market. Reuters reported in September that shortages and high HBM procurement costs were pushing several Chinese AI chipmakers to raise accelerator prices.

HBM capacity therefore cannot be analyzed only as memory capex. It is part of the production constraint for the systems that monetize AI infrastructure spending.

Japanese materials and Taiwanese integration survive the move

A second layer sits farther upstream and attracts less attention because the inputs are less visible than a memory fab.

Japan has long held unusually concentrated positions in semiconductor materials. A METI industry study published in 2023 estimated that Japan represented 48% of the global market across six major semiconductor material categories, including wafers and photoresist. The same year, Japan's industry minister said Japanese companies held roughly 90% of the global photoresist market. Those figures are historical rather than current market-share estimates, but they demonstrate the structural concentration from which today's supply chain developed.

Shin-Etsu Chemical illustrates the breadth of that materials layer. Its semiconductor portfolio includes silicon wafers, photoresists, photomask blanks, high-purity deposition materials, encapsulation products and other process materials. The company specifically identifies HBM and AI semiconductors as applications supported by its materials portfolio.

Altsets data shows Shin-Etsu supplying Micron. Micron represents 1.83% of Shin-Etsu revenue, while Shin-Etsu represents 1.34% of Micron COGS. Shin-Etsu also supplies TSMC: TSMC represents 4.02% of Shin-Etsu revenue, while Shin-Etsu represents 1.33% of TSMC COGS.

Neither relationship is large enough to establish technical single-source dependence, and the data does not identify which Shin-Etsu products are present in any specific HBM generation. What it does show is that a major Japanese semiconductor materials company sits upstream of both a memory producer and the Taiwanese foundry that has become deeply involved in HBM4 and AI packaging.

The Taiwan layer is even more direct.

SK hynix is using TSMC advanced logic technology for the HBM4 base die. TSMC's CoWoS platform then provides a route for integrating HBM stacks with GPUs and other accelerators. TSMC's published manufacturing footprint currently lists its advanced backend fabs in Taiwan, while its 2025 annual report says the company will continue investing in advanced packaging capacity at multiple Taiwanese locations.

Micron provides another example of how slowly manufacturing geography changes. Although it is an American company and is building major new US capacity, Reuters reported on September 15 that its largest manufacturing operation is currently in Taiwan, where the company produces DRAM and HBM.

The United States is beginning to address the packaging gap. TSMC and Amkor signed a 10-year agreement in June to expand advanced packaging capabilities in Arizona. Amkor subsequently expanded its planned Arizona campus to roughly $12 billion and separately announced a multi-year advanced packaging partnership with Nvidia.

Those projects could materially change the map over time. For now, however, the existing network and much of the associated expertise remain centered in Taiwan. Building an HBM fab before equivalent packaging capacity is available creates domestic wafer capability without fully domesticating the finished AI module.

The equipment layer is global even when the fab is American

A US memory fab also does not imply US production equipment.

ASML is based in the Netherlands and remains the dominant supplier of advanced lithography systems. Reuters reported this week that the company held approximately 94% of the lithography market in 2025 and that SK hynix, Samsung, TSMC and Intel are among the manufacturers preparing to use its High NA EUV systems.

Micron's supplier relationships show how broad the equipment requirement can become. Altsets currently records the following relationships:

These are company-wide relationships. They do not show that a given percentage of HBM manufacturing requires one vendor's tools, and they do not establish technical substitutability. Their value is economic: they show that a large memory manufacturer's cost structure is connected to multiple specialized suppliers even before packaging and customers are considered.

This is why the semiconductor reshoring story is not simply about replacing Asian fabs with American fabs. Some critical equipment is American, some Dutch, some Japanese. Materials come from another supplier set. HBM memory companies have their own geographic concentrations. Advanced integration adds Taiwan back into the network.

What US HBM investment actually changes

The strongest case for US HBM expansion is not that it creates semiconductor autarky. It is that it can turn a supply chain concentrated in a few physical manufacturing regions into one with more alternative production locations.

A Micron fab in Idaho can produce leading-edge DRAM even if capacity elsewhere is constrained. An SK hynix manufacturing operation in the United States could give the world's largest HBM supplier another production geography. Indiana adds advanced HBM packaging. Amkor and TSMC are building an American advanced packaging ecosystem. GlobalWafers is establishing US production of 300 mm silicon wafers, another input that historically has been heavily concentrated in East Asia.

But those projects should not be counted as equivalent steps toward independence. A wafer fab, an HBM stacking operation, a base-die foundry process and a CoWoS-class accelerator package solve different bottlenecks.

For investors, this creates a practical way to interpret semiconductor capex announcements. The important question is not simply whether an AI chip or memory product can be labeled American-made. It is which production stage has moved, which dependencies remain outside the facility, and whether those remaining dependencies are duplicated elsewhere.

On that basis, the US HBM buildout is meaningful but incomplete.

Korean companies still dominate HBM supply. Japanese companies remain important providers of semiconductor materials. TSMC is becoming more deeply involved in the HBM4 base die while its CoWoS platform integrates memory with AI processors. The existing advanced packaging footprint remains heavily Taiwanese even as Arizona capacity is built. A small group of specialized equipment manufacturers supplies the fabs themselves. And at the other end of the chain, companies such as Nvidia represent economically significant customers for the memory producers.

The likely outcome is not a fully American HBM supply chain. It is a geographically broader allied supply chain with more manufacturing in the United States.

That distinction is the investment finding. New US fabs can reduce the risk attached to a particular production location without removing the economic dependencies that connect AI accelerators to Korea, Japan and Taiwan. An investor treating every US semiconductor factory as equivalent reshoring will miss where the remaining bottlenecks actually sit.

Methodology and limitations

Sources

• Reuters, September 16, 2026, reporting on exploratory SK hynix and Intel talks concerning US memory manufacturing.
• SK hynix, August 28, 2026, Indiana HBM advanced packaging facility announcement.
• Micron, US manufacturing expansion information for Idaho and New York.
• Counterpoint Research HBM market-share data for Q2 2026, reported by Yonhap on September 3, 2026.
• Nvidia fiscal 2026 Form 10-K, including foundry, memory, CoWoS and geographic supply-chain disclosures.
• SK hynix and TSMC materials describing HBM4 base-die collaboration and CoWoS integration.
• Shin-Etsu Chemical semiconductor materials disclosures and Japan METI semiconductor-material market data.
• TSMC manufacturing and annual-report disclosures concerning advanced packaging in Taiwan.
• Amkor announcements covering its TSMC partnership, Arizona packaging expansion and Nvidia agreement.
• Reuters, September 14, 2026, on ASML's lithography position and High NA adoption by major semiconductor manufacturers.

Unique analytical contribution

The article separates HBM localization into fabrication, memory stacking, materials, base-die logic, advanced accelerator packaging, manufacturing equipment and downstream demand, showing that US fab construction reduces geographic production concentration without removing the economic and technological dependencies linking AI hardware to Korea, Japan and Taiwan.

How to Cite This

For research inquiries or data access: press@altsets.com

Go Deeper