Search AI Chips News Today

Manufacturing

TSMC 5.5-Reticle CoWoS Now in Production With 98% Yield; 14-Reticle Path Targets 20 HBM Stacks

TSMC AI chips news today: how TSMC CoWoS capacity expansion, 5.5-reticle production at 98% yield, and 9.5-reticle super-carriers scale HBM4 AI processors.

B-ad-semphotoofsu8. Photo: Fraunhofer ENAS, D. Neumann Uploaded by Enaswiki at en.wikipedia (CC BY-SA 3.0 de, Wikimedia Commons).

TSMC has confirmed that its 5.5-reticle CoWoS packaging entered production in 2026 with reported yields of 98%, and the company is planning a 14-reticle CoWoS platform for 2028 that would integrate about 10 compute dies and 20 HBM stacks. That package-scale roadmap, alongside hybrid bonding demonstrations from imec and EV Group, is resetting how much memory bandwidth AI accelerators can ship with. The global 2.5D/3D packaging market is valued at USD 14.3 billion in 2026 and forecast to reach USD 51.9 billion by 2035, as reported by WhatTheyThink and tracked by SEMI Global Semiconductor Alliance.

What limits do interposer reticle sizes impose on HBM stack counts?

A single reticle-sized interposer caps how many compute and HBM dies can sit side by side, so larger reticle counts directly increase the number of HBM stacks an accelerator can address. 2.5D packaging links dies side by side via an interposer or embedded bridge, while 3D packaging stacks dies vertically. TSMC's current 5.5-reticle CoWoS is the world's largest production interposer package and achieves 98% yields. The planned 14-reticle CoWoS for 2028 targets about 10 compute dies and 20 HBM stacks; TSMC then expects 24 HBM stacks in 2029. This stack count matters because memory bandwidth per accelerator scales with the number of HBM sites, not just HBM capacity. Buyers evaluating AI systems should pair package size with choosing memory for AI workloads to understand delivered bandwidth.

How does hybrid bonding change interconnect density and energy efficiency?

Hybrid bonding replaces solder microbumps with direct copper-to-dielectric bonds, enabling much finer pitches that scale interconnect density and lower energy per bit. imec and EV Group demonstrated wafer-to-wafer hybrid bonding at 200-nanometer copper pitch on May 28, 2026, with post-bond alignment accuracy below 40 nanometers. Hybrid bonding joins copper contacts and dielectric directly, which supports fine-pitch connections versus solder. TSMC's SoIC technology delivers 56 times the interconnect density and 5 times the energy efficiency versus 2015-era CoWoS. First-generation SoIC is in mass production; a 6-micron bonding pitch version is set for 2025, the N2-generation SoIC supports 6-micron stacking in 2028, and the A14 generation targets 4.5-micron. These pitch reductions are the mechanical basis for 3D chiplet stacking and influence how much logic can be placed above or below HBM.

Why does 5.5-reticle CoWoS production matter for current AI systems?

It is today's largest production interposer package, so it sets the practical upper bound for shipping high-bandwidth memory alongside monolithic or chiplet compute dies. TSMC confirmed 5.5-reticle CoWoS entered production in 2026, and the company projects AI accelerator wafer demand growing 11-fold between 2022 and 2026. CoWoS capacity is forecast to grow above 80% CAGR from 2022 to 2027. TSMC also raised its global semiconductor market forecast to over USD 1.5 trillion by 2030 from USD 1 trillion. Kevin Zhang, Deputy Co-COO at TSMC, said that "future AI accelerator performance will depend on the integration of transistor computing, advanced packaging, and high-speed interconnect technologies." For more on how these packaging decisions shape accelerator platforms, see the AI chipmakers coverage hub.

What yield economics should buyers track across advanced packaging suppliers?

Reported and rumored yield figures for Intel 18A, TSMC N2, Samsung SF2, and Intel EMIB-T show a wide spread that directly affects package cost and availability. Intel 18A yields improved to 85% from 65% in the prior quarter, trailing TSMC N2 at 90%. Samsung SF2 yields are reported at 50-60%. Intel's EMIB-T packaging reached a rumored 98% yield, up from 90% three months earlier, but that figure is unverified. Alleged EMIB customers include NVIDIA's Feynman GPU, Google's TPU HumuFish, and Amazon AWS Trainium 3, although these wins are still rumored. Intel is insourcing 80-90% of Nova Lake tiles and expanding 18A and Intel 3 capacity. Intel also plans glass substrate packaging at its Rio Rancho facility, enabling larger reticle dies through EMIB bridges. For hardware buyers, the AI Chip Buyer’s Guide recommends practical checks:

  • Confirm whether Intel EMIB-T 98% yield is vendor-verified or analyst rumor before building cost models.
  • Check whether design wins involving AMD, NVIDIA, Marvell, Microsoft, Micron, and OpenAI are confirmed or still rumored.
  • Account for package-level yield, not just die yield: 5.5-reticle CoWoS at 98% still leaves risk at 14-reticle and beyond.
  • Ask whether a supplier's hybrid bonding pitch roadmap (200 nm demonstration, 4.5-micron A14 SoIC) matches your chiplet I/O density requirements.
  • Treat advanced packaging capacity buildouts, such as Amkor phase two adding 60,000 square meters cleanroom, as lead indicators for supply availability.

Where is advanced packaging capacity actually being built, and what remains unannounced?

TSMC is expanding CoWoS and Arizona fabs, Intel is expanding 18A and Intel 3 capacity, and Amkor is adding a second Arizona cleanroom phase. Amkor's phase two adds 60,000 square meters of cleanroom space, bringing the campus to about 93,000 square meters; construction starts late 2027 and completion is targeted for end of 2029, with total planned investment of USD 12 billion. TSMC's Arizona first fab is producing, the second enters tool move-in in 2H26, and the third is under construction, with Arizona output expected to grow 1.8-fold by 2026. Intel 18A is in volume ramp, 18A-P is in risk production, and 14A is set for risk production in 2028 and volume in 2029. Several details remain unannounced or unverified: whether Intel's reported design wins with AMD, NVIDIA, Marvell, Microsoft, Micron and OpenAI are confirmed or still rumored; whether Intel EMIB-T 98% yield figure is verified; whether alleged EMIB customers NVIDIA Feynman, Google TPU HumuFish and AWS Trainium 3 are confirmed; full details of the imec/EVG hybrid bonding demonstration; and whether TSMC's CoPoS fan-out panel-level packaging timeline is set. These unknowns matter because they determine how much of the 14-reticle and HBM stack roadmap is firm supply versus planning. More coverage of these capacity expansions is available at the AI chip manufacturing hub.