TSMC and Amkor Finalise Advanced CoWoS Packaging Alliance to Relieve AI Silicon Bottlenecks
TSMC and Amkor Technology have finalised a multi-year packaging pact to scale CoWoS and 2.5D substrate assembly outside Taiwan for Nvidia and AMD accelerators.
By Marcus Vance, Semiconductors & Hardware
, updated 6 min read
Geographic Diversification for 2.5D Packaging Capacity
Taiwan Semiconductor Manufacturing Company (TSMC) and leading outsourced semiconductor assembly and test (OSAT) provider Amkor Technology have concluded definitive agreements on a multi-year manufacturing partnership. The alliance establishes shared production workflows for advanced 2.5D packaging, specifically scaling TSMC's proprietary Chip-on-Wafer-on-Substrate (CoWoS) and integrated fan-out technologies across Amkor's facilities in Peoria, Arizona, and Gwangju, South Korea.
Under the framework of the agreement, TSMC will manufacture cutting-edge compute chiplets on its 3nm and 2nm class logic nodes across its Taiwanese fabrication lines and upcoming Phoenix fabs, while Amkor will handle advanced wafer-level packaging, interposer placement, and final module testing. This cooperative pipeline directly addresses the most persistent operational bottleneck in artificial intelligence silicon delivery: back-end advanced packaging capacity.
Historically, TSMC conducted virtually all advanced packaging within its domestic Taiwanese gigafabs in Taichung, Tainan, and Chiayi. However, explosive demand for AI accelerators from customers including Nvidia, AMD, Apple, and cloud hyperscalers has overwhelmed front-end wafer fabrication lead times, leaving finished silicon wafers queued for months waiting for interposer integration.
CoWoS Architecture and the Interposer Constraint
The primary technical bottleneck in high-performance computing silicon is not logic gate fabrication, but the physical integration of processing cores with high-bandwidth memory (HBM). Modern architectures such as Nvidia's Blackwell and Rubin lines, alongside AMD's Instinct MI350 series, rely on passive silicon interposers measuring twice the reticle limit to link multiple compute dies to eight or twelve stacks of HBM3E and HBM4.
In standard CoWoS-S packaging, micro-bumps with pitches below 25 microns connect the compute dies to an underlying silicon wafer interposer, which provides high-density routing paths with sub-micron line spacing. Producing these interposers requires dedicated lithography equipment and chemical mechanical planarisation tools identical to front-end wafer fabrication. By licensing select process recipes to Amkor and co-locating engineering teams, TSMC can offload wafer sawing, micro-bump bonding, and organic substrate assembly without diluting proprietary intellectual property.
Amkor has committed an additional $2 billion to expand cleanroom floor space and automated die-attach equipment at its Peoria packaging hub, which sits within 40 kilometres of TSMC's Arizona Fab 21 complex. This physical proximity eliminates international air-freight transport delays and customs handoffs, establishing a domestic United States assembly corridor for defence and enterprise high-performance computing clients. The Peoria facility is scheduled to commence pilot qualification runs in early 2027.
Packaging Price Increases Reflect Severe Market Scarcity
The finalisation of the Amkor agreement coincides with TSMC notifying top-tier clients of revised price structures for advanced packaging runs scheduled throughout 2026 and 2027. Supply chain sources indicate that TSMC has instituted price increases ranging between 10% and 15% on CoWoS-S and CoWoS-L assembly lines, alongside selective price adjustments on leading-edge 3nm wafer allocations.
Industry analysts note that high-performance computing customers have largely absorbed these adjustments without hesitation. With artificial intelligence clusters trading at substantial premiums and enterprise demand remaining supply-constrained, hyperscalers prioritise guaranteed delivery allocations over per-unit assembly costs. TSMC executives stated during recent investor presentations that company-wide advanced packaging capacity will double again over the next twelve months, yet projected demand will continue to exceed available cleanroom capacity through late 2027.
The partnership also provides strategic reassurance for AMD and Nvidia. AMD chief executive Lisa Su highlighted during recent supply chain summits that multi-sourcing packaging partners is essential to support the global volume rollout of next-generation EPYC server processors and Instinct accelerator modules.
Long-Term Impact on Advanced Heterogeneous Integration
The formal collaboration between TSMC and Amkor signals a broader industry transition toward disaggregated, multi-vendor semiconductor manufacturing. As monolithic die yields decline on sub-2nm nodes, semiconductor architects are universally embracing chiplet-based heterogeneous integration, where memory, logic, optical transceivers, and power delivery circuits originate from different foundries and process nodes.
Looking beyond traditional silicon interposers, TSMC and Amkor are already running test vehicles on glass substrate packaging and panel-level fan-out architectures. Glass substrates offer superior flatness, higher thermal stability, and reduced signal loss across high-frequency interconnects compared to conventional organic bismaleimide-triazine (BT) cores.
For enterprise server purchasers and cloud infrastructure architects, the expansion of packaging capacity across Arizona and South Korea promises more predictable lead times for critical computing hardware. By systematically removing packaging constraints from the global silicon supply chain, the semiconductor industry is laying the physical infrastructure required to sustain the next decade of autonomous computing workloads.
Sources
- 1TSMC Investor Relations Quarterly Disclosuresinvestor.tsmc.com
- 2Amkor Technology Corporate Manufacturing Announcementsamkor.com