AI Chip Advanced Packaging, HBM, and Test Equipment: A Deep Dive

You are reading an earlier report. A newer report on this company was published on May 20, 2026: AI Content Copyright and Data Licensing: Investment Research

Lead
The real bottleneck in AI chip supply is the compound scarcity of high-end HBM, CoWoS / large-format advanced packaging, N3 / HBM base die, and test/probe capacity. The most certain revenue and margin upside sits with HBM makers, TSMC's 3DFabric, and test and hybrid-bonding equipment leaders. Rating Watch: track TSMC, SK hynix, Micron, ASE, Advantest, BESI, Samsung, Broadcom, FormFactor, and Unimicron, while treating ABF / glass substrate / broad AI PCB names as sympathy plays that still need orders and disclosed breakouts.

Core Conclusions

AI GPUs and AI ASICs are no longer a pure "advanced-process race." They are a systems-engineering contest of "advanced process + HBM + advanced packaging." TSMC explicitly positions CoWoS as a packaging solution for AI and supercomputing, Samsung defines HBM as the key memory underpinning AI infrastructure, and Google's published TPU v6e specs list HBM capacity and bandwidth directly as core parameters.
The true bottleneck in today's AI chip supply is not a single step but a compound one: "high-end HBM supply + CoWoS / large-format advanced packaging capacity + leading-edge logic / N3 and HBM base die capacity + test/probe capability." On its Q1 2026 earnings call, TSMC again stated plainly that its advanced packaging capacity is "very tight," while Micron and SK hynix have each publicly said AI-related memory demand is supply-constrained or "well beyond capacity."
The most direct and most certain revenue elasticity comes from three kinds of companies: first, HBM suppliers; second, leading foundries like TSMC with CoWoS / SoIC / 3DFabric capability; and third, the leaders in test / probe / advanced packaging equipment. By comparison, ABF substrates, PCBs, general OSAT, and most materials companies look more like "sympathy" or "cyclical" plays than core bottleneck assets.
On margins, HBM and leading automated test equipment (ATE) are the best positioned. Micron's fiscal Q2 2026 non-GAAP gross margin reached 74.9%, with both its cloud memory and core data center business unit gross margins at 74%; TSMC's Q1 2026 gross margin was 66.2% and operating margin 58.1%; and Advantest raised its full-year profit guidance on AI test demand.
On supply scarcity, the hybrid bonding, base logic die, and advanced testing tied to CoWoS and HBM4 are scarcer assets than commodity ABF or ordinary OSAT. TSMC has cited warpage, thermal limits, and large-reticle packaging challenges from oversized AI dies; BESI has already won HBM4 hybrid bonding orders from two leading memory makers.
The companies that have already formed a "real orders → capacity expansion → profit growth" loop are most clearly represented by TSMC, SK hynix, Micron, ASE, Advantest, and BESI. They map respectively to logic + packaging, HBM, HBM, packaging and OSAT capture, testing, and hybrid bonding equipment, the most critical links in the value chain.
Strong narrative but still-thin financial evidence is concentrated in two buckets: one is the "glass substrate / panel-level packaging / advanced thermal management" theme names, many still at validation, pilot, or distant-roadmap stage; the other is some PCB, materials, and second-tier OSAT names whose public disclosures make it hard to break out AI advanced packaging revenue and profit separately. These directions are worth tracking but should not be lumped in with the high-certainty HBM/CoWoS assets. This judgment is a research inference based on the disclosure coverage of this cycle.
The upgrade from HBM3E to HBM4 / HBM4E will push profit allocation further from "pure stacked DRAM" toward "DRAM die + advanced base die + advanced packaging + advanced testing + equipment/materials." TSMC has stated its 3nm capacity will go to HPC/AI, including HBM base dies; both Samsung and SK hynix emphasize HBM4's 2,048 I/O and its higher power / thermal management demands; and Micron has publicly stressed that HBM4E's customization options will deepen customer collaboration.
TSMC's CoWoS roadmap remains the dominant platform for training-grade AI accelerators. Intel EMIB/Foveros offers cost/architecture appeal in bridging and thermal management, but public evidence still sits mostly at potential talks and solution promotion; Samsung I-Cube/X-Cube has full memory + foundry + packaging integration potential, but its external large-customer production stickiness and market share still trail the TSMC/NVIDIA ecosystem.
In-house ASICs will not weaken the importance of HBM and advanced packaging; instead they will broaden demand from "single NVIDIA pull" to "multiple bloom points across cloud players." Broadcom expects Q2 2026 AI chip revenue of 10.7 billion USD and forecasts cumulative AI chip sales to exceed 100 billion USD by 2027; Marvell's custom AI silicon has entered volume production; and Google TPUs already treat HBM as standard configuration.
Inference growth will change the shape of demand but will not eliminate HBM demand. The reason is that agentic AI, long context, KV cache, and MoE routing are all lifting memory and bandwidth requirements. Micron has stated plainly that AI is making compute architectures "more memory-intensive," while TSMC treats rising token consumption as evidence that AI demand keeps strengthening.
In today's valuations, the AI expectations of popular names like SK hynix, Micron, TSMC, Advantest, BESI, and Broadcom are already largely priced in; ASE, some probe/substrate names, and a few second-tier advanced packaging capture players that have yet to be fully re-rated may still hold an expectations gap, but only if subsequent orders and AI revenue breakouts come through.
What is most likely to move from shortage to glut first is not top-tier CoWoS / SoIC / hybrid bonding, but secondary advanced packaging, mature substrate tiers, and broad AI PCB/materials capacity; the real thing to watch for is price competition in the legacy HBM3E generation and second-tier capture capacity after 2027. This conclusion is a research judgment based on the current pace of capacity expansion and technology-generation transitions.
The biggest risk is not a single technology substitution but the pace of cloud AI CapEx, the rebalancing of HBM supply and demand, overly fast advanced packaging expansion, and the redistribution of value-chain profit as the shift from training to inference unfolds. Public statements from TSMC, Micron, Advantest, and Broadcom all confirm that AI demand remains strong, yet the market's pricing of that strong demand is already very high.

