CoWoS (Chip-on-Wafer-on-Substrate) is TSMC’s advanced packaging service that mounts several chips, typically an AI processor and stacks of high-bandwidth memory (HBM), side by side on a silicon interposer that wires them together, then attaches the assembly to a substrate. TSMC lists CoWoS among its 3DFabric services for joining different chips into one high-performance package. It matters because an AI accelerator is not finished until it is packaged with its memory, so packaging capacity, not only wafer capacity, sets how many AI chips reach the market.
Instead of making one huge die, CoWoS joins separate dies in one package. Research on 2.5D graphics modules describes the fine-pitch, high-density wiring a silicon interposer provides between a GPU and four HBM stacks (Cho et al., 2016).
02
Memory bandwidth is the point
AI training and inference move enormous amounts of data between processor and memory. Placing HBM next to the processor on the interposer shortens those connections; SK hynix describes HBM as a critical component of the GPUs that train AI systems.
03
One supplier, two bills
TSMC makes the logic wafer and can also package it. Wafer fabrication was about 86% of TSMC's net revenue in 2025; the rest came mainly from packaging and testing, mask making, design and royalties, according to its annual report.
04
Part of a packaging family
TSMC groups CoWoS with its other 3DFabric technologies, such as InFO fan-out packaging and SoIC silicon stacking, in its annual report.
What changed in 2025-26
AI became TSMC's biggest market
High performance computing, which includes AI accelerators, rose from 43% of TSMC’s net revenue in 2023 to 58% in 2025. In the third quarter of 2026 TSMC’s revenue was $46.72 billion, up 41.2% year on year, and it guided to $52.3-54.3 billion for the fourth quarter.
Packaging gets its own capex
TSMC’s annual report says its 2025 capital expenditures included expanding capacity for specialty technologies and advanced packaging, including building and facility expansion for Fab 23 and Fab 24.
Packaging follows the fabs to Arizona
In October 2024 Amkor and TSMC signed a memorandum of understanding under which TSMC will contract turnkey advanced packaging and test services from Amkor’s planned facility in Peoria, Arizona; the companies named InFO and CoWoS as technologies to be used.
Memory makers build packaging too
SK hynix said in April 2024 it would invest an estimated $3.87 billion in West Lafayette, Indiana, in an advanced packaging plant and R&D facility for HBM, with mass production planned for the second half of 2028.
CoWoS and related packaging pieces
CoWoS
What it doesJoins logic and HBM side by side on an interposer
Who (source)TSMC (2025 annual report)
InFO
What it doesFan-out packaging, offered to smartphone customers; TSMC calls it industry-leading
Who (source)TSMC (2025 annual report)
SoIC
What it doesAdvanced silicon stacking of chips
Who (source)TSMC (2025 annual report)
HBM
What it doesStacked DRAM placed beside the processor
Who (source)SK hynix, Indiana plant from 2028
Outsourced packaging
What it doesTurnkey packaging and test for TSMC customers in Arizona
Who (source)Amkor (2024 agreement)
What it means for you
Investors: track packaging capacity as closely as wafer capacity; an AI chip that cannot be packaged cannot ship.
Buyers of AI compute: ask suppliers where their accelerators are packaged and how much packaging is contracted, not only which node they use.
Chip designers: plan for a multi-die package early; interposer size and memory stack count shape cost and delivery.
Policy watchers: count packaging sites, such as Amkor in Arizona and SK hynix in Indiana, when judging how much of the AI supply chain has moved.
Read TSMC's platform mix each quarter; the share from high performance computing is the cleanest proxy for AI demand at the foundry.
The Business Engineer lens
CoWoS shows that the AI bottleneck moves: first GPUs, then the step that joins processors to memory. The Business Engineer reads packaging as a toll point where TSMC earns a second time on the same chip, and where new capacity in the US and from memory makers is now being built.
Chip-on-Wafer-on-Substrate: chips are placed on a silicon interposer wafer, which is then mounted on a package substrate.
Why is CoWoS important for AI chips?
It joins the AI processor with stacks of high-bandwidth memory in one package; without that step the accelerator cannot be finished.
Who makes CoWoS?
TSMC offers CoWoS as one of its 3DFabric advanced packaging services; under a 2024 agreement Amkor will provide turnkey packaging and test for TSMC customers in Arizona.
How much of TSMC's revenue comes from AI-related chips?
High performance computing, which includes AI accelerators, was 58% of TSMC's net revenue in 2025, up from 43% in 2023.
The complete explainer, cleaned up and reformatted for mobile.
Even when logic and memory exist, GPUs cannot be completed without advanced packaging.
Why CoWoS Is Critical
Enables GPU + HBM integration on single package
Without CoWoS, chips are built but cannot be finished
Supply Chain Response
SK hynix US plant (Indiana): an estimated $3.87 billion advanced packaging plant and R&D facility for HBM, announced in April 2024, with mass production planned for the second half of 2028
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
