Report

Foundry Monthly Intelligence Report, September 2026

If you are interested in subscribing to the full report or would like to learn more, please contact us at [email protected].

Overview

This Foundry Monthly Intelligence Report provides timely, always-on market intelligence to keep you aligned with the latest developments in the foundry ecosystem. It includes ad hoc deep dives into emerging trends and market shifts, along with concise analyst takeaways that deliver real-time, decision-ready insights to support strategic planning and execution.

The Shift in Agentic AI Computing Infrastructure and Its Impact on Foundry and Advanced Packaging. Three Takeaways from OCP 2026 About the Shift in Agentic AI Computing Infrastructure and Its Impact on Foundry and Advanced Packaging

Agentic AI has increased the demand for CPU computing power for workflow management. Meanwhile, longer task cycles and larger context windows are driving the industry toward multi-tier memory and storage architectures that integrate HBM, CXL-based memory, and SSD storage.
As power consumption per AI rack continues to climb, reaching 140 kW, 240 kW, and even over 350 kW, high-voltage power delivery and liquid cooling solutions are becoming increasingly critical for managing power loss and deployment costs.
Advanced process nodes, chiplet technology, and 2.5D/3D packaging are helping AI chip manufacturers address challenges related to reticle size limits, production yields, and power consumption. Furthermore, the wider interface of HBM4 places more stringent requirements on the manufacturing and integration of base dies.

A14 Is Running Earlier Than Expected and Packaging Builds Out on Three Tracks

A14 yield learning is running ahead of our original expectations. We expect risk production in 2027, a ramp in H1 2028 and meaningful volume in H2 2028, roughly one to two quarters earlier than previously expected.
The initial capacity ramp for A14 will begin in Hsinchu before extending to Taichung. The Arizona facility is unlikely to contribute any capacity before 2030.
CoWoS remains the top priority for TSMC’s packaging business, with current capacity already fully utilized. Capacity expansion for SoIC is proceeding at a more measured pace, while CoPoS is unlikely to reach a meaningful scale of production before the second half of 2028.

Demand-driven capacity expansion may continue until 2030 for main Chinese foundries

Strong local demand is underpinning the expansion of SMIC in both mature and advanced process nodes on 12’’ wafers. We project that its capacity for advanced process nodes will reach 60–70 KWPM by 2030.
HHG is focusing on 12” mature-node capacity through its HLMC Fab 5 acquisition and the future ramp of Wuxi Fab 9B. Its total capacity could reach around 350 KWPM once Fab 9B is fully loaded.

TSMC Keeps Adding Capacity for N3: Fab Roadmap Update and What Customers Are Actually Asking For

Years after the N3 process node entered mass production, TSMC continues to expand its capacity, reflecting sustained demand from AI GPUs, AI ASICs, and base dies for HBM.
AMD has requested additional capacity from TSMC, amounting to approximately 25,000 wafers, for the production of its MI450 products in 2027. However, TSMC has yet to make a firm commitment to meet this incremental demand. Final shipment volumes will also depend on the availability of supporting CoWoS packaging and substrate capacity.
Apple has reduced its N2 and N3 wafer orders by 10–15% amid weaker iPhone Duo build plans and higher memory costs. We believe AI demand can absorb this reduction, although a longer consumer slowdown remains a risk.

Kategorie

Branche:

Semiconductors

Dienstleistung:

Foundry

Berichtstyp:

Report

Zeitraum:

Monthly

Branche:

Semiconductors

Berichtstyp:

Report

Dienstleistung:

Foundry

Zeitraum:

Monthly

Autor

William Li

William is a Research Analyst in Semiconductor and Components team, based in Taiwan. He has over 8 years of experience in global PC (personal computer) and semiconductor market. Before Counterpoint, he was an equity research analyst at Credit Suisse, focused on both technology and non-technology sectors as well as supporting Taiwan equity market strategy research. Prior, he worked for a Taiwan PE fund as a research analyst covering semiconductor and downstream components sector.

Brady Wang

Hi, I’m Brady Wang, a seasoned professional with over 20 years of experience in the high-tech industry, spanning semiconductor manufacturing, market intelligence, and strategic advisory roles. Currently, I serve as an analyst at Counterpoint Research, where I specialize in semiconductors with a focus on advanced applications such as automotive, server platforms, and cutting-edge process nodes. My core research centers on AI servers and their key components, including GPUs, custom accelerators, high-bandwidth memory (HBM), CPUs, and advanced packaging technologies. I also track the evolution of AI server architectures, interconnect technologies, and data center deployment trends. By combining deep technical knowledge with market insight, I help clients navigate the fast-changing AI infrastructure landscape and make strategic, data-driven decisions.

Overview

Table of Contents

The Shift in Agentic AI Computing Infrastructure and Its Impact on Foundry and Advanced Packaging. Three Takeaways from OCP 2026 About the Shift in Agentic AI Computing Infrastructure and Its Impact on Foundry and Advanced Packaging

Agentic AI has increased the demand for CPU computing power for workflow management. Meanwhile, longer task cycles and larger context windows are driving the industry toward multi-tier memory and storage architectures that integrate HBM, CXL-based memory, and SSD storage.

As power consumption per AI rack continues to climb, reaching 140 kW, 240 kW, and even over 350 kW, high-voltage power delivery and liquid cooling solutions are becoming increasingly critical for managing power loss and deployment costs.

Advanced process nodes, chiplet technology, and 2.5D/3D packaging are helping AI chip manufacturers address challenges related to reticle size limits, production yields, and power consumption. Furthermore, the wider interface of HBM4 places more stringent requirements on the manufacturing and integration of base dies.

A14 Is Running Earlier Than Expected and Packaging Builds Out on Three Tracks

A14 yield learning is running ahead of our original expectations. We expect risk production in 2027, a ramp in H1 2028 and meaningful volume in H2 2028, roughly one to two quarters earlier than previously expected.

The initial capacity ramp for A14 will begin in Hsinchu before extending to Taichung. The Arizona facility is unlikely to contribute any capacity before 2030.

CoWoS remains the top priority for TSMC’s packaging business, with current capacity already fully utilized. Capacity expansion for SoIC is proceeding at a more measured pace, while CoPoS is unlikely to reach a meaningful scale of production before the second half of 2028.

Demand-driven capacity expansion may continue until 2030 for main Chinese foundries

Strong local demand is underpinning the expansion of SMIC in both mature and advanced process nodes on 12’’ wafers. We project that its capacity for advanced process nodes will reach 60–70 KWPM by 2030.

HHG is focusing on 12” mature-node capacity through its HLMC Fab 5 acquisition and the future ramp of Wuxi Fab 9B. Its total capacity could reach around 350 KWPM once Fab 9B is fully loaded.

TSMC Keeps Adding Capacity for N3: Fab Roadmap Update and What Customers Are Actually Asking For

Years after the N3 process node entered mass production, TSMC continues to expand its capacity, reflecting sustained demand from AI GPUs, AI ASICs, and base dies for HBM.

AMD has requested additional capacity from TSMC, amounting to approximately 25,000 wafers, for the production of its MI450 products in 2027. However, TSMC has yet to make a firm commitment to meet this incremental demand. Final shipment volumes will also depend on the availability of supporting CoWoS packaging and substrate capacity.

Apple has reduced its N2 and N3 wafer orders by 10–15% amid weaker iPhone Duo build plans and higher memory costs. We believe AI demand can absorb this reduction, although a longer consumer slowdown remains a risk.