Value-Chain Landscape and Benefit Attribution

Value-Chain Landscape Table

Value-chain position	Sub-segment	Core products	AI demand driver	Revenue recognition	Key customers	Supply bottleneck	Margin profile	Representative companies	Listed/Unlisted	Benefit strength	Investment elasticity	Source
Memory core	HBM2E/HBM3/HBM3E/HBM4/HBM4E	TSV-stacked HBM, base die, MR-MUF/NCF solutions	Extreme demands on bandwidth, capacity, power, and size from training/inference chips	Shipments and long-term agreements/SCA	NVIDIA, AMD, cloud ASICs	DRAM die, base die, packaging yield, testing	High and rising, but still subject to the memory cycle	SK hynix, Micron, Samsung	Listed	5	5
Memory complementary	DDR5/LPDDR/enterprise SSD/CXL	Server memory, SSD, LP DRAM, CXL	Inference and data-layer storage expansion; KV cache, vector databases, capacity-tier migration	Module/SSD shipments	CSPs, OEMs, server makers	Bit supply crowded out by HBM	Mid-high, with higher volatility than HBM	Micron, Samsung, SK hynix, Kioxia	Listed	3	4
Wafer foundry	Leading-edge logic + HBM base die	N3/N4/N5/N2, HBM base die	GPU/ASIC logic scale-up and HBM4 custom base die	Wafer shipments	NVIDIA, AMD, Broadcom, cloud players	N3/N4, advanced packaging coordination	Very high	TSMC, Samsung Foundry, Intel Foundry	Listed	5	4
Advanced packaging core	CoWoS-S/L/R, SoIC, 2.5D/3D	Chip-on-Wafer, interposer, large-format packaging	Oversized GPU die + multiple HBM integration	Packaging service/turnkey	NVIDIA, AMD, ASIC design houses	CoWoS large format/warpage/thermal limits	High	TSMC 3DFabric	Listed	5	5
Alternative 2.5D/3D routes	EMIB/Foveros, I-Cube/X-Cube	silicon bridge, 3D stacking, Fo-PLP	Lower the cost of large silicon interposers, support chiplet/ASIC diversity	Foundry/package service	Intel internal and potential external customers, Samsung Foundry customers	External-customer qualification, ecosystem maturity	Depends on customer ramp	Intel, Samsung	Listed	3	4
OSAT capture	High-end packaging and test	flip-chip, SiP, partial CoWoS-class capture	TSMC overflow and regional division of labor	Packaging and test services	Fabless, IDM, foundry coordination	High-end equipment/customer qualification	Mid-high, scale-driven	ASE/SPIL, Amkor, JCET	Listed	4	4
Substrate	ABF substrate/high-layer-count substrate	ABF, HDI, hybrid substrate	Large packages, more power/signal layers, chiplet fan-out	Substrate shipments	Packaging houses, GPU/ASIC platforms	High-layer-count yield, size/warpage	Upper-mid but strongly cyclical	Ibiden, Shinko, Unimicron, AT&S	Listed	3	4
Interposer	Silicon interposer / RDL interposer	Si interposer, RDL, Local Silicon Interconnect	Rising HBM count and shrinking die-to-die distance	Packaging integration service	GPU/ASIC platforms	Area, yield, routing density	High technical barrier	TSMC, Samsung	Listed	4	4
Bonding/stacking	TSV, MR-MUF, Hybrid bonding	TSV, Cu-Cu bonding, NCF/MUF	HBM4, 16Hi, 3D logic stacking	Equipment/materials/process	HBM makers, foundry	pitch, thermal, yield	High barrier	SK hynix, Samsung, BESI, AMAT	Listed	5	5
Packaging equipment	die bonder, hybrid bonder, wafer bonding	bonding, thinning, flip-chip, hybrid bonding	Rising HBM4 and 3D packaging complexity	Equipment delivery	Memory makers, foundry, OSAT	Process precision and lead times	High elasticity	BESI, ASMPT, K&S, AMAT	Listed	4	5
Test equipment	ATE / burn-in / final test	SoC tester, memory tester	Rising AI chip test duration, complexity, and binning	Equipment delivery and services	IDM, OSAT, fabless	Test throughput, delivery	High gross margin	Advantest, Teradyne	Listed	4	4
Probe cards	HBM/SoC probe card	High pin count, high frequency, high-temperature testing	HBM stacking and high-end logic wafer testing	Probe card shipments	Memory makers, logic makers	Pin count and consistency	Mid-high	FormFactor, Technoprobe	Listed	3	4
Packaging materials	underfill, MR-MUF, NCF, ABF film	UF, MUF, NCF, Build-up film	Higher layer counts, higher thermal flux, finer pitch	Material shipments	Packaging houses, memory makers	Long qualification cycles	Mid-high	Ajinomoto, Resonac, Namics, Shin-Etsu	Mostly listed	3	4
Thermal management	TIM, liquid-cooling coordination	thermal interface materials, cold plate/liquid-cooling kits	Rising power/thermal density per rack	Material/system sales	Server OEMs, cloud players	Qualification and system integration	Medium	Boyd, Laird, Auras/AVC, etc.	Mixed	2	3	Research inference based on rising AI server thermal density; detailed financials were not individually verified this cycle.
Design services	ASIC/packaging co-design	HBM PHY, packaging co-design, chiplet design	Growing AI ASIC customer base, with packaging becoming the front-end design boundary	NRE + production royalties	Broadcom/Marvell ecosystem, cloud players	Design IP and production coordination	Mid-high	GUC, Alchip, Marvell, Broadcom	Mostly listed	4	4
Chip customers	GPU/AI ASIC	accelerators, switch chips, networking chips	Training/inference cluster expansion	Chip shipments	CSPs, enterprises, sovereign AI	Packaging and HBM coupling	Software/ecosystem determines profit	NVIDIA, AMD, Broadcom, Marvell	Listed	5	4
Cloud players and model companies	TPU/Trainium/Maia/in-house XPU	training/inference clusters	Model scale, token consumption, agentic AI	CapEx converted to orders	Google, AWS, Microsoft, Meta	Capital spending and upstream supply	Not an on-chain profit measure	Google, Amazon, Microsoft, Meta	Mostly listed/internal business units	5	4

Who Benefits Most Directly

The most direct beneficiaries are not "every semiconductor company connected to AI," but the links that convert AI chip expansion almost linearly into orders: HBM suppliers, TSMC CoWoS/3DFabric, the ASE/SPIL pulled by TSMC overflow demand, the AI test chain (Advantest / high-end probe cards), and the hybrid bonding equipment and key materials required for HBM4 / 3D packaging. TSMC itself discloses that advanced packaging is still very tight, ASE expects its advanced packaging business to double to 3.2 billion USD in 2026, and BESI has already won HBM4-related orders, showing that order transmission has moved from "story" to "financials."

