Can SK hynix Really Bring Memory Manufacturing to America?
September 16, 2026
Altsets
Research by Altsets Research
What proposed U.S. memory fabrication and planned HBM packaging could localize—and which Korean, Japanese, and Taiwanese dependencies would remain.
Data used:Altsets Supply Chain Intelligence: 90k+ entities, 400k+ relationships, 20+ years of history.
Key findings
- A potential Intel manufacturing agreement could add front-end DRAM fabrication to SK hynix's planned U.S. HBM packaging footprint.
- SK hynix's much larger Korean investment program indicates that U.S. capacity would diversify production rather than replace Korea as its manufacturing center.
- HBM4 logic, advanced packaging, specialized materials, and product qualification would preserve important dependencies on Taiwan and Japan.
For journalists and researchers: this casefile separates what a potential Intel manufacturing agreement could localize from the HBM dependencies that would remain abroad.
Yes, SK hynix can bring real memory manufacturing to America. It cannot, at least with the projects currently disclosed, bring the entire high bandwidth memory supply chain with it.
That distinction matters after Reuters reported on September 16, 2026 that SK hynix is in exploratory discussions with Intel about manufacturing memory chips in the United States. One scenario would involve leasing part of Intel's Ohio semiconductor complex. Another could involve a joint venture among SK hynix, Intel, and cloud computing companies seeking more secure memory supplies. SK hynix subsequently said it is considering ways to strengthen its global manufacturing footprint but that no specific cooperation, production plan, or agreement has been finalized.
The talks would represent a much larger step toward localization than SK hynix's existing American project. Its Indiana facility, which broke ground in August 2026, is an advanced packaging plant. The company says it will invest more than $4 billion there and expects mass production of next generation HBM in the second half of 2029. Packaging is a critical part of HBM manufacturing, but the DRAM dies inside the stack still have to be fabricated somewhere.
An Intel agreement could therefore fill a missing part of the U.S. chain by adding front end memory wafer production. Even then, an American SK hynix HBM stack would initially be better understood as a multinational product assembled across a more diversified geography than as a fully domestic semiconductor.
Takeaway: The strategic value of the Intel talks is not that they suddenly eliminate Asian semiconductor exposure. It is that they could remove one more geographic concentration from a product that has become unusually important to AI infrastructure.
At a glance
Relationship percentages are Altsets company level relationship estimates. Supplier Revenue % and Customer Cost % are different metrics and should not be compared as if they measure the same quantity.
What an Intel deal would actually change
HBM is not one manufacturing step. The finished product begins with advanced DRAM wafers, requires the dies to be connected vertically through thousands of through silicon vias, relies on increasingly complicated packaging and thermal materials, and in the newest generation incorporates a logic base die beneath the memory stack. That HBM package must then be integrated alongside a GPU or other accelerator through another layer of advanced packaging.
Where a U.S. SK hynix HBM chain could form
The announced and discussed projects address different manufacturing stages.
- 01Front end DRAM fabricationPotential Intel Ohio role if a manufacturing agreement is reached
- 02HBM stacking and advanced packagingSK hynix Indiana project, with mass production targeted for H2 2029
- 03HBM4 base die and advanced integrationTSMC collaboration remains important, with Arizona capacity still being built
- 04AI accelerator packageHBM is integrated alongside GPUs or other accelerators
A more complete U.S. chain is possible, but facility construction alone does not prove that a specific HBM generation or accelerator will use each U.S. site.
Caption: The Intel discussions matter because they could add front end DRAM fabrication to a U.S. chain that already includes planned SK hynix HBM packaging. The remaining logic, packaging, materials, and qualification dependencies make the chain multinational.
Source: Altsets
SK hynix's Indiana project addresses part of that chain. Its planned production line is focused on the advanced packaging stage, where multiple DRAM dies are stacked and turned into an HBM product. The company's proprietary MR-MUF process, for example, fills spaces between stacked memory dies with a protective molding material while addressing heat dissipation and warpage. SK hynix describes advanced packaging as central to maintaining HBM performance as stacks become denser.
Front end DRAM manufacturing is different. It requires an advanced wafer fab, extremely expensive deposition, etch, lithography, cleaning, inspection, and metrology equipment, along with a production process tuned specifically for memory.
This is why the Intel talks are more consequential than another packaging announcement. If SK hynix eventually uses Intel's Ohio campus for DRAM fabrication, the United States could host both HBM memory fabrication and SK hynix's downstream HBM packaging.
There is an important timing problem, however. Intel's most recent public Ohio schedule calls for its first fab there to be completed in 2030 and to begin operations in 2030 or 2031. Intel deliberately slowed construction as it sought to align spending with actual customer demand. Any SK hynix arrangement could conceivably alter that plan, but neither company has announced that it will.
The current story is therefore about strategic optionality, not near term capacity. For Intel, an anchor memory customer could create another route to monetizing a huge manufacturing site whose original timeline has already slipped. For SK hynix, Intel could provide a way to add American wafer capacity without independently constructing an entire greenfield memory fab.
