Who ASE is
ASE Technology Holding (3711.TW), NYSE:ASX) is the world's largest pure-play OSAT, with an electronics-manufacturing-services (EMS) leg beside the ATM (assembly, test and materials) book. It is not TSMC, and it does not own the CoWoS name. Jason Chang's July 2026 line is the one to keep: the most advanced wafer-level packaging still sits with the foundry; ASE takes the packaging, test and full-process work around that, and can buy a substrate and assemble EMIB or another scheme if the customer asks.
It is not on NVIDIA's published 800 VDC list. It belongs here because AI advanced packaging is now the growth engine of the listed holding company, and because the call is explicit about where ASE stops and the foundry starts.
Business mix
Two reported engines, plus materials inside ATM:
- ATM (packaging, test, materials) — NT$126.15bn in Q2 2026, 66% of group sales, +12% QoQ and +36% YoY. Gross margin 27.3%. Application mix: communications 41%, computing 30%, automotive / consumer / other 29%. Computing is the AI climb.
- EMS — the rest of the group. Q3 revenue guided +40% sequentially, of which a slice is memory-price inflation rather than unit growth. Operating margin guided 3.2–3.4% (about 3.7–3.8% ex that inflation).
- LEAP — ASE's advanced-packaging and advanced-test service line, not a GAAP segment. 2026 revenue guided above the prior US$3.5bn target, with "several hundred million" more dollars possible; 2027 is a double versus 2026.
Where it sits in the NVIDIA 800 VDC stack
ASE does not make the 800 V rail. It packages and tests the AI silicon that the rail will feed. The July 2026 call never named 800 V or HVDC.
On CoWoS specifically, management said ASE will keep raising CoWoS-related capacity, scale and efficiency, and will put EMIB or another backend on the roadmap if customers choose it. That is OSAT language: support the package the customer brings, including TSMC CoWoS overflow and ASE's own fan-out / panel-level families. It is not a claim that ASE operates TSMC's CoWoS line.
Dated alternatives on the same call:
- Full-process (turnkey) about US$300m in 2026, still too small to move group margin, guided to grow "significantly" in 2027. CPU, GPU and ASIC are in the mix; customer names in one to two quarters.
- Panel-level 310×310 mm fully automated line, production 2027 Q1. Described as complementary to the foundry's approach, same class of customer, similar reticle and line/space — the winner is capacity, performance and time-to-volume.
- Glass substrate — in evaluation with substrate vendors, the foundry and customers. No production in the next 12 months.
- CPO — small volume from end-2026; revenue and 2027 outlook in about two more quarters. Twenty years of work in-house; no hurry to size it.
Capex is the tell: US$10.5bn for 2026, the third raise of the year, mostly LEAP, with 13 new fab projects and eight acquisitions / retrofits sized through 2028–29. Free cash flow stays negative while that lands.
Coverage note
One note, Q2 2026 (30 July 2026). Dual-tagged on the CoWoS hub. US engineering sites in Fremont and San Jose are not a second memo.