Which specific AI packaging or substrate process will remain supply-constrained through 2028?

Research note · 28 September 2026

A substrate supplier can double its process capacity and still make the same number of packages. Ibiden has described precisely that pressure: larger, more multilayer GPU substrates can require twice the semi-additive processing capacity for an unchanged unit count. That is the starting point for deciding which AI packaging constraint survives through 2028. [1]

Our strongest candidate is customer-qualified semi-additive processing, or SAP, for large, high-layer-count ABF organic substrates. Specifically, we mean the repeated copper build-up and fine-line patterning used to manufacture the carrier beneath the interposer and chips. Our confidence that this remains constrained through late calendar 2028 is moderate. The manufacturing mechanism is persuasive; the timing of relief depends on qualification, yields and demand. [2][3]

The organizing idea is simple: the work required per package is growing alongside the number of packages. Factory expansion must cover both. Large-format CoWoS-L redistribution-layer and local-bridge processing is our second candidate; qualified low-thermal-expansion glass cloth is a credible upstream risk, with a less secure case for scarcity lasting through 2028.

The substrate grows faster than the package count

An advanced AI package needs several kinds of wiring. The interposer connects the compute dies and high-bandwidth memory; the organic substrate carries the assembled package and provides its connection onward. CoWoS-L uses redistribution layers and local silicon interconnects, while the underlying substrate contains its own multilayer routing. More interposer capacity therefore leaves a separate manufacturing requirement beneath it. [4][3]

SAP builds and patterns copper wiring through repeated processing steps. Larger footprints consume more processing area, and additional build-up layers require more work on that area. Ibiden’s explanation is unusually useful because it connects a product-generation change directly to factory load: the same number of newer GPU substrates can consume twice the SAP capacity. Its FY2025 results materials also anticipate total SAP demand exceeding industry supply as substrates become larger and more multilayered. [1][2]

The geometry makes that pressure tangible. TechSearch gives a Blackwell substrate example of 70 × 76 mm, with a 7-2-7 construction, and describes future AI substrates of 100 × 100 mm or 120 × 120 mm. Multiplying the dimensions gives future footprints of 1.88–2.71 times the Blackwell example. These are future design cases; their significance is the increase in area before adding any extra layers or yield losses. [3]

We think this makes qualified output at the required footprint and layer count the useful measure of capacity. Counting pieces obscures the transition. A production line delivering smaller substrates cannot be credited with the same output of the largest AI carriers simply because its equipment is installed. The evidence selects SAP as the process family under pressure; it does not identify one universal machine-level bottleneck within it. [2][1]

A doubling of capacity can support surprisingly little unit growth

The next question is whether expansion can outrun that increase in work. Ibiden has announced approximately ¥500 billion of investment in high-performance IC package substrates over FY2026–FY2028. Its separate project notice describes approximately ¥220 billion with sequential operation and mass production planned from FY2027. Those are substantial additions, but the investment timetable extends across the period in which new package designs are raising the manufacturing requirement. [5][6]

Our break-even test gives the supply response considerable room: assume qualified SAP process capacity doubles between 2026 and late 2028. If average work per package rises by 50–100%, that expansion supports only 0–15.5% annual package-unit growth, with yields unchanged. The upper workload case is informed by Ibiden’s generation-change commentary; the lower case allows for a partial transition across the product mix. [1]

Demand could readily test that threshold. A quoted UBS estimate puts NVIDIA GPU units at 8.8 million in 2026 and 10.8 million in 2027, an increase of 22.7%. In a stress case that repeats that growth for a second year and uses it as a proxy for the target substrate market, a 50–100% increase in work per package would leave 11.5–33.6% of process demand unmet, even after capacity doubles. That range describes the stated scenario; its purpose is to show how much the answer depends on product mix and demand growth. [7][1]

Our conviction rests on this interaction, rather than a precise industry deficit. The supply response can be large and still struggle to catch up. Equally, slower unit growth, a gentler transition to larger substrates or better yields can clear the market.

