Unimicron Technology

Phase A — Understand the business

Company Overview

Unimicron is the world's largest printed-circuit-board (PCB) maker by revenue and one of the top-tier makers of IC substrates — the fine-pitch organic carriers (FC-BGA / ABF substrates, FC-CSP, wire-bond BGA) that sit between a packaged silicon die and the motherboard. It is the substrate/PCB arm that was spun out of United Microelectronics (UMC), which remains a strategic shareholder (see Lens 9). FY2025 revenue was NT$131.24B (~US$4.25B), +13.75% YoY.

What it sells (three buckets). Sources conflict on the exact mix — flag this:

• One profile: ~60% PCB / ~30% IC substrate / ~10% other components.
• A 2024 breakdown: substrates 61% / PCB + HDI 38% / flex + RF 1%. These are not reconcilable as stated; the truth is that IC substrate is now the larger and far more valuable half, and the "60% PCB" figure looks like a stale group-level split. Treat the mix as roughly half substrate / half PCB, with substrate carrying the margin and the AI story.

The thing that matters: Unimicron is a top-3 global ABF (Ajinomoto Build-up Film) substrate supplier. ABF substrate is the organic carrier for high-performance logic — CPUs, GPUs, AI accelerators, ASICs and networking switch silicon. As AI packages grow in die size and layer count, ABF substrate has become a named chokepoint in the AI-compute backend, adjacent to (and gating alongside) TSMC's CoWoSCoWoSTSMC’s method of packing a processor and its memory onto one carrier so they sit close together. Supply of it has been a hard limit on how many AI chips can be built. advanced packaging.

Customers. Fabless logic and their OSAT/foundry packaging partners. Management states Unimicron is the "main supplier for 4 major US cloud-service providers" and that ASIC-related orders are growing "exponentially". Named end-demand drivers: Nvidia (GPU), AMD, Broadcom & Marvell (custom ASIC / networking), and hyperscaler custom silicon (Google TPU, Amazon Trainium, Meta MTIA) — all consumers of high-layer ABF. our figures is empty, so specific revenue concentration by customer is n/a — not disclosed / not sourced.

Contract structure. Substrate is sold on POs against qualified designs, not take-or-pay. In an up-cycle the leverage inverts to the supplier: high-end ABF lead times ran >30 weeks in 2025 and the industry is sold out into 2026-27, letting Unimicron raise substrate prices to pass through Ajinomoto's film hikes and the fibreglass shortage.

Supply Chain

Map (upstream → Unimicron → end buyer), named at every node:

Upstream raw materials (Unimicron's own dependencies):

• Ajinomoto Fine-Techno — the ABF dielectric film itself. Controls ~95–98% of CPU/GPU-grade ABF film; a functional monopoly with 18–24-month qualification for any substitute — an insurmountable switching cost that Ajinomoto used to hike prices ~30% into the AI upcycle. This is Unimicron's single most important upstream chokepoint.
• Fibreglass / glass-cloth (E-glass, low-DK/low-CTE grades) — in structural shortage through 2026, forcing substrate price increases; suppliers include Nan Ya, Taiwan Glass, Nittobo.
• Copper foil, prepreg/laminate (Nan Ya Plastics et al.), dry film, plating chemistry, laser-drill & lamination equipment.

Unimicron (substrate + PCB fab) — plants in Taiwan (Guangfu, Yangmei), China (Kunshan, Suzhou), and an expanding Thailand PCB site.

Downstream:

• → Advanced packaging / OSAT & foundry: TSMC (CoWoS), ASE, Amkor — the substrate is integrated with the interposer/die here. Note Unimicron pivoted its Yangmei line from EMIB toward CoWoS to chase utilization.
• → Fabless logic: Nvidia, AMD, Broadcom, Marvell, and hyperscaler custom-silicon teams.
• → End buyers: the 4 major US CSPs (Microsoft, Amazon, Google, Meta) and the AI-server ODMs.

Chokepoints: Unimicron is itself a chokepoint (top-5 ABF makers ≈ 74% of the market) but is squeezed from above by Ajinomoto's film monopoly and the fibreglass crunch — its input costs are set by parties with more pricing power than it has. Ground: frames the equipment→foundry→packaging→hyperscaler flow; ABF substrate is the layer that wiki still under-weights (listed only as "glass-substrate packaging" on the watch list).