By comparison, ABF, PCB, general OSAT, and ordinary materials will still benefit, but more as "second-order" beneficiaries. These links typically need to wait until platform shipments ramp, packaging solutions stabilize, and customer qualification completes before revenue shows up clearly, and they more easily fall back into cyclical competition after large-scale expansion. Samsung H-Cube publicly explains that large-format 2.5D will use an ABF/HDI combination, but at this stage the industry's public disclosures still emphasize HBM, advanced packaging, and test bottlenecks rather than ABF being the sharpest global constraint.

Demand, Technology Roadmaps, and Bottlenecks

Why AI GPUs and AI ASICs Cannot Live Without HBM

The reason is simple: models keep getting bigger, and parameters, activations, KV cache, expert routing, and long context all raise "the amount of data that must be fed to the chip per unit of compute." In its fiscal Q2 2026 prepared remarks, Micron stated directly that AI is making compute architectures "more memory-intensive," and listed longer context, deeper reasoning chains, and multi-agent orchestration as core drivers of growing memory demand. Google's official TPU v6e documentation also lists 32GB HBM and 1,638 GiB/s of HBM bandwidth directly as per-chip baseline specs.

HBM is not simply "large-capacity DRAM." It vertically stacks multiple DRAM dies via TSV and delivers ultra-wide I/O and higher energy efficiency to the logic die over a much shorter packaging distance. Samsung's current product pages state that a single HBM3E stack can reach 1,180 GB/s, while HBM4 adopts 2,048 I/O, 12H, and up to 13.0 Gbps, corresponding to up to 3,300 GB/s; SK hynix says its HBM4 doubles bandwidth versus the prior generation through 2,048 I/O, with energy efficiency improved by more than 40%.

How Bandwidth, Capacity, and Stack Height Drive Performance

Bandwidth determines whether matrix multiplication and attention modules can keep the compute "fully fed," capacity determines how much of the model/context can stay in near-end memory, and power determines whether the system can stack more compute under a given rack power and liquid-cooling envelope. Raising the stack from 8Hi to 12Hi and on toward 16Hi is essentially a way to increase per-stack capacity without meaningfully expanding the packaging boundary, but the cost is that thermal, warpage, bonding, and test difficulty all rise in step. Both SK hynix's HBM4 and Samsung's HBM4 set 12Hi as the current mainstream direction, while Micron has publicly shown an HBM4 16-high sample.

How Much HBM a Single AI GPU or ASIC Needs

Public data shows high-end training accelerators have entered the era of "over 100GB of HBM per chip." NVIDIA's GB200 Grace Blackwell Superchip is officially specced with two Blackwell GPUs sharing 372GB HBM3E; the HGX B200 platform is officially specced at 8 GPUs totaling 1.4TB of total memory; Google TPU v6e is 32GB HBM per chip; Google Ironwood is reported in public media as 192GB HBM3e per chip; and AMD MI325X is reported in public media at 256GB HBM3E. From this, HBM demand density at the training-grade platform level is already in the mainstream range of roughly 175GB–256GB per accelerator, and cloud players' in-house TPUs/ASICs have not escaped their HBM dependence either, only differing somewhat in capacity structure.

The Upgrade Logic from HBM3E to HBM4

HBM3E's core is pushing per-pin speed and stack height higher; HBM4's core is raising interface width, base die complexity, power integrity, and system co-design all at once. SK hynix publicly stresses HBM4's 2,048 I/O, >10Gbps operating speed, Advanced MR-MUF, and 1bnm process; Samsung emphasizes HBM4 using a 4nm logic base-die; and Micron has publicly tied HBM4 volume shipment and its future HBM4E custom option to the next generation of AI platforms. In other words, HBM4 is not "a slightly faster HBM3E," but more like an integrated system upgrade of "memory + logic + packaging + testing."

How HBM4 Will Reshape Competition Among the Three Memory Makers

SK hynix still holds a first-mover advantage. Reuters, citing Macquarie data, says SK hynix held roughly 61% HBM market share in 2025, with Samsung at 19% and Micron at 20%; for 2026, another Reuters report describes Samsung's share at roughly 22% and SK hynix at roughly 57%. While the latter is a sell-side/media figure rather than formal company disclosure, it is enough to show the landscape is still "one giant, two strong players."

But in the HBM4 phase, Micron's catch-up is stronger than the market expected early on. Micron has publicly stated that HBM4 36GB 12H began volume shipment in Q1 2026, aimed at NVIDIA Vera Rubin, and that HBM4E will reach volume production in 2027; this means Micron is no longer just a "second-tier filler," but has a chance to take share upward through process nodes, base die, and customer customization. Samsung's advantage lies in its integration of memory, foundry, and packaging, but from current public evidence, the pace of qualification for its external training-grade platforms still trails SK hynix and Micron.

Why CoWoS Became a Bottleneck

Because HBM is not a DIMM plugged into a motherboard; it must be co-packaged with the GPU/ASIC over an extremely short distance. TSMC's official CoWoS page states clearly that CoWoS-S accommodates logic chiplets and HBM via a large-format silicon interposer; the interposer can reach up to about 3.3x reticle (roughly 2,700 mm²), and larger designs must shift to CoWoS-L or CoWoS-R. The more HBM stacks, the bigger the die, and the wider the I/O, the more the interposer area, power delivery layers, thermal design, and warpage management grow exponentially in complexity.

More importantly, on its Q1 2026 call, TSMC again stressed that the current "main approach" is still large-format CoWoS, that advanced packaging capacity remains very tight, and that it must partner with OSAT to expand capacity; at the same time the company is advancing CoPoS and larger-package pilots. This shows that even when front-end wafers are sufficient, high-end back-end packaging can still choke GPU/ASIC shipments.

