Advanced Packaging Is Becoming as Important as the Chip

September 16, 2026

Altsets

Research by Altsets Research

Share

AI accelerators are turning advanced packaging from a back-end manufacturing step into a strategic constraint, shifting economic importance toward CoWoS capacity, substrates, HBM integration, testing, bonding equipment, and specialized packaging materials.

Data used:Altsets Supply Chain Intelligence: 90k+ entities, 400k+ relationships, 20+ years of history.

Key findings

  • TSMC is moving CoWoS toward packages larger than 14 times reticle size as AI package complexity increases.
  • ASE announced a new advanced AI packaging hub in Kaohsiung expected to add more than 113,000 square meters of factory space.
  • Altsets estimates Nvidia represents 27.88% of SK Hynix revenue while SK Hynix represents 27.33% of Nvidia COGS.
  • Altsets estimates Nvidia represents 17.62% of Micron revenue while Micron represents 14.00% of Nvidia COGS.
  • The advanced packaging opportunity extends into substrates, packaging materials, bonding equipment, and test.
  • Increasing value per package raises the economic importance of yield, inspection, testing, and precision assembly.

The next AI chip bottleneck is increasingly not the transistor. It is the package around it.

Nvidia's Vera Rubin platform entered full production in 2026 just as TSMC is pushing CoWoS toward much larger package sizes and ASE is adding new advanced packaging capacity for AI and high-performance computing. TSMC says CoWoS remains the majority of its advanced packaging activity, while its roadmap is moving toward packages larger than 14 times reticle size and exploring new substrate technologies to support them.[1][2] The investment implication is that the value chain around an AI accelerator is widening. TSMC and ASE matter, but so do substrate producers, testing companies, bonding equipment vendors, and materials suppliers such as Ajinomoto, Resonac, Advantest, and Besi.

This is not simply more semiconductor spending. It is a change in where manufacturing difficulty sits.

The bottleneck is shifting from making the die to reliably integrating logic, HBM, substrates, interconnects, and materials into one high-value package.

27.88%
Nvidia share of SK Hynix revenue
Altsets Supplier Revenue %
27.33%
SK Hynix share of Nvidia COGS
Altsets Customer Cost %
17.62%
Nvidia share of Micron revenue
Altsets Supplier Revenue %
14.00%
Micron share of Nvidia COGS
Altsets Customer Cost %
14x+
Reticle-size package roadmap
TSMC is moving toward packages larger than 14 times reticle size
113,000 m²+
Planned ASE AI packaging hub factory space
Kaohsiung project announced with WUS Printed Circuit

The package is becoming part of the architecture

A conventional processor could be treated largely as a chip mounted onto a substrate. An advanced AI accelerator is closer to a system assembled inside a package. A large GPU or accelerator die must communicate with several stacks of high-bandwidth memory through extremely dense interconnects, while power delivery, thermal performance, signal integrity, warpage, and yield all have to remain within acceptable limits.

CoWoS is one solution to that problem. TSMC places multiple dies and HBM around an interposer or redistribution structure and then connects the assembly to a package substrate. Its CoWoS family now includes silicon-interposer, redistribution-layer, and local-silicon-interconnect approaches, while package dimensions continue to grow.[1]

Exhibit 1

The AI accelerator package is becoming a system

Logic, HBM, interconnects, substrates, and materials must work as one manufacturing stack

  1. 01
    Logic die
    Large GPU or accelerator die provides the core compute.
  2. 02
    HBM stacks
    High-bandwidth memory sits close to the logic and adds memory capacity and bandwidth.
  3. 03
    Interposer and interconnect
    Dense wiring, redistribution, bumps, or bonding connect logic and memory.
  4. 04
    Package substrate
    The assembled package connects to power and the system board through a large advanced substrate.
  5. 05
    Materials and test
    Underfill, molding, thermal materials, inspection, and test protect yield and reliability.

This is a simplified structural diagram. It does not represent every package architecture or imply that each platform uses the same materials or interconnect technology.

