II

Advanced packaging & CoWoS

The GPU shortage has repeatedly been a packaging shortage: 2.5D/3D interposer and hybrid-bonding capacity, not wafer starts, sets accelerator output.

Advanced packaging is the chokepoint inside the chokepoint. When people say "GPU shortage," they almost always mean a CoWoS shortage: the 2.5D/3D interposer, micro-bump, and increasingly hybrid-bonding steps that stitch a logic die to its high-bandwidth memory stacks. Wafer starts have rarely been the binding constraint on accelerator output this cycle — back-end assembly has. TSMC's CoWoS line ran fully booked through 2025 and into 2026, and reports of Google cutting its 2026 TPU build target from ~4M to ~3M units on constrained CoWoS allocation are the cleanest illustration that this step, not transistors, gates units shipped.

Who owns it

The choke is unusually concentrated. TSMC owns the dominant CoWoS process and is scaling capacity from roughly 75,000-80,000 wafers/month at end-2025 toward 120,000-130,000 by late 2026. Because even that is not enough, TSMC is outsourcing an estimated 180,000-190,000 wafers/year to Amkor and 60,000-80,000 to SPIL, which is what turns the OSATs — Amkor and ASE — from cyclical assemblers into structurally bid capacity. One layer up sits the toolchain: Besi for hybrid bonders and thermo-compression bonding, and the packaging inspection/metrology duopoly of Onto Innovation and Camtek, whose tools qualify and yield-manage every interposer and HBM stack.

What breaks it

The near-term risk is not demand — it is that capacity finally catches up. TSMC's roughly-quadrupling of CoWoS output and OSAT additions are explicitly framed as ending the crunch, which is bullish for units shipped but compresses the scarcity premium on packaging slots. The second risk is technology transition: CoWoS-S/L gives way to panel-level and CoWoP, and 2.5D micro-bump interconnect gives way to hybrid bonding — each transition reshuffles which tool vendor wins. The third is hyperscaler internalization and circular financing: every custom-ASIC program (Google TPU via Broadcom, AWS Trainium via Marvell, Microsoft Maia) competes for the same CoWoS pool, and vendor-to-customer investments in the accelerator layer can flatter demand signals that ultimately still bottleneck at one back-end line.

What forces the reprice

The metrology/equipment names reprice on order intake, not just revenue, because bookings lead the capacity build. Besi's Q1-26 orders of EUR 269.7M more than doubled year-on-year on hybrid-bonding bookings, with hybrid-bonder unit orders more than doubling sequentially past the 2024 peak and the customer base widening to 20. Onto guided advanced packaging to grow more than 50% in 2026; Camtek booked over $105M of orders in June 2026 alone, including over $50M from a leading HBM player. For the OSATs, the reprice is margin mix: ASE expects advanced packaging to double to ~$3.2B in 2026 with full-process CoWoS targeting ~$300M, and Amkor posted record advanced-packaging revenue in 2025 and is spending $2.5-3.0B of capex in 2026 on top of a $7B Arizona campus that will package TSMC Arizona / Apple silicon.

Be honest about size and earliness

This is a real chokepoint but a small slice of most of these income statements. Advanced packaging is only an estimated ~7-10% of TSMC's total revenue, so TSM is a wafer-foundry story with a packaging kicker, not a pure-play. At ASE, full-process CoWoS at ~$300M sits inside a >$3B advanced-packaging line that is itself a fraction of group revenue. The purest exposure is in the toolmakers — Besi (hybrid bonding), Onto and Camtek (packaging inspection/metrology) — but those are smaller-cap, lumpy-order businesses (Camtek's entire 2025 revenue was ~$496M) where a single quarter's bookings swing the narrative. Hybrid bonding at high volume is still early: die-to-wafer hybrid bonding was only ~$275M in 2025, so much of the Besi thesis is forward adoption, not current run-rate. Treat this as a structurally tight, mostly-Taiwan-anchored supply chain where the listed pure-plays are small and order-driven, not as a basket of large-cap AI compounders.

Who owns the choke

AMKRcore

Amkor Technology, Inc.

$35.29+3.6%

FY2025 net sales were $6.71B with record advanced packaging and computing revenue; Q1 2026 was a record $1.68B (+27% YoY). Amkor broke ground on a $7B Arizona advanced-packaging campus that will package TSMC Arizona / Apple silicon, with 2026 capex of $2.5-3.0B.

Bull
If CoWoS demand stays bid, Amkor's record $6.71B sales, $2.5-3.0B capex and $7B Arizona campus turn it from cyclical assembler into structurally booked packaging capacity.
Bear
Advanced packaging is one line in a cyclical OSAT, and capacity additions across TSMC and OSATs could compress the scarcity premium it depends on.