Ranking the Real Bottlenecks Today

My ranking is: high-end HBM supply > CoWoS / large-format advanced packaging > leading-edge logic / N3-N4 and HBM base die > test/probe > high-end ABF/substrate > ordinary packaging equipment and general materials.

There are three bases for this ranking. First, TSMC says plainly that advanced packaging is very tight, and its 3nm expansion has already listed HBM base die separately as a demand source; second, both Micron and SK hynix put AI demand and supply constraints at the center; third, the order changes at Advantest and BESI show that test and bonding equipment are becoming "shadow bottlenecks." ABF still matters, but compared with 2021–2022, it is no longer the single sharpest constraint on today's AI upstream.

Transmission to NVIDIA, AMD, Broadcom, Marvell, and Cloud ASICs

NVIDIA remains the largest single source of demand pull for HBM and CoWoS, because the published specs of the Blackwell/GB200 platform already show extremely high HBM density and ultra-large-scale NVLink interconnect. AMD's MI300/MI325/MI350 series have also entered the high range on per-chip HBM capacity, which matters because it opens up greater HBM and packaging share space for a second supply chain. Broadcom and Marvell broaden demand from "selling GPUs" to "selling custom XPUs / switching / interconnect," with Broadcom even providing an outlook of cumulative AI chip sales exceeding 100 billion USD in 2027, showing that cloud players' in-house ASICs are diffusing HBM and advanced packaging demand from a single customer to multiple hyperscalers.

Google TPU v6e is already an HBM platform visible in official documentation; as Ironwood, Trainium, Maia, and others keep evolving, in-house ASICs will not make HBM demand disappear, but will change the sub-segment mix of packaging routes: some will go to TSMC CoWoS/SoIC, and some may accept relatively more cost-effective options like EMIB/bridge/RDL/FO-PLP.

Will Inference Growth Weaken HBM and Advanced Packaging

It will not simply weaken them; it will differentiate them. Low-cost, low-latency inference ASICs and marginally better inference GPUs will shift part of the demand from "extreme training-grade HBM configurations" toward structures that "emphasize cost/power ratio"; but agentic AI, long context, retrieval augmentation, and KV cache will also keep pushing many inference workloads toward high-bandwidth memory. Micron has explicitly tied inference expansion to full-stack demand across HBM, LP, DDR, and SSD, and TSMC treats rising token consumption as a supporting factor for the AI accelerator CAGR continuing to approach the "higher 50s."

Three Scenario Assumptions

Dimension	Conservative	Base	Aggressive
Assumption	Cloud CapEx growth slows, training-to-inference migration faster	Both training and inference grow, token consumption keeps rising	Sovereign AI + hyperscaler + ASIC concurrent expansion
AI GPU/ASIC shipment pace	Growth slows but does not turn negative	High prosperity continues	Sharp upside
HBM demand	Still grows, but skewed toward HBM3E / mid-capacity	HBM3E→HBM4 transition accelerates	HBM4/HBM4E ramps fast
CoWoS/advanced packaging demand	High-end CoWoS still tight, second-tier packaging eases	CoWoS / large packages stay tight	CoWoS / 3D packaging tighter still
ABF/substrate demand	Limited improvement	High-end substrate improves	High-end ABF tight again
Equipment/materials demand	Test/maintenance stronger than new equipment	Hybrid bonding / test equipment benefit most	Equipment lead times extend
Core beneficiary links	Test, HBM leaders, a few OSAT	HBM, CoWoS, test, Hybrid bonding	Whole-chain resonance, but bottleneck assets strongest
Representative companies	Advantest, Micron, ASE	TSMC, SK hynix, Micron, BESI	TSMC, SK hynix, Micron, ASE, Advantest, BESI
Key risks	CapEx slowdown, legacy-generation price cuts	Yield, expansion pace	Overheated valuations, later glut

These three scenarios are research inferences based on public statements from TSMC, Micron, Broadcom, Google, and others, not the companies' own guidance. TSMC maintains a long-term AI accelerator revenue CAGR approaching the "higher 50s," Micron clearly believes 2026 data center DRAM/NAND bit TAM will for the first time exceed 50% of industry TAM, and Broadcom has publicly given a strong mid-term outlook for AI chip revenue.