Competitive Advantages (moats)

Real, but narrower than the tape implies.

• Scale + breadth (genuine moat). Largest PCB maker globally and the broadest ABF customer base; the most comprehensive layout across high-end ABF, server substrates and AI-GPU substrates among the merchant players. Capital intensity (record NT$34B 2026 capex) and multi-year qualification cycles are a barrier to new entrants.
• Qualification lock-in (moat). Substrate designs are co-developed and qualified per package; requalifying a second source costs quarters. Once in a GPU/ASIC socket, incumbency is sticky.
• The hole in the moat (critical). In the highest-value tier — 14+ layer FC-BGA for GPU-class packages — Unimicron has NOT demonstrated volume yield; it and Nan Ya are "price-competitive at lower layer counts" but confined to the commodity/mid-layer segments, while Ibiden's edge is yield maturity at the layer counts that matter. High-end FC-BGA share: Ibiden ~35%, Shinko ~18%, Unimicron ~14%, AT&S ~10%, Nan Ya ~5%; some sources put Ibiden at 70–80% of the AI-specific substrate market. So the moat is strongest where the dollars are weakest, and weakest where they are richest.

Bargaining power: strong over customers today (sold out, >30-week lead times, raising prices), weak over suppliers (Ajinomoto/fibreglass dictate input cost). That asymmetry only holds while the cycle is tight — it reverses hard in a glut (see 2023-24, Lens 8).

Segments

our figures is empty → all figures.

By product (revenue mix): ~half IC substrate / ~half PCB+HDI (mix conflict flagged in Lens 1). Substrate is the margin and growth engine; within it, ABF for AI applications was ~40% of substrate revenue in 2025, guided to ~60% in 2026. Within PCB, AI-related revenue rises from 55–60% (2025) to 65–70% (2026). Blended AI-related revenue is guided to exceed 60% of the company in 2026, up from ~30% in 2025 — the single clearest number in the file: this is now an AI-infrastructure supplier, not a diversified PCB house.

By geography: China ~40%, USA ~25%, Taiwan ~15%, rest ROW.

Trend & cause: Revenue accelerating (FY24 +11% → FY25 +14% → Q1'26 +24.5% YoY) as the AI-substrate mix climbs and the trough-year commodity PC/handset drag fades. The mix shift is the thesis: AI moves from ~10% of the ABF end-market (2015) toward ~75% by 2030.

---

Phase B — Measure performance

Earnings Result (latest print + the cyclical backbone)

The 5-year cycle in one table:

This is a deep cyclical: 2022 was a once-in-a-cycle peak (GM 35.9%, EPS 20.08), then two brutal years (EPS collapsed −83% to 3.34 in 2024) before the current recovery. Conflict: one source reports FY2025 net income at NT$7.55B (+36%) vs stockanalysis's NT$6.67B — the NT$6.67B ties to EPS 4.38 × 1,523M shares and is used as primary; the NT$7.55B is likely total consolidated income incl. minorities vs attributable-to-parent.

The inflection — Q1 2026 (quarter ended 31-Mar-2026):

• Revenue NT$37,446M, +24.5% YoY, +7.9% QoQ.
• Net income NT$5,381M, +482% YoY, +37% QoQ — a single quarter nearly matched all of FY2025 (NT$6.67B).
• EPS NT$3.28 (vs NT$0.60 a year earlier).
• Gross margin 18.0% (vs 13.4% YoY, 15.8% prior Q) — 460bps YoY expansion; operating margin ~7%.
• +65% EPS surprise, +2.7% revenue surprise vs consensus.
• Confirming the H2-loaded shape: 4Q25 profit alone surpassed the first three quarters of 2025 combined.

Guidance/outlook: management targets record 2026 revenue exceeding the 2022 peak (NT$140.5B); H2 2026 > H1; utilization >90%; AI mix >60%. BofA raised 2026 operating-profit estimates +6% and 2027 +27% on stronger ABF margins. Balance-sheet flags: net debt/EBITDA <2x; total assets NT$231.4B, total liabilities NT$67.7B (equity ~NT$163.7B) — modest leverage, but a convertible bond is outstanding and Unimicron ran a capital increase (dilutionDilutionIssuing new shares, so each existing share owns a smaller slice of the same company.) with UMC participating (Lens 9/10).