Source: Altsets research using TSMC disclosures

That changes the economic structure of semiconductor manufacturing. Adding another HBM stack does not only create demand for more memory. It can require more interposer area, a larger substrate, additional micro-bumps or bonding steps, more underfill and molding material, more difficult thermal management, and more expensive testing. A larger package also increases the cost of failure late in the manufacturing process because several already valuable components have been combined before final validation.

Packaging therefore becomes a multiplier. More expensive logic and more HBM create more demand for the infrastructure needed to connect, protect, test, and assemble them.

Altsets relationship data shows how large the memory leg of this system has already become. Nvidia represents 27.88% of SK Hynix revenue, while SK Hynix represents 27.33% of Nvidia COGS. Nvidia also represents 17.62% of Micron revenue, while Micron represents 14.00% of Nvidia COGS. These relationships are not product-specific and do not prove that every dollar reflects HBM, but they show that the GPU and memory supply chains are economically intertwined before the packaging layer is even considered.

The important shift is that advanced packaging physically joins those previously distinct supply chains.

The beneficiaries extend beyond TSMC

TSMC is the obvious exposure because CoWoS sits directly inside its manufacturing stack. ASE Technology is another. In May, ASE announced a new advanced AI packaging hub in Kaohsiung with WUS Printed Circuit, specifically citing growing demand for chip stacking, chiplet integration, CoWoS, and FOCoS. The project is expected to add more than 113,000 square meters of factory space.[2]

But the more interesting investment map sits one layer below the packaging houses.

Package substrates are becoming larger and more technically demanding as AI packages add more compute and HBM. Public companies such as Ibiden, Unimicron Technology, and Shinko Electric participate in the high-end substrate ecosystem. Ajinomoto occupies a less obvious position through Ajinomoto Build-up Film, the insulating material used within advanced package substrates. The company is continuing to develop ABF for increasingly advanced semiconductor packages and photonic-electronic co-packages.[3]

Materials intensity rises elsewhere too. Resonac sells a portfolio that reaches across the back end of advanced semiconductor manufacturing, including nonconductive film for HBM, liquid underfill, molding compounds, thermal interface materials, solder resist, and materials used in packaging substrates. These are small components relative to the value of a finished AI accelerator, but that is exactly why the economics are interesting. A low-cost material can still become strategically important if package yield or reliability depends on it.

Equipment is another second-order exposure. Besi sells semiconductor assembly equipment and has explicitly connected improving demand to 2.5D AI data-center applications and expanding hybrid-bonding use cases in logic, memory, and co-packaged optics.[4] Hybrid bonding becomes more relevant as manufacturers try to connect dies at tighter pitches than traditional bump-based approaches allow.

Testing also becomes more valuable as the package becomes more complex. Advantest is developing test systems specifically for increasingly complex AI, HPC, chiplet, memory, and silicon-photonics devices. The reason is straightforward: combining expensive logic, HBM, interposers, and substrates raises the value at risk inside each completed package. Finding defective components before final assembly and validating the finished system become economically more important.

Exhibit 3

The packaging beneficiaries extend below the packaging house

Selected companies named in the article, grouped by their role in the stack

LayerSelected companiesRole described in the article
Packaging capacityTSMC, ASE TechnologyCoWoS and outsourced advanced packaging capacity
Package substratesIbiden, Unimicron Technology, Shinko ElectricHigh-end substrates for larger and more complex packages
Substrate materialsAjinomotoAjinomoto Build-up Film used within advanced package substrates
Packaging materialsResonacNonconductive film, underfill, molding compounds, thermal materials, solder resist, and substrate materials
Bonding equipmentBesiAssembly equipment and hybrid-bonding exposure
TestingAdvantestTest systems for AI, HPC, chiplet, memory, and silicon-photonics devices

These companies occupy different parts of the packaging workflow and should not be treated as interchangeable exposures.

