← research bank

EMIB Shifts the Packaging Bottleneck to Substrates

emib-substrate-shift · conviction medium · status playing-out · horizon 2027 · as of 2026-07-22

EMIB adoption moves the AI-packaging bottleneck from CoWoS interposers to high-end FC-BGA substrates by 2027, re-rating the substrate cluster (Ibiden, Shinko, Unimicron, AT&S, LG Innotek) while eroding the pure CoWoS-scarcity premium at the margin.
Rests on filed figures, not on modelled shares. 17 premises (5 entity, 4 signal, 3 field, 3 edge, 2 weight); no derived cell is involved, so the undisclosed supply weights that put a range on other pages in this bank cannot move this one.

Exhibits

Exhibit 1Relative performance, indexed to 100How the names in this thesis have traded against SOXX.
36621320ATS.VI 7394062.T 572011070.KS 353SOXX 225TSM 17612mo, indexed to 100 at start · dashed = SOXX benchmark

Series available as data/emib-substrate-shift.csv

Exhibit 2Who pays ABF substrate and build-up film supply, and who keeps the moneyCapturers average 10.8% operating margin against payers' 37.9% — the owners of the scarce thing earn LESS than the names it constrains.
IBIDEN Co., Ltd.14.0%Ajinomoto (ABF)12.6%Unimicron Technology Corp.5.9%NVIDIA Corporation64.0%Advanced Micro Devices11.8%

Green/blue = model marks it as CAPTURING the rent (unbound and supplies the scarce good); faded = PAYING it (bound severe or moderate). Operating margin, live.

Exhibit 3What the conviction is actually made ofEach premise and the number it composes to. A conjunction of plausible premises is far weaker than any of them.
The AI-packaging bottleneck is real and TSMC-concentrated todayCoWoS (Chip-on-Wa… — supply constraint seve… †85.0%Taiwan Semiconductor Manufacturing Company…80.0%COMPOSED (and)80.0%

† 1 premise marked supporting — shown and arguable, but the conclusion does not depend on them, so they are not multiplied into the composed figure. Citing a filed figure should not cost conviction.

One gating premise, so the conclusion is exactly as strong as it. The figure is an ordering device, not a calibrated probability — see how the numbers are made.

Weakest link: Taiwan Semiconductor Manufacturing Company — 90% supply share of CoWoS (Chip-on-Wafer-on-Substrate) at 0.80 — A 0.9 supplier weight concentrates the scarcity rent at one vendor, which is the condition the variant says gets competed around. The weight is a desk

EMIB is a credible second path, not vaporwareEMIB (Embedded Multi-die Interconnect Bridg…85.0%EMIB (Embedded Multi-die Interconnect Bridg…100.0%EMIB (Embedded Multi-die… — maturity ramping †65.0%SK Hynix — signal 2026-07-2170.0%COMPOSED (and)59.5%

† 1 premise marked supporting — shown and arguable, but the conclusion does not depend on them, so they are not multiplied into the composed figure. Citing a filed figure should not cost conviction.

60% if the 3 gates are independent, 70% if they move together. They are claims about one industry, so the truth is between and nobody can say where. Treat this as an ordering device rather than a calibrated probability — the ranking of premises is the information, not the level.

Weakest link: SK Hynix — signal 2026-07-21 at 0.70 — The dominant HBM maker testing HBM-on-EMIB means the path plugs into the main demand artery. Held at 0.70: SK Hynix discloses no production timeline a

If EMIB scales, the scarce step migrates to FC-BGA substratesEMIB (Embedded… — supply constraint moderate †70.0%LG Innotek supplies IBIDEN Co., Ltd.100.0%Advanced Packaging Bottl… — signal 2026-07-…75.0%Emib Substrate Shift — signal 2026-08-0375.0%Elite Material Co., Ltd. (EMC)55.0%COMPOSED (and)30.9%

† 1 premise marked supporting — shown and arguable, but the conclusion does not depend on them, so they are not multiplied into the composed figure. Citing a filed figure should not cost conviction.

31% if the 4 gates are independent, 55% if they move together. They are claims about one industry, so the truth is between and nobody can say where. Treat this as an ordering device rather than a calibrated probability — the ranking of premises is the information, not the level.

Weakest link: Elite Material Co., Ltd. (EMC) at 0.55 — EMC is the CCL layer the thesis runs through, and it carries NO financials — every field is null because the entity was built from a channel note. It

The bottleneck migration pulls a PASSIVES layer with it, and that layer is supply-constrained independentlyMultilayer ceramic capacitor (high-capacita…70.0%Murata Manufacturing Co., Ltd.85.0%Silicon capacitor (in-package decoupling)60.0%COMPOSED (and)35.7%

36% if the 3 gates are independent, 60% if they move together. They are claims about one industry, so the truth is between and nobody can say where. Treat this as an ordering device rather than a calibrated probability — the ranking of premises is the information, not the level.

Weakest link: Silicon capacitor (in-package decoupling) at 0.60 — In-package silicon capacitors take the power-integrity role MLCC cannot fill at accelerator current density, with Powerchip named as exclusive supplie

HBM integration gives the migration a demand anchorSK Hynix — 62% supply share of High-Bandwid…75.0%SK Hynix supplies NVIDIA Corporation100.0%SK Hynix — signal 2026-07-2170.0%COMPOSED (and)52.5%

52% if the 3 gates are independent, 70% if they move together. They are claims about one industry, so the truth is between and nobody can say where. Treat this as an ordering device rather than a calibrated probability — the ranking of premises is the information, not the level.