Market reaction: the stock has multi-bagged (~3–4x in a year) on the substrate upcycle — it now trades as a high-beta AI-backend derivative (Lens 8).

Earnings Calls (sentiment trend)

transcripts/ is empty →, drawn from monthly-revenue commentary, results coverage and analyst notes across ~4 quarters:

• Tone has shifted from "recovery/uncertainty" (early 2025) to "structural sold-out upcycle" (mid-2026). Early-2025 language leaned on tariff uncertainty and CoWoS-client pivots; by 2026 it is capacity, sell-out, price increases, record capex, "three-year super-cycle."
• Recurring phrases (2026): "AI-related revenue >60%", "H2 stronger than H1", "record revenue/capex", "ASIC orders growing exponentially", "raising substrate prices."
• What they stopped saying: the 2023-24 refrains of "inventory correction", "weak PC/handset demand", "under-utilization." The commodity-drag narrative has been replaced by an allocation narrative.
• Read: management sentiment is confidently bullish and demand-constrained — the tell to watch is whether the language ever slips back toward "digesting capacity" as the NT$34B build lands (that would be the top-signal).

Comps

Merchant ABF/substrate peers. Multiples are with source/date or n/a; never fabricated.

The comps punchline: Unimicron trades at a large premium to Ibiden (~200x vs ~27x trailing; ~58x vs Ibiden-cheaper on forward) despite Ibiden being the superior high-end/AI-yield asset. The market is paying up for Unimicron's breadth, ASIC exposure and earnings-acceleration, not for high-end leadership. Two of the five majors (Shinko private, AT&S subsidized) are effectively off the public risk/reward board — supply is more concentrated than the headline "5 players / 74%" suggests.

Stock-Price Catalysts (5-yr, >5% moves)

Pattern from the price history:

• 2021–2022 melt-up → 2022 peak on the first substrate shortage (COVID-era compute + auto), EPS to NT$20.08.
• 2022 H2 → 2024 crash as the shortage flipped to a glut (PC/handset inventory correction) — EPS −83%, the stock de-rated for ~2 years. This is the single most important risk lesson in the name.
• 2025 → 2026 multi-bag (~3–4x) on the AI-substrate upcycle. Specific catalysts that moved it:
• Morgan Stanley Overweight, TP raised to NT$1,225, "golden decade," 105% 2025-28 EPS CAGR, ABF supply gap widening to ~22%.
• BofA upgrade to Buy on ABF-margin uplift.
• Monthly revenue prints (Taiwan discloses monthly): e.g., Jan-2026 NT$12.767B, +34.5% YoY.
• Ajinomoto +30% film hike and fibreglass shortage — read as scarcity-validation, not cost-fear.
• UMC capital injection (Dec-2025) and capex raises (NT$19.4B→25.4B→34B).
• What the market actually reacts to: AI-capex sentiment (Nvidia/hyperscaler capex), substrate supply-gap datapoints, monthly revenue momentum, and analyst upcycle calls. It is a leveraged, high-beta proxy on the AI backend — it will overshoot in both directions.

---

Phase C — Judge people & books

Management

• Governance transition at the inflection (flag). Long-time chairman Tzyy-Jang (T.J.) Tseng retired / was dismissed on 13-Feb-2026 when corporate director UMC reappointed its representative; S.C. Chien became acting chairman. A leadership change precisely as the super-cycle and record capex land is a governance watch-item — continuity of the capacity strategy is now on a new team.
• UMC parentage / skin in the game. Unimicron is a UMC-affiliated company (spun from UMC's PCB arm); UMC remains a strategic shareholder and bought 6.03M shares for NT$700M in Dec-2025 and participated in the capital increase. That is a strategic holder adding at the top of the cycle — a confidence signal, but also the related-party vector (Lens 10).
• Capital-allocation history. Textbook cyclical: reinvests aggressively into up-cycles (record NT$34B 2026 capex, 65% substrate / 35% PCB), funds it with a mix of internal cash, convertible bond and equity (dilution), pays a thin dividend (NT$2.00, ~0.2% yield). ROE is cycle-distorted — spectacular in 2022, single-digit in 2024. The open question is return discipline: building record capacity into peak pricing is exactly what set up the 2023-24 collapse.
• Archetype: professional-manager / corporate-parent (UMC) governance, not founder-led. Implication: competent operators executing an industry upcycle, but with the classic cyclical-capex temptation and no founder to override consensus at the top.

