Biwin bets ¥4.5B on Dongguan wafer-level packaging — module house climbs the stairs

What happened

On September 23, 2026, the South China Morning Post (Howard Liu, Beijing) reported that Biwin Storage Technology will invest 4.5 billion yuan, about US$672 million, in the third phase of its advanced packaging and testing project in Dongguan, Guangdong. That sum is larger than the 3.09 billion yuan Biwin put into phases one and two combined. The story rests on Biwin’s Tuesday stock filing and on remarks from earlier investor meetings.

The Post is plain about where the project stands. Including the first two phases, it had yet to generate mass-production revenue. Work was still in customer sampling, process verification, and small-batch production. The company estimated that monthly capacity would more than triple, to 10,000 wafers, by the end of 2027. That capacity figure is a company estimate, not an audited run rate.

Biwin already runs a traditional memory and system-in-package base in Huizhou, also in Guangdong, and buys chips from outside suppliers. The Post frames Dongguan as a move into more sophisticated wafer-level work. In the same piece, Biwin is pushing fan-out packaging for stacked memory, a separate assembly that integrates 10 chiplets so compute and memory sit together, and customer work on near-packaged optics. Those are research and customer-engagement notes. They are not shipment proof.

Other module houses are moving at different speeds. Shenzhen Techwinsemi has brought some wafer-die and packaged-chip testing in-house and still sends most packaging outside. Longsys said in June that its Suzhou plant could produce more than 1 million miniature SSDs a month, using system-in-package assembly to put the controller, flash, power management, and other parts in one module. Sinolink Securities, in an August note cited by the Post, argues that module makers now earn more of their keep from product definition, supply-chain management, and customer development, and that packaging is becoming a performance and reliability factor for AI devices rather than only a last assembly step. Yole Group, via SEMI and cited by the Post, forecasts global advanced-packaging revenue of US$46.1 billion in 2024 growing to US$79.4 billion in 2030.

Biwin’s Chinese name is 佰维 (Bǎi-wéi), literally “a hundred dimensions,” which fits a module house trying to think in wafers instead of sticks.

Why it matters

AI hardware still ends in packaging. Taiwan’s CoWoS queue is the famous bottleneck. Malaysia is taking in HBM-related shipments from Korea. China’s module houses are trying to move from board-level assembly into wafer-level work so they keep more of the AI memory and edge-device bill. Biwin’s phase-3 check is large next to its own earlier phases. Revenue from the Dongguan lines has not started.

This is a storage company spending on packaging. Korea’s HBM-related exports to Malaysia, which the customs lane put up about 5.7-fold, are the other geography on the same bottleneck. Listed packaging houses such as JCET and Tongfu are the profit-side version of that theme, because AI packaging demand is already showing up in their results. Biwin is not taking CoWoS work from TSMC. More than 10,000 wafers a month at the end of 2027 is not capacity running today.

What to watch is the first disclosure of mass-production revenue from the Dongguan phases; whether the end-2027 guide above 10,000 wafers a month survives into later filings; whether fan-out or multi-chiplet work shows up under customer names in a teardown; and whether Longsys and Techwinsemi match the spending or stay with test and system-in-package work.

The Taiwan queue this project does not replace is TSMC’s CoWoS allocation.