Segment Breakdown and Competitive Landscape

Segment Breakdown Matrix

Segment	Segment logic	How AI demand becomes revenue	Current supply-demand/price/margin	Technical barrier/yield challenge	Catalysts over the next 12–24 months	Key risks	Investment appeal
HBM	The most core memory form for AI compute	Directly tracks GPU/ASIC shipments	Persistently tight supply, strongest profitability	TSV, stacking, thermal, testing	HBM3E/4 ramp	Cycle reversal	5
HBM3E	Current mainstream high-end generation	Tied to Blackwell, MI325/350, etc.	Still the main revenue source	12H yield and thermal	12H/16H scale-up	Generation switch after 2027	5
HBM4/HBM4E	Succeeds HBM3E, more complex base die	Tied to Rubin / next-gen ASICs	Higher ASP and added value	2048 I/O, base die, testing	2026–2027 customer ramp	Production ramp	5
DRAM/NAND/enterprise SSD	Driven by the AI data layer	Server, SSD, capacity-tier migration	Recovering but more cyclical than HBM	Bit supply allocation	Data center NAND acceleration	Price volatility	3
CoWoS	The main platform for large chips + multiple HBM	Volume packaging directly recognizes revenue	Still tight, strong pricing power	Scale, warpage, thermal	TSMC expansion, U.S. packaging footprint	Overly fast expansion	5
SoIC	3D logic stacking, long-term upgrade	High-end 3D integration and power optimization	Still more medium-to-long-term	Bonding, thermal, EDA/co-design	Larger reticle and maturing 3D routes	Slow customer ramp	4
EMIB/Foveros	Intel's bridge/3D route	Large upside if external foundry customers ramp	Public external volume orders still limited	Ecosystem and qualification	EMIB-T / external large-customer trial orders	Customer wins below expectations	3
Samsung I-Cube/X-Cube	Integrated memory + foundry + packaging	Elasticity if it locks in AMD/cloud players	Full solution set, share still to improve	Both 2.5D/3D need qualification	HBM4 + external foundry customers	Ecosystem weaker than TSMC	3
2.5D/3D packaging	The main battlefield of the chiplet era	Tracks AI chip enlargement	High-end tight, commodity no longer scarce	Signal/power/thermal/warpage	Large-format platform iteration	Second-tier capacity glut	4
Chiplet	ASIC diversification and system-level division of labor	NRE + production platform	Long-term upward	Co-design is hard	Cloud players' in-house ASIC acceleration	Software/interconnect not standardized	4
Silicon interposer	Core bridge for training-grade platforms	Recognized together with CoWoS	Supply constrained by CoWoS	Area and yield	Rising HBM count and larger dies	High cost	4
RDL interposer	Cost/size compromise	Better suited to some ASICs and large packages	Rising penetration over the medium-long term	Line width/spacing and reliability	CoWoS-R/I-CubeE advancement	Performance below pure Si interposer	3
ABF substrate	Power/routing foundation	Tracks rising high-end package complexity	High-end improving, industry overall no longer extremely tight	Layer count, warpage, size	Rising AI substrate mix	Capacity backwash	3
Glass substrate	Long-term route, not the current profit mainline	If adopted, changes large-package cost/precision	Still more experimental/pre-qualification	Reliability/process chain immature	Pilot lines/sample validation	Slow commercialization	2
Hybrid bonding	The core upgrade point of HBM4/3D packaging	Equipment/materials/NRE benefit directly	Orders have appeared, revenue inflection nearing	<4µm pitch, cleanliness, alignment	HBM4, logic 3D chip volume production	Production yield	5
TSV	The base process for HBM	Driven directly by HBM growth	Steady growth	Deep holes, copper fill, reliability	HBM4/16Hi	Partial substitution by new routes	4
Packaging test	Complexity and test duration rise in step	Service revenue and turnover rise	High-end capacity tighter	Thermal test, burn-in, parallelism	Blackwell/HBM4 lengthen test time	Equipment delivery bottleneck	4
Probe cards	The hidden bottleneck of wafer sort	HBM/advanced logic testing	Demand strengthening	High frequency, high pin count, thermal stability	High-end SoC/HBM iteration	Long customer qualification	4
Test equipment	Mandatory complement to AI chip shipments	Equipment revenue + services	Strong growth, high gross margin	Software, parallelism, precision	AI tester continued pull	Capital spending volatility	5
Packaging equipment	The shovel of advanced packaging upgrades	Equipment delivery recognizes revenue	High-end stronger than commodity	Precision, capacity, delivery	hybrid bonder, wafer bonder	Cyclicality	4
Packaging materials	Adhesion/fill/reliability	Continued shipments after qualification	Mid-high gross margin but fragmented	Long qualification cycle	HBM4 and 3D solution switch	Customer price pressure	3
Thermal management materials	Essential for high-thermal-flux AI servers	System part-number adoption	Big opportunity but fragmented	Reliability and system coupling	GB200/liquid-cooling penetration	Not a bottleneck, heavy competition	2
AI chip packaging design services	Front-back co-design becomes a hard requirement	NRE + production platform fees	Elasticity depends on large-customer wins	Packaging/PHY/power co-design	Cloud-player ASIC proliferation	Customer concentration	4

The scores above are research scores by this report based on public disclosures, not company-disclosed data. High-scoring segments share only five traits: a short link between revenue and AI shipments, hard yields, scarce capacity, long customer qualification, and difficulty of substitution. The core segments that follow from this are only five: HBM, CoWoS / large-format advanced packaging, HBM4 / Hybrid bonding, test equipment / probe cards, and high-end OSAT able to capture overflow demand.

Conclusions on Route Competition

TSMC CoWoS remains the de facto standard for training-grade AI chips. Its advantage is not a single-point process but the coordination of front-end logic, HBM base die, CoWoS-S/L/R, SoIC, system-level co-design, and capacity scheduling within one system. In Q1 2026, TSMC was still stressing that large-format CoWoS is the current main supply while advanced packaging capacity is very tight, which in itself reflects its moat.

The strategic value of Intel EMIB/Foveros lies in that, when a full large silicon interposer is not needed, a bridge structure may be more attractive on cost, thermal, and modularity; there have also been reports of potential cooperation with Google/Amazon, SK hynix, and others. But as of this report's research, public evidence is still mostly "talks/testing/promotion" rather than already-formed large-scale external production revenue, so it should still be regarded as a potential alternative route rather than an already-formed main chain.

Samsung's I-Cube/X-Cube route has good technology reserves, with I-CubeE, H-Cube, and X-Cube hybrid copper bonding in particular showing its full-stack footprint in RDL, ABF, large format, and 3D stacking. But from an investment standpoint, a "viable" technology route does not equal "already realized" financial benefit. To truly close the gap with TSMC, Samsung still needs to simultaneously win more external large-customer qualifications, HBM supply share, and a volume-production track record in high-end packaging.

On the OSAT side, ASE is better positioned than most peers, because it is both the world's largest OSAT and explicitly named by Reuters, with its SPIL a key packaging supplier for NVIDIA AI chips, while also publicly setting a target to double its advanced packaging business to 3.2 billion USD in 2026. Amkor's U.S. advanced packaging campus has strategic significance, but its production start in 2027–2028 makes its elasticity for the next two years' results less clear than ASE's. JCET, Tongfu, Huatian, and others reflect more of a regional-substitution and long-term-upgrade direction, and are still hard to put on par with the CoWoS main chain in the near term.