The economic incentives are becoming stronger because SK hynix's largest AI customers are also concentrated in the United States. Altsets maps SK hynix as a supplier to Nvidia. In the current relationship estimate, Nvidia represents 27.88% of SK hynix revenue, while SK hynix represents 27.33% of Nvidia COGS.
Those figures should not be interpreted as HBM shares. They are company level relationship estimates rather than product level measurements. They do show why locating more production near major American AI customers can matter economically. A relationship of that scale gives both sides a reason to care about capacity availability, logistics, policy risk, and geographic resilience.
South Korea would remain the center of SK hynix memory production
The strongest evidence against interpreting the Intel talks as a wholesale relocation is SK hynix's own capital spending.
The company is simultaneously expanding an enormous Korean memory manufacturing network. Its M15X facility in Cheongju is designed for next generation DRAM including HBM, and its neighboring M15 facility has been expanding TSV capacity. SK hynix has explicitly described the combination as a way to maximize HBM production capability.
The expansion does not stop there. In August 2026, SK hynix approved approximately 54 trillion won of additional investment in new fabs at Yongin and Cheongju. The planned Yongin Y2 fab alone represents 35.2 trillion won of investment and is intended to manufacture HBM and other next generation DRAM products. It is part of a much larger Korean expansion program that SK hynix says could eventually involve 600 trillion won in the Yongin semiconductor cluster.
Packaging is also expanding domestically. SK hynix is building P&T7 in Cheongju as a dedicated advanced packaging facility for AI memory including HBM. That facility sits within a Korean manufacturing system that already connects Icheon, Cheongju, and future Yongin capacity.
What the spending implies: American capacity would diversify production rather than replace Korea.
For investors, this distinction changes how geopolitical risk should be modeled. A U.S. fab could reduce the number of HBM units that must begin their manufacturing journey in South Korea, particularly if SK hynix eventually duplicates important DRAM processes in Ohio. It would not make Korean production irrelevant. Korea is still receiving much larger commitments to both front end DRAM capacity and HBM packaging.
The Korean supply chain also has decades of accumulated process knowledge, supplier relationships, technical labor, equipment support, and high volume manufacturing experience. Duplicating a fab building does not instantly duplicate that ecosystem.
HBM4 introduces a separate Taiwan dependency
The geography becomes more complicated with HBM4 because the next generation of memory is increasingly merging memory manufacturing with advanced logic.
SK hynix and TSMC agreed in 2024 to collaborate on HBM4. Previous SK hynix HBM generations used the memory maker's own process for the base die at the bottom of the HBM stack. Beginning with HBM4, SK hynix is adopting TSMC's advanced logic process for that base die. The companies are also working together to optimize SK hynix HBM for TSMC's CoWoS technology, which integrates HBM and processors such as GPUs on a common package.
This matters because producing the DRAM dies in America would not automatically localize the HBM4 base die or the final accelerator package.
TSMC itself is gradually changing that geography. The company already manufactures advanced logic wafers in Arizona, and its U.S. expansion now includes planned advanced packaging capacity. TSMC said in 2026 that it expects an Arizona packaging facility supporting technologies including CoWoS to open around 2029. Until equivalent U.S. capacity is qualified for the actual products SK hynix and its customers need, however, Taiwan remains deeply embedded in the HBM4 architecture.
This creates an unusual transition period. By the end of the decade, the United States could theoretically contain SK hynix HBM packaging in Indiana, TSMC advanced logic and packaging in Arizona, and potentially SK hynix memory wafer production through Intel in Ohio. That would be a much more complete AI semiconductor manufacturing chain than exists today.
But those projects are being constructed on separate schedules by separate companies, and advanced semiconductor supply chains are qualification specific. A facility existing in the United States does not prove that a particular Nvidia accelerator, HBM4 base die, or SK hynix memory generation will immediately use it.
Taiwan risk therefore declines only when production is actually transferred and qualified, not when construction spending is announced.
Japan remains a materials dependency
Japan represents a different kind of HBM exposure. Its role is less visible than Korea's fabs or TSMC's packaging, but Japanese companies still control specialized materials used inside advanced memory manufacturing.
JX Advanced Metals provides a particularly direct example. The Japanese company says its semiconductor sputtering targets are used as wiring materials in HBM and support production of SK hynix's latest generation products. SK hynix gave JX Advanced Metals a Best Partner Award in 2025 for its supply support.
The geographic chain behind those targets is itself multinational, but important upstream work remains in Japan. JX says the upstream manufacturing processes for its semiconductor sputtering targets are carried out at facilities in Ibaraki Prefecture, while downstream machining also occurs at sites in South Korea, Taiwan, and the United States. In March 2026, the company opened a new Hitachinaka factory in Japan and began trial operation of additional semiconductor sputtering target capacity.