Source: Altsets research using company disclosures

The second-order bottleneck is yield

Investors often frame advanced packaging capacity as a simple volume constraint: Nvidia wants more accelerators, therefore TSMC needs more CoWoS lines. The deeper constraint is yield across an increasingly complicated assembly.

Every additional die, HBM stack, interconnect layer, substrate connection, and bonding step creates another place where manufacturing variability can destroy the value of the finished package. That makes packaging materials, inspection, testing, and precision assembly equipment part of the same bottleneck.

Exhibit 4

Package complexity turns yield into a second-order bottleneck

More integrated value means more value at risk when a late manufacturing step fails

Complexity sourceWhat becomes harderSecond-order exposure
More HBM stacksMemory integration and interconnect densityPackaging capacity, bonding, testing, and underfill
Larger package areaWarpage, routing, substrate design, and mechanical stabilitySubstrate producers and substrate materials
Tighter interconnect pitchAssembly precision and connection reliabilityHybrid-bonding and precision assembly equipment
Higher value per assembled packageLate-stage defects become more expensiveInspection and test systems
Higher power densityThermal and mechanical reliabilityThermal interface and encapsulation materials

The table describes the mechanism discussed in the article. It does not assign failure probabilities or forecast company margins.

Source: Altsets research

It also produces different investment exposures. TSMC benefits from selling scarce packaging capacity. ASE benefits when customers need additional outsourced assembly and test capacity. Substrate manufacturers benefit from larger and more difficult packages. Ajinomoto and Resonac participate through materials whose importance grows with package complexity. Besi participates as interconnect pitches shrink and bonding technology changes. Advantest participates because more expensive heterogeneous packages demand more sophisticated test flows.

The common factor is not simply AI demand. It is increasing value per package combined with increasing difficulty per package.

Conclusion

Advanced packaging is becoming as strategically important as the chip because modern AI processors are no longer economically complete when the wafer leaves the fab. Their performance depends on integrating logic, HBM, interposers, substrates, thermal materials, and thousands of high-density connections into one reliable system.

That shifts part of the semiconductor bottleneck from lithography and wafer capacity into the back end. CoWoS capacity is the most visible constraint, but the broader investment exposure includes ASE, substrate manufacturers, Ajinomoto and Resonac in materials, Besi in bonding equipment, and Advantest in testing.

As AI packages become larger and more heterogeneous, the companies that make the pieces work together can gain economic importance even when they never manufacture the GPU itself.

Sources

  1. TSMC, "Q2 2026 Earnings Conference Transcript," July 16, 2026, https://investor.tsmc.com/english/encrypt/files/encrypt_file/reports/2026-07/57b65edbfe6e480e74abe202be983ecbde79e934/TSMC%202Q26%20Transcript.pdf

  2. ASE Technology Holding, "ASE and WUS Announce Strategic Collaboration to Build Advanced AI Packaging Hub in Kaohsiung," May 8, 2026, https://www.aseglobal.com/press-room/ase-and-wus-announce-strategic-expansion

  3. Ajinomoto Co., "Message from the President and CEO 2026," 2026, https://www.ajinomoto.com/sustainability/ir/ceo_2026.php

  4. BE Semiconductor Industries, "Besi Increases Long-Term Financial Targets at 2026 Investor Day," June 18, 2026, https://www.besi.com/investor-relations/press-releases/details/be-semiconductor-industries-nv-increases-long-term-financial-targets-at-2026-investor-day/

How to Cite This

According to Altsets Supply Chain Intelligence (altsets.com), Nvidia represents 27.88% of SK Hynix revenue while SK Hynix represents 27.33% of Nvidia COGS, illustrating how the memory and accelerator supply chains are already economically intertwined before advanced packaging joins them physically.

For research inquiries or data access: press@altsets.com

Go Deeper

Altsets Research

Get new Altsets research when we publish.

Original supply-chain research, market casefiles, and new data findings.

Research updates only. Unsubscribe any time.

Methodology

Read the methodology for this research.