Company Tiering, Scoring, and Key Research List

Company Tiering and Investment Priority

Tier	Company	Rationale
Tier A	TSMC	Core platform for AI logic + HBM base die + CoWoS/SoIC; highest order certainty and pricing power.
Tier A	SK hynix	HBM leader; Q1 2026 revenue and profit surge, with demand explicitly "exceeding capacity."
Tier A	Micron	HBM4 volume shipping, clear HBM4E roadmap, record margins and cash flow.
Tier A	ASE	One of the most direct captures of TSMC overflow, with clear advanced packaging targets.
Tier A	Advantest	AI test equipment leader; demand already reflected in raised profit guidance.
Tier A	BESI	The most core equipment elasticity for HBM4 / Hybrid bonding.
Tier B	Samsung Electronics	Full-stack technology, large potential, but HBM/advanced packaging financial realization still weaker than SK hynix/TSMC.
Tier B	Broadcom	Not a direct seller of HBM/packaging, but strong AI ASIC order certainty; one of the most important signal lights on the demand side.
Tier B	Marvell	Custom AI already in volume production, but customer concentration and competition remain key constraints.
Tier B	FormFactor	Probe cards sit at a shadow-bottleneck position, but public financial breakouts are less complete than Tier A.
Tier B	Ibiden / Shinko / Unimicron	High-end substrates have a benefit path, but remain fundamentally cyclical, and the AI share needs continued validation.
Tier B	Amkor	High value in the U.S. advanced packaging option, but financial realization over the next two years is slower than the market story.
Tier C	Applied Materials	Its BESI stake hints at a bet on hybrid bonding, but its own AI packaging revenue transmission is more indirect.
Tier C	ASMPT / K&S / Onto	Related to packaging equipment/inspection, but public evidence is less direct than BESI/Advantest.
Tier C	JCET	Long-term advanced packaging upgrade opportunity, but lacking sufficient public evidence of high-end global AI GPU/HBM production chains. This item is a research inference.
Tier C	Shennan Circuits / WUS Printed Circuit / Suntak / Kinwong	AI server and high-layer-count boards have a benefit path, but are far from HBM/CoWoS monetization. This item is a research inference.
Tier D	Glass substrate theme cluster	Distant route, long validation, weak near-term profit evidence. Based on this cycle's verification, not included in the high-certainty list.
Tier D	Broad "AI packaging materials" theme cluster	Without clear HBM/CoWoS qualification and separately disclosed revenue, treated for now as narrative over financials.

Scoring Model

I use the following weights: AI demand direct exposure 25%, value-chain position and pricing power 20%, technical barrier 15%, capacity scarcity 15%, financial quality 10%, valuation reasonableness 10%, future catalysts 5%. This model is a research framework, not factual disclosure.

Rank	Company	Total score	Explanation
1	TSMC	91	Directly controls front-end logic, HBM base die, and CoWoS, with advanced packaging still tight.
2	SK hynix	90	HBM leader + strongest financial realization + clear expansion.
3	Micron	88	Successful HBM4/HBM4E catch-up, with very strong margins and free cash flow.
4	Advantest	84	AI tester shadow bottleneck, high profit elasticity.
5	ASE	83	The clearest AI advanced packaging capture among OSATs.
6	BESI	82	HBM4 / Hybrid bonding tools are scarce, orders already validated.
7	Broadcom	80	Not in the packaging chain, but one of the strongest demand signals from cloud ASIC expansion.
8	Samsung Electronics	78	Strong full-stack capability, but HBM/packaging realization relatively lagging.
9	FormFactor	77	Probe cards have a hidden-bottleneck quality, but financial breakouts are limited.
10	Unimicron	75	Clear substrate-direction benefit, but cyclicality and valuation need balancing.
11	Ibiden	74	Same as above.
12	Marvell	73	Clear custom AI growth, but higher customer concentration and competition.
13	Amkor	71	Improving strategic position, but cash flow realization is back-loaded.
14	Applied Materials	70	Benefits through its hybrid bonding footprint, but not a pure AI packaging elasticity name.
15	JCET	63	Trackable long-term, but insufficient near-term evidence of high-end AI chains.