Mitsui Chemicals provides another example closer to HBM process engineering. Its ICT materials subsidiary received SK hynix's 2025 AI First Mover Award after its ICROS protective tape helped address technical challenges in the HBM manufacturing process. Mitsui says the tape is used across semiconductor wafer processing, dicing, and packaging, and its Nagoya Works in Japan manufactures ICROS Tape.
These are not interchangeable commodity inputs. Semiconductor materials are engineered around process specifications and qualification requirements. Japan's broader semiconductor materials industry retains strong positions in products including silicon wafers, photoresists, cleaning chemicals, and advanced packaging materials.
The investment implication is not that every Japanese semiconductor supplier becomes an HBM bottleneck. It is that relocating a memory fab does not relocate the qualified materials ecosystem underneath it. Some suppliers can add American or Korean processing capacity while maintaining upstream production and intellectual property elsewhere.
That makes materials localization a slower and less visible stage of semiconductor onshoring than building a fab.
What localizes, and what remains geographically distributed
The relevant dependency is different in each geography.
| Geography | Role in the HBM chain | What is changing | What remains |
|---|---|---|---|
| United States | Planned HBM packaging, possible DRAM fabrication, expanding advanced logic and packaging | Indiana, Ohio, and Arizona could cover more stages by the end of the decade | Product qualification and actual production transfer are still required |
| South Korea | Core SK hynix DRAM fabrication, TSV capacity, and advanced HBM packaging | Large new fab and packaging investments continue | The center of SK hynix memory production remains in Korea |
| Taiwan | HBM4 base die and CoWoS ecosystem | TSMC is beginning to duplicate some advanced capabilities in Arizona | HBM4 and final accelerator integration remain tied to qualified Taiwan capacity today |
| Japan | Specialized semiconductor materials used in advanced memory manufacturing | Some suppliers operate downstream processing outside Japan | Important upstream production and process know-how remain in Japan |
Caption: U.S. investment can reduce geographic concentration without eliminating the established Asian manufacturing and materials ecosystems underneath HBM.
Source: Altsets
The investment case is about redundancy, not independence
The reported SK hynix and Intel talks are meaningful precisely because the United States currently lacks a complete SK hynix HBM production chain. Indiana addresses advanced packaging. A manufacturing agreement with Intel could eventually address front end memory fabrication. TSMC and other packaging investments in Arizona are beginning to address the logic and system integration side.
Taken together, these projects could materially reduce the number of critical AI memory manufacturing stages concentrated in East Asia.
They do not create immediate supply chain independence.
South Korea would remain SK hynix's principal capacity expansion zone and a major source of HBM DRAM and packaging. Taiwan remains tied to the HBM4 base die and CoWoS ecosystem, although TSMC is beginning to duplicate those capabilities in Arizona. Japanese companies continue to supply specialized materials used in HBM manufacturing, with important upstream production still located in Japan.
This makes the Intel discussions more interesting for Intel than a simple headline about bringing memory production back to America suggests. An agreement could give Intel an external customer for manufacturing infrastructure, add another component to the emerging U.S. AI hardware ecosystem, and give SK hynix a faster route toward geographic diversification.
For SK hynix, however, the strategy appears to be duplication rather than migration. Its Korean investment commitments are too large to support the idea that the center of gravity is moving to America.
Research conclusion: The most realistic end state is a parallel supply chain. Some HBM could eventually be fabricated, stacked, and integrated in the United States while larger volumes continue to run through established Asian manufacturing hubs.
That is still strategically important. Supply chain resilience does not require every component to come from one country. It requires enough independent production paths that the failure of one geography does not stop the entire system.
The Intel talks would move SK hynix closer to that objective. They would not finish the job.
Methodology and limitations
Altsets relationship data used in this analysis describes economic relationships between companies. Supplier Revenue % measures how important the customer is to the supplier's revenue. Customer Cost % measures how important the supplier is to the customer's cost of goods sold.
The SK hynix and Nvidia figures cited above are relationship level estimates and should not be interpreted as the share of Nvidia products containing SK hynix memory, the share of SK hynix revenue specifically attributable to HBM, or a measure of technical substitutability. A supplier relationship does not establish that another supplier can replace it.
The displayed relationship is also not intended to represent either company's complete supply chain. Public company disclosures, company announcements, government information, and supplier disclosures were used separately to identify manufacturing stages and geographic dependencies.
How to Cite This
According to Altsets Supply Chain Intelligence (altsets.com), SK hynix can move meaningful parts of HBM production to the United States, but critical dependencies would remain in South Korea, Japan, and Taiwan.
For research inquiries or data access: press@altsets.com
Go Deeper
Inspect the company relationships and supply chain paths that sit behind SK hynix's AI memory exposure.
Follow a manufacturing or supplier disruption through upstream and downstream company relationships.
Review Supplier Revenue %, Customer Cost %, Relationship Size, and the limits of company level relationship estimates.