Taiwan’s National Science and Technology Council has approved a plan by DRAM maker Nanya Technology to build an advanced 3D wafer packaging and testing plant in Pingtung Science Park, in the island’s south. The facility will handle the back-end work that turns standard DRAM wafers into stacked, high-bandwidth memory modules, the kind of packaging that has become a bottleneck in the AI hardware buildout.
The Southern Taiwan Science Park Administration announced the approval this week. According to the statement, Nanya will focus on through-silicon via technology, known as TSV, and wafer stacking, which let DRAM dies be connected vertically and packed into compact 3D structures. The goal is more capacity and bandwidth per module with lower latency, which is what memory suppliers must deliver to keep pace with AI accelerators that consume ever more data per second.
Packaging Is the New Battleground
Memory makers historically sold commodity DRAM and left advanced packaging mostly to others. That division of labor has collapsed as AI changed the economics. High-bandwidth memory, the stacked DRAM that sits next to GPU dies in accelerator assemblies, is now one of the most profitable and most supply-constrained segments in the industry. Advanced packaging lines are booked out, and every major memory producer is racing to add back-end capacity where it can, because a fast front-end fab is useless in this market if the output cannot be stacked and assembled.
Nanya is a mid-sized player in this market, smaller than Samsung, SK hynix and Micron, and it has been repositioning for exactly this shift. In July the company said it already offers wafer-to-wafer bonding for customized memory chips and is developing memory module solutions with logic attached to the silicon, which extends a DRAM manufacturer’s usual job into territory normally held by packaging specialists and module houses. The Pingtung approval turns that stated direction into a concrete site, with the local science park authority confirming Nanya is aiming for vertical integration of DRAM into 3D structures to lift capacity and bandwidth while cutting latency.
How the Investment Fits Nanya’s Plans
The new plant is the back-end part of a wider buildout that Nanya laid out earlier this year. In August the company’s board approved capital spending of up to NT$346.6 billion, roughly US$10.7 billion, for its Fab 5A front-end fab in Linkou through 2029, including extreme ultraviolet lithography for 10nm-class process nodes. Fab 5A is planned at a full capacity of about 45,000 wafers per month, with wafer input starting in the second half of 2027 and total investment estimated near US$16 billion when fully deployed. First-phase capacity is set at 35,900 wafer starts per month.
Nanya also raised its 2026 capital expenditure budget this year from NT$52 billion to as much as NT$69.7 billion, an increase of about 34 percent, partly to prepay for equipment headed into Fab 5A. Taiwanese media reports from August said the company had been eyeing sites in both Yunlin and Pingtung, with total investment in new fabs expected to exceed the NT$300 billion committed to Fab 5A. The NSTC approval of the Pingtung packaging and testing project confirms at least one of those sites is moving from exploration into execution.
The customer list suggests the capacity will find buyers. In March, SanDisk, Cisco, Kioxia and SK hynix’s Solidigm unit together invested NT$78.72 billion in Nanya through a private placement, an unusual cross-company arrangement that secures future DRAM capacity for buyers who need it. Those companies compete with Nanya in some product lines but still bought in, which says something about how tight memory supply has become and how much the customers value guaranteed allocation over cash today.
Order signals back the direction too. Nanya’s August 2026 revenue came in at NT$44.69 billion, and the company has scheduled its third-quarter earnings conference for October 12, where investors will expect detail on how the Pingtung project slots into the capital plan and when the first line should start producing tested modules.
Context: The Memory Squeeze
The approval lands while memory markets run hot. Micron this week reported record revenue for its fiscal fourth quarter, beating estimates, and projected another record quarter ahead, pointing to data center demand that keeps conventional memory and high-bandwidth memory sold out. Consumers are already seeing the effect at retail, with phone and PC makers passing through higher memory costs. Industry watchers have warned the shortage stretches into 2027, and every credible announcement of new capacity draws scrutiny from buyers burned by two years of price increases.
Whether the Pingtung plant matters for near-term supply is a different question. Advanced packaging facilities take years to build out, and Nanya’s own Fab 5A front-end output only starts arriving in volume from late 2027. The plant is a bet on the shape of demand several years out, one that assumes AI memory consumption keeps growing and that customized, vertically integrated memory products become the norm rather than the exception. Nanya has picked a lane where a mid-sized player with strong engineering can realistically compete, and the back-end investment is the part of the stack where the fight is currently being decided. The question for its customers is whether the added stack capacity arrives before the next shortage cycle peaks, and for its rivals, whether this comes at the cost of front-end funding they had assumed.