Detailed Table of Key Listed Companies

Company	Market	Sub-segment	AI benefit path	Key customers/supply chain	Latest public evidence	Expansion/CapEx	Valuation and expectations observation	Conclusion
TSMC	Taiwan stock / U.S. ADR	Foundry + CoWoS/SoIC	Directly monetizes GPU/ASIC logic, HBM base die, and CoWoS packaging	NVIDIA, AMD, Broadcom, cloud players	1Q26 revenue 35.9 billion USD, HPC 61% of mix, advanced packaging very tight, AI accelerator CAGR maintained in the high-50s range.	2026 CapEx pointing to the high end of 52 billion to 56 billion USD.	Valuation high but core-asset quality strongest.	Strong beneficiary, high certainty, not cheap.
SK hynix	Korean stock	HBM	HBM3E/HBM4 sold directly to AI platforms	NVIDIA a key customer, also serving the broader AI ecosystem	Q1 2026 revenue 52.6 trillion KRW, operating profit 37.6 trillion KRW, demand exceeding capacity; HBM4 volume-production readiness complete.	Announced a 19 trillion KRW advanced packaging plant investment in 2026.	Share price and market cap heavily re-rated, expectations already high.	Strong beneficiary, high certainty, watch for overheated valuation.
Micron	U.S. stock	HBM+DRAM+SSD	Full-stack benefit across HBM4/4E, data center DRAM/NAND/SSD	NVIDIA, CSP/OEM	FQ2'26 revenue 23.86 billion USD, non-GAAP gross margin 74.9%, HBM4 already shipped to Rubin, HBM4E in volume production in 2027.	FQ2'26 net CapEx 5 billion USD.	Very strong earnings elasticity, but partly reflected.	Strong beneficiary, high elasticity.
Samsung Electronics	Korean stock	HBM+foundry+packaging	Can benefit in an integrated way if external AI customers ramp	AMD has signed an AI memory cooperation MOU	HBM4 product page already shows a 4nm logic base-die roadmap; I-Cube/X-Cube/H-Cube solutions complete.	Expansion and customer qualification still need tracking.	The market has priced in a reversal, but realization needs validating.	Medium beneficiary, high potential, high execution risk.
ASE	Taiwan stock / U.S. ADR	High-end OSAT	Captures AI packaging and test plus TSMC overflow	SPIL a key supplier for NVIDIA AI chips	Expects advanced packaging business to double to 3.2 billion USD in 2026; Q4 revenue T$177.9 billion, net profit up 58%.	Continues to sharply raise equipment and facility CapEx in 2026.	Versus TSMC/HBM leaders, expectations may not be fully priced in.	Strong beneficiary, with an expectations gap.
Advantest	Japanese stock	Test equipment	Rising AI SoC/HBM test time and complexity	NVIDIA ecosystem, IDM, OSAT	Raised FY2026 operating profit guidance to 300 billion JPY on AI test demand.	Supply-chain improvement supports delivery.	High-quality leader, valuation usually not low.	Strong beneficiary, high certainty.
BESI	European stock	Hybrid bonding equipment	The most core equipment for HBM4/3D stacking	Leading memory makers, Asian foundry	Already won HBM4 hybrid bonding orders from two leading memory makers.	AI orders continue, Applied holds a 9% stake.	AI expectations clearly priced in.	Strong beneficiary, high elasticity, valuation hot.
Broadcom	U.S. stock	AI ASIC/interconnect	Cloud players' in-house XPU expansion pulls HBM and packaging demand	Google, Microsoft, Amazon, Meta, etc.	Expects 2026Q2 AI chip revenue of 10.7 billion USD, with AI chip sales exceeding 100 billion USD in 2027.	Driven by customer CapEx.	Strongest demand signal, but not a direct HBM/packaging seller.	Medium beneficiary, high certainty.
Marvell	U.S. stock	Custom AI silicon/interconnect	Trainium/custom compute and interconnect ramp	AWS and other hyperscalers	Custom AI silicon already in volume production; relationship with Amazon strengthening.	M&A to bolster AI fabric.	Large swings in market expectations.	Medium beneficiary, high risk and high elasticity.
Amkor	U.S. stock	OSAT / U.S. advanced packaging	Scarcity of domestic U.S. advanced packaging	Apple and NVIDIA as lead customers for the Arizona project	Arizona campus producing from 2028, investment expandable to 7 billion USD.	Capacity realization back-loaded.	Strategic value exceeds near-term results.	Medium beneficiary, long-term option value.
FormFactor	U.S. stock	Probe cards	HBM/high-end SoC testing raises probe requirements	Major memory and logic makers	The company's long-term positioning is high-end testing and probing; but no latest AI-breakout financials were found this cycle.	Needs continued validation.	May hold an expectations gap.	Medium beneficiary, needs ongoing validation.
Ibiden	Japanese stock	ABF substrate	AI large-package high-layer-count substrate demand	GPU/CPU/packaging houses	Clear public path, but no latest AI revenue breakout obtained this cycle.	Expansion and mix improvement need tracking.	Cyclical quality stronger than bottleneck quality.	Medium beneficiary, cyclical.
Shinko Electric	Japanese stock	ABF substrate	Same as above	GPU/ASIC platforms	Same as above.	Same as above.	Same as above.	Medium beneficiary, cyclical.
Unimicron	Taiwan stock	ABF substrate	High-compute server / AI substrate	Packaging houses / chip platforms	Same as above.	Same as above.	May hold an expectations gap, but needs validation.	Medium beneficiary.
AT&S	European stock	High-end substrate	HPC/AI high-layer-count substrate	Packaging houses / CPU/GPU platforms	Path holds, but AI packaging's separate weight in financials not fully disclosed.	Needs validation.	More of a cyclical reversal.	Medium beneficiary.
Applied Materials	U.S. stock	Equipment	Benefits via the hybrid bonding ecosystem	Strengthening coordination with BESI	Holds a 9% BESI stake, betting on hybrid bonding.	Broad equipment coverage, AI packaging is incremental.	Good company, but AI packaging revenue is not the mainline.	Indirect beneficiary.
ASMPT / Onto	HK stock / U.S. stock	Packaging equipment/inspection	AI packaging process control and inspection	OSAT, packaging houses	Onto's guidance is supported by advanced AI demand, but evidence this cycle is less than BESI.	Advanced packaging orders need tracking.	May hold an expectations gap.	Indirect beneficiary.
JCET	A/H	OSAT	Domestic advanced packaging upgrade	Domestic GPU/ASIC/SoC	Insufficient public evidence of high-end global AI chains.	Needs further validation.	Valuation reflects domestic-substitution expectations more.	May benefit long-term.
Shennan Circuits	A stock	High-speed boards / substrate-related	AI server and switch board-level demand	Servers/OEMs	Reasonable benefit, but not the HBM/CoWoS main chain.	Needs further validation.	More like AI infrastructure complement.	Indirect beneficiary.

Brief In-Depth Judgments on Key Listed Companies

TSMC: This is the most core "bottleneck asset" in the entire theme. It does not simply benefit from GPU shipments; it captures AI logic, HBM base die, CoWoS, and future 3D packaging upgrades all at once. What truly needs tracking is not total revenue but HPC mix, advanced packaging capacity, CoWoS lead times, and N3 allocation to HBM base die.

SK hynix: This is one of the strongest profit-elasticity names in the theme. It has upgraded from "AI narrative beneficiary" to "financial realizer," with the key being that HBM has restructured its profit base; but precisely because the market already fully recognizes this, research focus should shift to whether HBM4 share holds, whether packaging expansion can hold yield, and whether the post-2027 generation switch erodes margins.

Micron: The market underestimated its catch-up speed, but it can no longer be viewed through the old framework. Micron's early HBM4 volume shipment, HBM4E customization, 5-year SCA, and parallel data center SSD benefit show that its story is no longer just "playing catch-up," but "share gains + product structure upgrade + cash flow quality improvement."

ASE: If you can only pick one company among OSATs, its priority should clearly be above most peers, because it has already given an advanced packaging business scale target and is publicly named in connection with NVIDIA. The research focus is not "whether it can do advanced packaging," but the pace at which advanced packaging's share of total revenue/gross margin rises.

Advantest: Testing is the most easily underestimated "shovel" this cycle. Advanced AI chips cannot simply ship once packaged; ATE and system-level test time, binning, and reliability verification are all rising. Its drawback is that the market already knows this, so one should focus more on delivery pace and new tester types than on a generic "AI beneficiary" pitch.

BESI: This is one of the purest equipment-elasticity names in the HBM4/3D packaging upgrade. The deciding factor is the pace at which hybrid bonding moves from "tech demo" to "customer volume production." Reuters has already provided HBM4 order validation, so subsequent research should focus on two threads: the pace of orders expanding from memory to logic, and the timing of revenue recognition shifting from booking to shipment.