CoWoS-L is the strongest alternative, but expansion must be judged by geometry

There is direct evidence that packaging already limits shipments. On TSMC’s July 2026 earnings call, C.C. Wei said: “our packaging capacity is so tight that now it limits my customers’ growth.” That gives the current shortage a clear commercial consequence. [8]

CoWoS-L also has a concrete demand case. TSMC says its first 3.5-reticle format has been in volume production since 2024. AMD specifies CoWoS-L for its MI455X offering, alongside 12 HBM4 stacks. More compute and memory must be interconnected within a larger assembly, raising the importance of large-format redistribution layers and local bridges. [4][9]

The supply response is aggressive. A September industry report relayed a forecast for aggregate CoWoS capacity to reach roughly 260,000 wafers per month by end-2028, double its cited end-2026 level. The estimate comes through several reporting layers, and the aggregate figure leaves the mix of CoWoS variants and qualified output unresolved. [10]

Geometry could still consume much of that growth. Using a separate 2026 estimate of 140,000 wafers per month, the rise to 260,000 implies 86% more nominal capacity. Under a hypothetical change in average interposer size from 5.5 to eight reticles, with process share, packing and yields constant, package throughput would rise only about 28%: 260/140 multiplied by 5.5/8. That is why we retain CoWoS-L as the strongest rival to the substrate call. [11][10]

We nevertheless prefer SAP as the durable constraint because Ibiden ties increasing workload directly to its substrate manufacturing process. TrendForce expects CoWoS-L to remain the mainstream large-AI-chip packaging solution through 2028, but technology leadership and capacity scarcity are different propositions. Interposer expansion—including a move toward panel processing—still leaves the need for a compatible finished organic carrier. [1][12][4][13]

Glass cloth could move the constraint upstream

The substrate thesis raises a further question: can manufacturers obtain the materials needed to use their new process capacity? Nittobo identifies significantly rising demand for both thick and thin low-thermal-expansion T-glass used in semiconductor package substrates. It has also announced approximately ¥15 billion of glass-cloth investment, with production scheduled to begin in fiscal-2026 Q4. Demand and a meaningful supply response are both visible. [14][15]

We regard qualified low-CTE glass cloth as a credible upstream risk, but we are less confident that it remains the binding constraint through late 2028. The decisive comparison is material-specific: customer-qualified cloth deliveries against the core constructions that substrate makers actually require. Applying substrate-area growth to glass demand and assuming a qualification ramp produces too wide a range of outcomes to select glass ahead of SAP.

The discriminating observation would be substrate factories with available SAP capacity that cannot operate because qualified cloth is missing. That would move the constraint upstream and strengthen the investment case around Nittobo’s capacity. Until then, we give greater weight to Ibiden’s direct account of the processing load.

HBM is another potential restriction, but our evidence does not select hybrid bonding as a universal packaging bottleneck through 2028. Micron expected an additional facility to contribute meaningfully to HBM packaging capacity from the first half of calendar 2027. Product-specific stacking and qualification developments matter more here than a broad assertion that memory remains scarce. [16]

Ibiden offers the clearest process exposure; returns still depend on execution

For investors, Ibiden is the most direct exposure supported by this investigation: its manufacturing commentary identifies the rising SAP load, and its ¥500 billion investment plan commits capital to high-performance substrates for AI and other high-performance servers. The opportunity and the execution burden sit together. [1][5]

Nittobo represents the upstream glass hypothesis. TSMC represents interposer and assembly capacity whose conversion into finished shipments still requires compatible carriers. NVIDIA, AMD and custom-ASIC customers bear the consequences of allocation, delays and potentially higher package costs. These exposures sit at different points in the same production chain. [15][14][4][3][9]

We would not turn that process ranking into a stock ranking. Our work does not establish comparable substrate-specific margins, incremental returns or valuations across Ibiden, Unimicron, AT&S and their peers. Scarcity creates an opportunity to earn more, but large capital commitments and yield-learning costs can absorb it. The investment question is how much qualified good output each new unit of capital ultimately delivers.