Forensic Red Flags

our figures/filings empty → income-statement forensics are; no 10-KForm 10-KA company’s audited annual report to the US regulator. The most complete thing it publishes. Item 3 exists (no SEC filer).

• Cyclicality is the headline "red flag," not fraud: gross margin has swung 35.9% → 13.9% in three years. Reported earnings quality is hostage to price/utilization, both of which the customer and Ajinomoto influence more than Unimicron does.
• Net-income definition conflict (NT$6.67B vs NT$7.55B for FY2025) — parent vs consolidated/minority interest. Not a manipulation flag, but a reminder to standardize on attributable-to-parent (EPS-consistent NT$6.67B).
• Dilution / capital structure: convertible bond (conversion price trimmed) + capital increase during the run-up mean share count and the conversion overhang can dilute the per-share upside the multiple is capitalizing. Watch share-count creep vs the 1,523M base.
• Capex-vs-cash-flow divergence: record NT$34B capex against still-recovering FCFFree cash flowCash left after paying to run and maintain the business. Unlike profit, it is hard to flatter with accounting choices. means the build is partly financed (debt/convert/equity) — the balance sheet is healthy now (net debt/EBITDA <2x) but the cushion thins if the cycle rolls before the capacity pays back.
• Inventory/receivables: AI orders "doubled," utilization >90% — demand-pull, no visible channel-stuffing signal; but this is, unaudited-by-me.

Regulatory findings (required):

• SEC EDGAR (LR, AAER): none possible — Unimicron has no CIK and is not an SEC filer. Zero SEC findings.
• Non-SEC enforcement (web): the only material hit is a 2018 environmental fine of CNY 380,000 (~US$56K) against the Kunshan (China) subsidiary for local environmental-regulation breach — immaterial. No FTC/DOJ/antitrust actions against Unimicron. Distinguish from the parent: UMC (not Unimicron) pleaded guilty in a 2020 US DOJ trade-secret case (Micron/Fujian Jinhua); relevant to parent-governance perception only.
• 10-K Item 3: n/a — no SEC filer.
• Verdict: No material regulatory or legal findings against Unimicron — verified via SEC EDGAR EFTS (LR, AAER: 0), web enforcement search (one immaterial 2018 China environmental fine), and the no-CIK limitation, as of 2026-07-10.

---

Phase D — Project & stress-test

Forward Projection (FY2026–FY2028 EPS)

Bottom-up from FY2025 actual (EPS 4.38) + Q1'26 actual (EPS 3.28, GM 18.0%) + guidance (record revenue, H2>H1, AI mix >60%). All outputs; inputs labeled.

FY2026 (base): Q1 3.28 actual; assume QoQ growth ~8–12% (H2>H1 guide) → Q2 ~3.6, Q3 ~4.1, Q4 ~4.6 → EPS ≈ NT$15.5. Revenue ~NT$160B (+22%), GM ~18–19%.

• Bear ≈ NT$11 (momentum stalls, ASIC ramps slip, GM gives back to ~16%).
• Bull ≈ NT$18 (full sell-out pricing, GM to 20%+, high-end mix improves).

FY2027 (base) ≈ NT$22; bull ~NT$28, bear ~NT$13. FY2028 (base) ≈ NT$30; Morgan Stanley's 105% 2025-28 CAGR implies EPS ≈ NT$38, the bull anchor.

Valuation read at NT$875: base fwd P/E ~58x (2026), ~40x (2027), ~29x (2028); even on MS's bull NT$38 2028 EPS the stock is ~23x 2028 — i.e., the tape already discounts ~2 years of hyper-growth as near-certain. That is the entire risk/reward: you are paying a growth multiple on top of peak-of-cycle earnings acceleration.

(No our model logged — unattended watchlist mode.)