[1] [2] [3]

ASXwatch

ASE Technology Holding Co., Ltd. (ADR)

$32.52+2.2%

ASE expects its advanced packaging and testing (LEAP) business to roughly double to ~$3.2B in 2026, with full-process CoWoS targeting ~$300M of revenue, driven by TSMC outsourcing high-end CoWoS and broad AI/HPC demand.

Bull
If TSMC keeps outsourcing high-end CoWoS, ASE's advanced-packaging line roughly doubling to ~$3.2B with full-process CoWoS at ~$300M lifts its margin mix.
Bear
Full-process CoWoS at ~$300M sits inside a >$3B line that is itself a fraction of group revenue, diluting the AI-packaging exposure.

[1] [2] [3]

BESIYwatch

BE Semiconductor Industries N.V. (ADR)

$164.59+1.2%

Q1 2026 orders hit a record EUR 269.7M (more than doubling YoY), with hybrid-bonder unit orders more than doubling sequentially across a customer base of ~20 in logic, memory and photonics. Besi anticipates the first hybrid-bonded 16-high HBM4E stacks in 2026 and HBM5 to use hybrid bonding exclusively; SK hynix, Applied Materials and Besi are publicly co-optimizing the flow for next-gen HBM.

Bull
If HBM4E 16-high and HBM5 adopt hybrid bonding, Besi's record EUR269.7M orders and SK hynix/Applied Materials co-optimization make it the purest hybrid-bonding play.
Bear
Die-to-wafer hybrid bonding was only ~$275M in 2025 and HBM4 stayed on microbumps, so the thesis is forward adoption on a lumpy-order small-cap ADR.

[1] [2] [3] [4] [5]

CAMTwatch

Camtek Ltd.

$98.42+3.6%

FY2025 revenue was a record $496.1M; in Q1 2026 Camtek booked >$90M of orders from leading OSATs mainly for CoWoS-like AI packaging, and in June 2026 received >$105M in orders including >$50M from a leading HBM player. Its Hawk system targets HBM, chiplets, hybrid bonding and wafers with up to 500M micro-bumps.

Bull
If HBM and CoWoS wafer starts scale, Camtek's record $496M, June 2026's >$105M orders and HBM-tuned Hawk inspection platform ride packaging-metrology demand.
Bear
A single quarter's bookings swing the narrative for this small-cap, where a record-year ~$496M makes it lumpy and order-driven, not an annuity.

[1] [2] [3]

ONTOwatch

Onto Innovation Inc.

$201.02+0.7%

2025 revenue hit a record $1.005B (Q4 a record $267M); in February 2026 the company disclosed a new volume purchase agreement estimated at over $240M with a leading HBM manufacturer for Dragonfly 2D inspection and 3D bump metrology systems supporting that customer's capacity expansion through 2027.

Bull
If HBM and 2.5D packaging capacity expands, Onto's record $1.005B revenue, >$240M HBM purchase agreement and >50% advanced-packaging growth ride the inspection/metrology duopoly.
Bear
HBM and packaging are slices of revenue for this small-cap, where lumpy order timing through 2027 swings the narrative quarter to quarter.

[1] [2] [3] [4]

TSMcore

Taiwan Semiconductor Manufacturing Co. (ADR)

$362.62+0.8%

Sole high-volume foundry for leading-edge AI silicon; began 2nm N2 volume production Q4 2025 at Fab 22, with 90%+ of leading-edge tape-outs on N2. Q1 2026 revenue $35.9B (+40.6% YoY), HPC 61% of revenue, 66.2% gross margin. Over 90% of capacity in Taiwan.

Bull
If frontier AI silicon demand holds, TSMC's sole-source N2 volume production, 61% HPC revenue and scaling CoWoS capacity make it the foundry every accelerator routes through.
Bear
Over 90% of capacity sits in Taiwan as a single-point-of-failure, and the AI demand pulling its stack is partly circular vendor-financed buildout.

[1] [2] [3] [4] [5] [6]

Catalyst calendar

  • 2026-07-16TSMC Q2 2026 earnings and conference callmediumTSMC's CoWoS capacity ramp pace and advanced-packaging revenue commentary set the supply ceiling for the entire accelerator chokepoint.
  • 2026-07-23Besi Q2 2026 results (order intake)mediumBesi's hybrid-bonder and TCB order intake is the cleanest leading indicator of the 3D-stacking transition inside advanced packaging.
  • 2026-07-27Amkor Technology Q2 2026 earningsmediumConfirms whether TSMC CoWoS outsourcing volumes and the $7B Arizona build are translating into sustained advanced-packaging revenue and margin.
  • 2026-10-13SEMICON West 2026 (Moscone Center, San Francisco, Oct 13-15)lowThe packaging/assembly toolmakers (Besi, Onto, Camtek) showcase next-gen hybrid-bonding and panel-level capacity and roadmap shifts that determine which vendor wins each transition.

View these names in the dashboard →