South Korean High Bandwidth Memory (HBM) exports to Malaysia have reached roughly $1.3 billion while shipments to Taiwan fell below $3 billion, down from over $5 billion months earlier. The shift coincides with Intel's newly completed Pelican advanced-packaging plant in Malaysia and reports that TSMC (台積電) has begun routing overflow CoWoS orders to Intel's Malaysian facility.
What do the trade numbers show about the HBM shift between Malaysia and Taiwan?
South Korean High Bandwidth Memory (HBM) exports to Malaysia have reached about $1.3 billion while exports to Taiwan have fallen below $3 billionCITE:E1. SemiAnalysis Chipbook export data shows Malaysia-bound HBM shipments now valued at roughly $1.3 billionCITE:E1, while Taiwan-bound HBM shipments dropped to under $3 billion, down from more than $5 billion just months earlierCITE:E2.
| Metric | Value |
|---|
| HBM exports to Malaysia | ~$1.3 billion |
| HBM exports to Taiwan (current) | Below $3 billion |
| HBM exports to Taiwan (prior, months earlier) | Above $5 billion |
What is Intel's Project Pelican in Malaysia, and what can it do?
Intel (英特爾) completed its Project Pelican advanced-packaging plant in Malaysia at the end of 2025, with total investment reaching $7 billionCITE:E3. The Pelican facility includes die sort and die prep functions and supports Intel's EMIB and Foveros advanced-packaging process flowsCITE:E4.
How is Intel expanding advanced-packaging capacity beyond Malaysia?
Intel has started EMIB assembly operations at Amkor's plant in Incheon, South Korea, to absorb orders from large AI customersCITE:E5. With the Pelican plant online, Intel's packaging capacity — previously confined to its New Mexico site in the United States — has doubledCITE:E6.
How tight is TSMC's CoWoS capacity, and is Intel picking up the overflow?
TSMC's (台積電) CoWoS advanced-packaging capacity is fully booked, and some back-end orders have reportedly overflowed to Intel's Malaysian plant, which is providing partial capacity support to serve shared customers — a break from past industry-ecosystem normsCITE:E7.
How are TSMC and ASE Technology responding to Intel's advanced-packaging push?
TSMC Chairman C.C. Wei (魏哲家) said TSMC focuses on front-end advanced packaging and welcomes more vendors supplying back-end capacity to provide flexibility where TSMC is shortCITE:E8. Wei said that based on media reports, the outlook for the relevant technology looks promising, and that he hopes the other party succeeds in sharing part of TSMC's capacity pressure, since new alternatives give customers greater flexibilityCITE:E8. ASE Technology (日月光投控) Chief Operating Officer Tien Wu (吳田玉) said CoWoS, EMIB and other packaging options are not mutually exclusive substitutes for one anotherCITE:E10. Wu said that when customers choose Intel's EMIB-T technology, ASE Technology can still procure substrates and provide back-end assembly and final test services for those customers, without being confined to a single technology pathCITE:E10.
What progress has EMIB-T made, and who else is working on it?
Unimicron (欣興) said it has secured customer commitments and is working with two Japanese firms to mature the technology, targeting 2027 mass production of EMIB-T-related applicationsCITE:E9. Intel CEO Lip-Bu Tan (陳立武) said Intel's next-generation EMIB-T packaging technology is on track to secure reliable yields, and that because substrate supply is very tight, several customers have already prepaid substrate feesCITE:E11. Tan said that with CoWoS capacity widely known to be short, Intel is in a unique position to offer support, and called the customer prepayments encouragingCITE:E11.
What this means
The trade data — $1.3 billion in HBM now flowing to Malaysia versus Taiwan's drop from over $5 billion to under $3 billionCITE:E1CITE:E2 — lines up with the timeline of Intel's $7 billion Pelican plant completing at the end of 2025CITE:E3 and reports of CoWoS orders overflowing to that same Malaysian facilityCITE:E7. TSMC's own chairman frames the overflow as welcome relief rather than a threatCITE:E8, while ASE Technology positions itself as agnostic across CoWoS and EMIB-TCITE:E10. Intel's claim of customers prepaying substrate fees for EMIB-TCITE:E11 and Unimicron's 2027 mass-production target with two Japanese partnersCITE:E9 point to the same underlying condition described across every account: back-end advanced-packaging supply, not chip design, is the current bottleneck.