Weakest link: SK Hynix — signal 2026-07-21 at 0.70 — Same signal as the credibility leg, cited again here for the DEMAND anchor rather than for path viability. NOTE: reusing one signal across two conclus

The variant

Consensus

CoWoS capacity — TSMC ~90% of 2.5D AI packaging — is the single binding constraint on accelerator shipments through at least late 2027; TSMC's expansion (120k-140k wpm in 2026, +60% by 2027, two Arizona packaging fabs in a $265B US buildout) is the read on AI supply, and substrate makers are secondary beneficiaries of a TSMC-led cycle.

Variant

Intel's EMIB scales as a real second path (~90% yields, Google TPU v8e 2H27, Meta CPU 2H28, customers prepaying substrate suppliers, SK Hynix testing HBM-on-EMIB); because EMIB replaces the silicon interposer with bridges in a large FC-BGA substrate, the scarce step migrates to high-end ABF substrate supply — already at full utilization with the film input repricing +30% — and the substrate cluster becomes the levered play while the pure CoWoS-scarcity narrative erodes at the margin.

Differentiator

Consensus watches interposer wafers-per-month; the variant watches substrate qualification, prepayments, and ABF pricing. The tell the market is missing: Intel's customers are prepaying SUBSTRATE suppliers, not bridge capacity — the constrained input in the EMIB path is the substrate, and every EMIB win converts interposer scarcity into substrate scarcity.

Stance
Bottleneck migration over interposer scarcity: every EMIB win converts CoWoS interposer demand into FC-BGA substrate demand, so own the substrate incumbents (Ibiden first) rather than the TSMC-scarcity narrative — hedged for the 2027 CoWoS capacity landing and the 2027-28 substrate capex wave.

Recommendations

Second order

Third order

Indicators

Tripwires

Falsifiers

Open questions

Reasoning chain

The AI-packaging bottleneck is real and TSMC-concentrated today VALID
premises

A severe-constraint supply step held ~90% by one vendor concentrates scarcity rent at TSMC — the condition the variant says gets competed around.

EMIB is a credible second path, not vaporware VALID
premises

A ramping Intel-owned packaging path with an HBM co-development signal from the dominant HBM maker means the alternative is being engineered, not just discussed. QUALIFIED 2026-08-04 BY INTEL'S OWN PRODUCT. Clearwater Forest uses TWELVE EMIB bridges embedded in organic substrate to join three Foveros base dies, for 29 total silicon pieces in one package. Each bridge is process steps and yield loss, so at that count the cost advantage EMIB is supposed to hold over a silicon interposer largely evaporates — a point made in the same source that documents the design. THE THESIS ASSUMES EMIB IS THE CHEAPER PATH. If Intel needs twelve bridges to make 18A economic, EMIB is credible as ENGINEERING and unproven as ECONOMICS, and those are different claims. What would settle it: a disclosed cost-per-package comparison against CoWoS at equivalent die count, or Intel winning external EMIB business on price rather than on capacity access.

If EMIB scales, the scarce step migrates to FC-BGA substrates VALID
premises

EMIB replaces the interposer but consumes high-end substrate; a small competing supplier set (incumbents + one entrant) is where the queue re-forms — scarcity moves to the slowest-scaling input. The 2026-08-03 channel evidence upgrades this from FORECAST to OBSERVATION: accelerator packages are reported waiting on substrate and a named TPU program slipped one to two quarters specifically on substrate yield, and a delay attributable to substrate rather than to logic or interposer is the cleanest available evidence the scarce step has already moved. Held back from a larger re-rating by two details in the SAME source: the announced capacity is REPURPOSABLE (optional, not committed), and the CCL price increases are largest in the LOWER grades — a capacity-reallocation signal, not a demand signal.

The bottleneck migration pulls a PASSIVES layer with it, and that layer is supply-constrained independently VALID
premises

The thesis has been argued at the substrate layer alone. Packaging change pulls a component layer with it, and that layer has its own constraint: board-level bulk capacitance at 30-week lead times AND a new in-package category ramping at +68.8% YoY. CRITICALLY THIS IS EXPANSION, NOT SUBSTITUTION — silicon capacitors take the package role while MLCC keeps the board role, so advanced packaging grows the total capacitor opportunity rather than moving it, and neither name is disrupting the other. That makes Murata a supply-constrained incumbent rather than a contested one, which is a BETTER position than the substrate cluster this thesis is built on occupies. WHAT WOULD CHANGE IT: the 800V distribution shift in the 3DHI roadmap moves conversion closer to the die and moves capacitance into the package with it. That is the mechanism by which complementary becomes substitutive, and it is the thing to watch rather than current share.

HBM integration gives the migration a demand anchor VALID
premises

The #1 HBM supplier to the #1 accelerator maker validating HBM-on-EMIB means the alternative path plugs into the main demand artery rather than a niche.

Sources

Write-up

Pre-filled skeleton: emib-substrate-shift.md