A functioning chip is not yet a functioning computing system. It must receive power, communicate with memory and other processors, and release heat without damaging itself. Packaging makes those connections possible. As AI systems combine more silicon and memory in tighter spaces, the package increasingly determines what the designer can build and how much useful performance the customer receives.
The investment question
The packaging investment question is where increasing complexity creates durable value: with the foundry coordinating the system, the outsourced assembly specialist, the equipment supplier or the materials provider. Demand can grow for all of them while the strongest profits accrue to only a few. The key is control of a difficult, qualified production step and the ability to scale it economically.
Our view is that advanced packaging has become an architectural capability, but the category must be separated from conventional assembly. A supplier’s exposure to AI cannot be inferred from total packaging revenue. Product mix, customer qualification, equipment utilisation and the division of work with foundries determine the financial outcome.
How advanced packaging works
A conventional package protects a die and connects it to a circuit board. Advanced packaging can integrate multiple dies, memory stacks and dense interconnects into a single functional unit. In 2.5D integration, dies commonly sit beside one another on an interposer or another connection structure. In 3D integration, dies are stacked vertically with connections between layers.
The distinction matters because each approach trades density, communication distance, heat removal, cost and manufacturing difficulty. TSMC’s CoWoS platform integrates logic and memory using several interconnect options. Its SoIC technology addresses dense three-dimensional integration. These are different capabilities within a broader packaging portfolio, not interchangeable labels for all AI assembly.
Hybrid bonding connects surfaces through both dielectric bonding and metal connections, enabling very fine interconnect pitches. The process requires demanding surface preparation, alignment and cleanliness. Other approaches remain useful where their cost and performance fit the application. A transition towards denser integration does not mean every product immediately adopts the most advanced bonding method.
Market structure and competitive advantage
| Participant | Role in the package | Commercial test |
|---|---|---|
| Foundries | Integrate manufacturing and advanced packaging | End-to-end qualification and roadmap control |
| Outsourced assembly and test providers | Package, assemble and test customer devices | Execution, scale and differentiated mix |
| Equipment suppliers | Bond, place, process and inspect components | Precision, throughput and process wins |
| Substrate and materials suppliers | Provide physical and electrical foundations | Reliability, specification and available capacity |
TSMC, Intel and Samsung provide advanced integration capabilities connected to their manufacturing strategies. ASE and Amkor are major outsourced assembly and test participants. ASE’s VIPack platform illustrates how an independent provider can offer a portfolio of advanced integration technologies. Equipment suppliers such as Besi serve another part of the chain, earning revenue from the tools used to assemble products.
Competition often involves both partnership and overlap. A foundry may control important integration steps while working with outsourced providers elsewhere. Customers can divide work across several companies, but accountability becomes more complicated when a finished package fails. The supplier coordinating design rules, testing and yield improvement can gain commercial influence beyond the individual step it performs.
Economics: the value of a good package
Packaging economics depend on qualified units, complexity, materials, cycle time and yield. Larger or denser packages can command more revenue, but also consume more equipment time and expensive inputs. A supplier may need substantial capacity investment before the customer’s accelerator reaches volume. High demand therefore does not automatically produce strong free cash flow.
Yield has an especially important financial role because packaging can combine several valuable components. If assembly fails late, the loss may include already functional logic and memory. Testing individual dies before assembly reduces this risk, but cannot eliminate defects introduced during integration or problems that appear only when the combined system operates.
Consider an illustrative package containing components costing 100 before assembly. If final assembly yield improves from 90% to 95%, the component cost per good package falls from about 111 to 105, excluding recoveries and other costs. This is not an estimate for a particular product. It shows why customers may pay more for a process that improves final yield and reduces the loss of expensive components.
Capacity metrics also require care. Packages vary greatly in area, complexity and process route. A stated increase in wafer-equivalent capacity or tool count does not translate directly into an equal increase in finished accelerator output. Ask what product mix, utilisation and yield assumptions sit behind the capacity figure.
AI and hyperscalers: packaging shapes the machine
AI accelerators require close connections between logic and high-bandwidth memory. As systems demand more memory bandwidth and larger integrated designs, the package becomes a constraint on performance, power delivery and heat removal. Adding memory stacks can change the required package dimensions and integration complexity, with consequences throughout the supply chain.
Hyperscalers influence these requirements through their own accelerator designs and their purchases from merchant suppliers. The same cloud operator may support several architectures, each requiring distinct packaging qualification. A shift towards custom silicon can therefore expand the number of programmes without immediately freeing the shared production resources those programmes need.
The linkage also runs downstream. A denser package can raise power density at the server level, requiring changes in cooling and electrical infrastructure. Better package interconnects can reduce data-movement energy, but the final benefit depends on the complete system. Packaging, compute, power and cooling should be analysed as related engineering decisions rather than separate demand forecasts added together without adjustment.
Current market debates — September 2026
The current debate is whether advanced-packaging capacity catches up with demand and how much of the expansion reaches independent providers. Amkor’s second-quarter 2026 results describe progress in AI and high-performance computing programmes alongside capacity expansion. This supports an opportunity beyond foundry-owned operations, but the value depends on the specific work awarded and the timing of production.
The equipment signal is also relevant. Besi’s July 2026 results identify hybrid bonding, photonics and data-centre applications among growth drivers. Orders for assembly equipment indicate investment plans; they should not be treated as proof that all resulting capacity is already qualified or fully utilised.
The constructive case is that larger AI packages and additional architectures absorb capacity as it arrives. The countercase is that multiple suppliers expand against overlapping forecasts, while improved yields increase effective supply. Watch qualification milestones and customer production ramps rather than relying on headline construction announcements or assuming every packaging shortage is equally persistent.
Structural debates: open chiplets and integration
Open chiplet interfaces could allow customers to combine building blocks from more suppliers. The UCIe consortium’s specifications address die-to-die connectivity and related management capabilities. Standardising an interface helps interoperability, but does not automatically solve thermal design, product validation, commercial liability or the need for compatible operating characteristics.
The investment implication is nuanced. An open ecosystem could broaden the design market and benefit independent integration providers. It could also increase the value of a trusted coordinator that makes diverse components work together. Technical openness does not necessarily imply fragmented profits.
A second debate concerns the economics of scaling package size. Larger packages can integrate more capability but face physical, thermal and manufacturing constraints. Alternatives such as bridges, fan-out and different substrate approaches compete on total cost and reliability. The winner may vary by application rather than emerging as one universal architecture.
What to watch
Track qualified capacity for named product classes, advanced-packaging revenue mix, final yields, customer concentration and cash investment. Distinguish tool orders from productive capacity and compare package growth with the accelerator and memory roadmaps that drive it.
The strongest packaging business should earn attractive returns by making complex systems manufacturable and reliable. Its advantage becomes most visible when customers launch difficult new products and choose the same partner again for the following generation.
Explore this sector
Semiconductors & chipmaking — sector overview
Related sectors: AI infrastructure & data centres · Data platforms & analytics
