Morgan Stanley Raises ABF Substrate Supply-Demand Gap Forecast to 9.3% by 2027, Predicting Price Hikes Exceeding 25%
nashnova research
Morgan Stanley raised its 2027 ABF substrate supply-demand gap from 6.2% to 9.3%, forecasting price increases above 25% — but explicitly below the buy side's ~50% expectation; the demand revision is driven not by Nvidia GPUs but by "other ASICs" including Microsoft custom chips and Chinese AI silicon.
A second upgrade in three months — how big is the gap now?
On October 14, Morgan Stanley updated its ABF substrate model — a high-end circuit board that carries AI chips. The 2027 supply-demand gap widened from 6.2% to 9.3%; the 2028 gap from 9.4% to 13.7%.
This is the second upward revision since the model launched in July. The driver is demand, not any loosening on the supply side.
This means → Morgan Stanley's own numbers keep getting tighter, reinforcing — not softening — the undersupply thesis.
Why isn't Nvidia the star of this demand upgrade?
The revision centers on "other ASICs": custom chips from Microsoft and peers, plus Chinese AI silicon from Huawei, Cambricon, Hygon, and T-Head. The 2027 forecast for this category jumped from 4.6 million to 6.4 million units — the single largest variable in the entire model.
Nvidia GPU assumptions are unchanged at 7.2–8.1 million units across 2026–2030. AMD GPUs were cut. Google TPU and Amazon Trainium saw only minor tweaks.
In plain terms = the incremental engine for substrate demand has shifted from "Nvidia alone" to "a crowd of companies building their own AI chips" — each one needs substrates, and the total adds up.
Over the long term, AI-related applications rose from ~10% of ABF substrate market value in 2015 to ~60% in 2025, projected above 80% by 2030. AI ASICs alone climb from 9% in 2025 to 30% in 2030 — the largest single growth pocket.
$14.3 billion in expansion — why isn't it enough?
In the past four months, expansion announcements piled up: AT&S committed €1.5–2 billion in Malaysia with AMD and Marvell; Samsung Electro-Mechanics announced ₩6.78 trillion across Sejong and Vietnam; Nan Ya PCB approved NT$46.8 billion for a new plant; Unimicron acquired Bridgestone's Taiwan site for ~NT$10.1 billion. Morgan Stanley totals these at roughly $14.3 billion.
Per its supply-chain research, each $1 invested supports $0.50–0.60 in annual revenue at full ramp — so $14.3 billion translates to $7.2–8.6 billion in incremental annual revenue. But the market grows from $9.5 billion in 2026 to $20.9 billion in 2030 — an increment of $11.4 billion, exceeding the capacity ceiling of announced investments.
This means → the money is real, but output cannot keep up with demand growth. The gap is structural, not a funding problem.
How many suppliers can actually ship new capacity in 2027?
After a plant-by-plant review, Morgan Stanley identifies only two facilities ramping in 2027: Unimicron's Guangfu Phase 2 and Kinsus (景碩) k6b. Kinsus k6b adds roughly 30% capacity.
Nan Ya PCB gains 5–10% in 2027 via bottleneck relief; its new plant arrives in H2 2029. Samsung Electro-Mechanics Sejong starts volume production in September 2028. Toppan's Singapore fab just opened and will contribute large-format, high-layer substrates in 2027, but Morgan Stanley does not quantify it separately.
Larger package sizes, higher layer counts, embedded passives, and Intel's EMIB-T all depress usable output — nominal capacity up X% means effective output well below X%.
In plain terms = factories are being built, but the boards themselves are getting harder and bigger, so each production line delivers fewer finished units than before.
A 25% price hike below buy-side expectations — who bears the risk?
Morgan Stanley forecasts 2027 ABF substrate prices rising >25%, but states explicitly this is below the buy side's ~50% expectation. In 2026, raw-material cost pass-through drives pricing; in 2027, supply tightness replaces cost pass-through as the dominant pricing force.
Under Morgan Stanley's assumptions, Unimicron's IC-substrate gross margin rises from 15.2% in 2025 to 30.8% in 2026; Nan Ya PCB's ABF margin from 17.3% to 33.7% — each up ~16 percentage points.
Morgan Stanley reiterates Overweight on both, lifting targets to NT$1,775 (Unimicron) and NT$2,000 (Nan Ya PCB), with Unimicron preferred. Its 2027 EPS estimate for Nan Ya PCB is NT$48.25, below consensus at NT$51.55.
This means → if the 25% call proves right, the real earnings-miss risk sits with buy-side positions priced for 50% hikes — and with Nan Ya PCB itself.
Can this cycle deliver — and where is the uncertainty?
Chinese AI chip shipments face policy, advanced-node capacity, and HBM supply constraints — the hardest part of the entire demand forecast to verify. If this segment underdelivers, the 2027 gap narrows materially but does not fall back to the old model's 6.2%.
Customer-funded expansion has appeared: Samsung Electro-Mechanics' investment includes client financing and medium-term volume guarantees; AT&S's Kulim project is fully backed by long-term commitments from AMD and Marvell. This reflects a supply squeeze severe enough for customers to pre-pay and lock volume.
But Morgan Stanley also warns: the last cycle's version of this model ended with capacity running ahead of demand. The 2030 gap widens to 25.9% under the new model — even counting every announced expansion, the industry still needs more capacity. That is the central uncertainty of this cycle.
市场有风险，内容仅供研究参考，不构成投资建议。