JPMorgan: ABF Substrate Utilization to Exceed 100% by 2027, Four AI Supply Chains Under Simultaneous Strain
nashnova research
JPMorgan forecasts global ABF substrate utilization will breach 100% in 2027, with orders exceeding all available capacity; each new generation of AI chips is physically scaling by more than ten times, putting substrates, high-speed CCL, MLCC, and test equipment under simultaneous pressure — a supply-demand mismatch now reshaping pricing power and margins across the chain.
Demand up ~40% a year, supply up ~10% — where does the gap come from?
JPMorgan's ABF demand model, measured by chip area times layer count, projects global demand at ~11,308 in 2025, ~21,150 in 2027, and ~29,565 in 2028 (billion sq mm), with year-on-year growth of 34% → 39% → 40%. Nearly all incremental demand comes from AI servers and switches, rising from 26% of total demand in 2024 to 73% by 2028.
Supply growth is far more restrained: 9.3%, 10.2%, and 12.1% year-on-year for 2026–2028. Unimicron and Ibiden alone account for over half of new capacity through 2028; all other makers are adding single-digit increments.
This means → the utilization curve steepens sharply: ~65% in 2024 → ~72% in 2025 → ~86% in 2026 → past 100% in 2027 → ~130% in 2028. Crossing 100% means orders exceed total capacity — some demand must be delayed or squeezed out.
Why does every new AI chip generation amplify substrate demand?
From Nvidia's Hopper to Feynman, total substrate area per chip (body area × layer count) rises from 38,280 sq mm to 432,000 sq mm — a more than tenfold increase.
In plain terms = chips are getting larger and stacking more layers, so every single chip "consumes" exponentially more substrate area — this is the physical root cause behind the supply-chain crunch.
Equipment data confirms the trajectory: global demand for substrate stepper lithography tools fell to ~40% of 2023 levels in 2025. JPMorgan expects a rebound from 2026, but new tools won't yield effective capacity until 2027 or later — lining up exactly with the 100% utilization crossover.
T-glass fiber allocation — who gets the material gets the high-end orders?
AI substrates require low-thermal-expansion T-glass fiber cloth — a specialized glass-fiber material — with Nitto Boseki as the primary supplier. JPMorgan estimates 35%–40% goes to Unimicron, 30%–35% to Ibiden, and all remaining makers split the last quarter to third.
This means → second-tier makers such as Nan Ya PCB and Kinsus may have production capacity but cannot run at full utilization on high-end products if they lack T-glass fiber — material allocation, not capacity, is the binding constraint on who wins large-format, high-layer AI orders.
The shortage is also reshaping contract structures. JPMorgan ranks five types of long-term agreements; "capacity reservation fees" score highest — customers pay upfront to lock future capacity without tying the payment to specific orders, giving makers both cash flow and pricing flexibility. In plain terms = substrate makers' revenue visibility is stretching from quarter-by-quarter to year-by-year.
How much can pricing and margins rise?
JPMorgan estimates next-generation AI accelerator ABF unit prices can easily exceed the prior generation by over 50%, driven by higher layer counts, larger area, and more complex packaging. Indexed to Unimicron's 2018 unit-area price at 1, the price index is projected at ~3.6 in 2027, surpassing the prior cycle peak of ~3.3 in 2022.
With pricing and utilization rising simultaneously, margins jump: Unimicron's ABF segment margin is forecast to climb from ~28% in Q1 2025 to ~62% by Q4 2027; company-wide net profit is expected to grow 357% in 2026 and another 132% in 2027, with gross margin rising from 13.9% in 2025 to a projected 43.0% in 2027.
This reflects a rare "volume-and-price" supercycle in substrates — not pricing alone, not shipment volume alone, but both compounding in the same window, producing margin leverage well beyond what linear extrapolation would suggest.
High-speed CCL — how large is this parallel supply-chain strain?
JPMorgan estimates the HPC high-speed CCL — copper-clad laminate, the base material for printed circuit boards — market at $4.25 billion in 2025 → $7.6 billion in 2026 → $12.77 billion in 2027, a 73% CAGR. Within that, M8-grade CCL grows at a 124% CAGR, its share rising from 46% in 2025 to 77% by 2027 — nearly all market growth is captured by M8.
Per-chip CCL value jumps with each generation: Hopper ~$45, Blackwell ~$210, Rubin ~$410, Rubin Ultra with backplane ~$580. At the rack level the spread is even wider: an H100/H200 HGX rack runs ~$4,000 in CCL, while a Vera Rubin Ultra rack with backplane exceeds $200,000 — the backplane alone multiplies per-rack CCL value by more than four times.
Layer count is another multiplier: a 100G switch PCB runs ~20–24 layers, while a 1.6T switch runs ~40–48 layers. In plain terms = doubling layers doubles the CCL sheets consumed per board, while lamination difficulty and scrap rates also rise — cost does not simply scale at 2×.
Where are the ceiling risks for this super-cycle?
JPMorgan explicitly flags two counter-arguments. First, glass core: the industry is advancing multi-layer thick-core organic CCL solutions; if this path succeeds, glass-core adoption may be pushed beyond 2030, a positive signal for incumbent makers like Unimicron.
Second, CoWoP packaging — a structure that removes the substrate entirely, placing the interposer directly on a CoWoP board connected to the motherboard via a socket — meaning some chips would no longer consume ABF substrates. JPMorgan provides no penetration timeline but lists it as a structural risk. This means → the 2028 ABF demand figure of 29,565 is not a certainty — technology-path evolution is the upper-bound constraint on the demand model.
Compared to earlier ABF reports from Goldman Sachs and Morgan Stanley, JPMorgan adds three new quantified dimensions: T-glass fiber allocation ratios, capacity-reservation-fee contract structures, and Unimicron's quarterly ABF margin path to ~62% by Q4 2027. The direction of the supply-demand gap is now confirmed by multiple houses; the real debate is: when will alternative packaging technologies like CoWoP penetrate — and that will determine how long this substrate super-cycle lasts.
市场有风险，内容仅供研究参考，不构成投资建议。