Samsung Electronics: Its difficulty is not a weak route but slow realization. Its technology and product routes are already in place, but on the investment side one cannot give a high score simply on "full-stack" as a label; the true inflection is external large-customer volume verification and recovering HBM share, not merely publishing a product page.

Broadcom / Marvell: These two are more like "demand weathervanes" in this theme than direct HBM/packaging names. Their importance lies in that, if custom AI keeps beating expectations, the demand base for upstream HBM, advanced packaging, and testing will broaden, no longer betting solely on NVIDIA.

Supplementary List of Unlisted and Private-Market Names

In this cycle's online verification, public disclosures from unlisted advanced packaging chain companies are generally less complete than those of listed companies, so the following list is included only as a direction for further research and not in this report's high-confidence conclusions: Absolics (glass substrate), Ayar Labs (optical interconnect / packaging co-design), Lightmatter (optical interconnect / AI interconnect), NHanced Semiconductors (U.S. advanced packaging OSAT), Tenstorrent (AI chip / system-level packaging co-design). These names are worth continued tracking, but their customers, funding rounds, valuations, and production pace need separate verification.

Risks, Open Questions, and Tracking Metrics

Risk Matrix

Risk	Transmission path	Most fragile link	Observation metric
Cloud AI CapEx slowdown	GPU/ASIC orders delayed, packaging and test scheduling compressed first	Second-tier packaging, substrate, materials	Broadcom AI revenue, cloud CapEx, TSMC AI guidance
GPU/ASIC shipments below expectations	HBM pull delayed	Legacy HBM3E generation, second-tier OSAT	NVIDIA/AMD platform delivery pace, TSMC HPC mix
HBM supply glut	ASP declines, margins fall back	Non-leading product lines at Samsung/Micron/SK hynix	HBM contract prices, DRAM/NAND bit supply plans
HBM4 yield below expectations	Affects volume ramp and customer qualification	HBM4, Hybrid bonding equipment	HBM4 sample-to-volume cycle
Overly fast CoWoS expansion	2027–2028 capacity pressure	Second-tier advanced packaging capture players	TSMC/ASE/Amkor expansion and utilization
ASICs replacing GPUs shift value allocation	More demand migrates to low-cost 2.5D/bridge	Pure GPU-bound narrative stocks	Broadcom/Marvell/custom XPU mix
Export controls/geopolitics	Regional supply-chain fragmentation	China-related chains and cross-region coordination	Changes in U.S. rules on advanced packaging/AI chips
Valuation bubble	Even good results struggle to move share prices	High-heat leaders	Degree of divergence between market cap and earnings growth

The Metrics Most Worth Tracking

First, watch TSMC HPC mix, CoWoS capacity, and advanced packaging lead times. If advanced packaging is still tighter than the front end, CoWoS remains the core bottleneck.

Second, watch the sample-to-volume speed of SK hynix and Micron's HBM4 / HBM4E, which determines who captures the most HBM4 profit.

Third, watch ASE's advanced packaging revenue share, the best metric for judging whether OSAT moves from "benefit narrative" to "financial realization."

Fourth, watch Advantest's AI tester shipments and orders and BESI's hybrid bonding shipments. Testing and bonding are often the last bottlenecks the market sees.

Fifth, watch Broadcom / Marvell's custom AI revenue. If cloud players' ASICs keep ramping, HBM and advanced packaging demand will keep broadening rather than depending only on NVIDIA.

Open Questions and Limitations

This report has three limitations that should be stated clearly. First, precise cross-market valuation metrics such as forward PE, EV/EBITDA, and FCF yield are not available with a consistent definition in public disclosures for all companies, so for some Japanese, Taiwanese, Korean, and A/H names, this report uses "valuation status / needs further validation" in place of precise multiples. Second, many materials, substrate, and OSAT companies do not separately disclose their AI advanced packaging revenue share, so their "degree of AI benefit" is more a research judgment based on order-transmission paths than a company's explicit statement. Third, for several China-based and unlisted companies on the user list, public materials provide insufficient evidence of whether they have entered the NVIDIA / high-end HBM production chain, so this report does not include them in the high-certainty core list.

Final Conclusions

HBM and advanced packaging have shifted from being AI chips' "accessories" to the "main battlefield." At this point in 2026, what truly merits research is not who "has an AI concept," but who possesses revenue elasticity, margin elasticity, order certainty, capacity scarcity, and long-term moats. If the whole chain is compressed into the five segments most worth watching, I would choose: HBM, CoWoS/3DFabric, HBM4/Hybrid bonding, test equipment/probe cards, and leading OSATs with genuine advanced packaging capture capability.

The ten listed companies most worth adding to the further-research list are: TSMC, SK hynix, Micron, ASE, Advantest, BESI, Samsung Electronics, Broadcom, FormFactor, and Unimicron. The first six lean toward "high certainty," and the latter four toward "an expectations gap but needing continued validation."

The three points most easily misunderstood by the market are: first, AI demand growth does not mean every packaging stock makes money; second, ASICs replacing GPUs does not mean HBM and advanced packaging demand declines, but more likely a diversification of the customer base; third, ABF / glass substrate / materials are important directions, but the strongest bottleneck today remains HBM, CoWoS, and high-end testing, not every "high-end substrate / new material" concept.

Over the next 6–12 months, what should be tracked first is not the share price but the data: HBM3E/HBM4 contract prices and the sample-to-volume pace, TSMC CoWoS capacity and lead times, TSMC N3 capacity allocation to HBM base die, ASE advanced packaging revenue growth, Advantest tester shipments, BESI hybrid bonding orders, Broadcom/Marvell custom AI revenue, and cloud-player CapEx guidance.

If I narrow the subsequent research directions further, I recommend prioritizing four topics: HBM4/HBM4E value-chain profit redistribution, the expansion pace of TSMC CoWoS and SoIC, the production realization of hybrid bonding equipment and materials, and whether test equipment / probe cards will upgrade from a shadow bottleneck to an explicit one. These four directions get closer to the investment mainline that can actually translate into earnings than a generic discussion of "AI packaging concepts."

This report is based on public information and does not constitute investment advice. Markets carry risk; invest with caution.

Mentioned Tickers

2330 MU 3711 6857 AVGO MRVL AMKR FORM 4062 6967 3037 AMAT 0522 ONTO INTC