Bull vs Bear

Bull case. ABF substrate is a structural, multi-year chokepoint in the AI backend: AI moves from ~10% of the ABF end-market (2015) to ~75% by 2030; the top-5 makers hold ~74% and two of them (Shinko, AT&S) are off the public board; lead times >30 weeks; Ajinomoto's +30% film hike and the fibreglass crunch validate scarcity and hand Unimicron pricing power. Unimicron has the broadest AI-substrate layout and is the main supplier to 4 US CSPs with exponential ASIC order growth — perfectly placed for the custom-silicon/networking wave (Broadcom, Marvell, TPU/Trainium/MTIA) even if Nvidia flagship GPU substrate skews to Ibiden. Earnings are inflecting (+482% Q1), margins expanding (13.4%→18.0% GM YoY), H2>H1, record capex funding the next leg. Morgan Stanley "golden decade," BofA Buy.

Bear case (permanent-impairment candidates).

1. Cyclical over-build. The industry is adding record capacity (Unimicron NT$34B; Ibiden ¥500B / 2.5× capacity by 2028) into peak pricing — the exact setup that produced the 2023-24 glut and an 83% EPS collapse. When AI-backend capex digests, a price-taker with half-commodity revenue de-rates violently.
2. High-end yield gap. Unimicron is #3 in high-end FC-BGA (~14%) and unproven at 14+ layer GPU-class yield; if the value migrates to the hardest packages, Ibiden/Shinko capture the richest dollars and Unimicron is left defending mid-layer/ASIC where competition (Nan Ya, Korean players) is fiercer.
3. Input squeeze. Ajinomoto (95%+ film) and the fibreglass oligopoly can claw back margin the moment demand softens — Unimicron's supplier bargaining power is weak.

Pre-mortem (18 months out, thesis broke): AI-server capex paused for a digestion quarter, hyperscaler ASIC ramps slipped two quarters, mid-layer ABF pricing rolled over as the 2026-27 capacity waves (Unimicron + Ibiden + Nan Ya) all landed at once, and the NT$34B build turned into under-utilized fixed cost — the ~58x forward multiple compressed to ~20x on lower-than-hoped 2027 EPS. A −50%+ drawdown from NT$875 is entirely inside this name's own recent history.

Multiples too high? On trough/normalized earnings, yes — ~200x trailing, ~58x forward-2026 for a deep cyclical is pricing perfection. Justifiable only if the "golden decade" (no glut, sustained high-teens/20% GM, mix to high-end) is real and durable.

Contrarian view (what the market refuses to see): The Street is buying Unimicron as the Nvidia-GPU-substrate AI play. It is really the ASIC/networking-substrate play — its GPU-flagship position is #3 behind Ibiden. The bull thesis is more right about the demand vector (ASIC/custom silicon is where hyperscaler volume and Unimicron's wins actually are) and more wrong about the entry price than consensus frames it.

Devil's Advocate (short-seller)

• What structurally breaks the money machine: a return to over-supply. This is a commodity-adjacent, price-taking cyclical dressed as a secular AI winner; the last upcycle (2022) ended in a two-year, −83% earnings depression. Every major is expanding at once into the same 2027-28 window.
• Revenue concentration: disclosed customer concentration is opaque (our figures empty), but "4 US CSPs" + AI ASIC means exposure to a handful of hyperscaler capex budgets — a single hyperscaler digestion quarter (or an in-housing/second-sourcing decision) hits hard.
• Weak moat where it counts: unproven at 14+ layer GPU-class yield — the premium tier is Ibiden's. Unimicron's moat is scale/breadth, not top-end technology leadership.
• Most dangerous underestimated competitor: Ibiden (yield leader, cheaper stock, ¥500B expansion) — and, longer-dated, glass substrate (Intel/Samsung/AT&S and a 15-name supply chain) as an architectural successor that could strand organic-ABF capacity built in 2026-27.
• Worst capital-allocation risk: debt/convert/equity-funded record capex + dilution at the top of the cycle — the pro-cyclical build that history punishes.
• Assumptions the NT$875 price requires: no glut through 2028, GM sustained high-teens+, AI mix keeps climbing, high-end yield closes, and the multiple doesn't compress. If 2027 growth disappoints 20–30%, base EPS drops toward ~NT$15-17 and a de-rating to ~25-30x forward implies a NT$400-500 stock (−40% to −55%).
• Single permanent-impairment scenario (plausibility): a 2027-28 organic-ABF glut coincident with faster-than-expected glass-substrate adoption at the high end — moderate probability, high severity. More likely than bulls admit precisely because everyone is building.

Management Questions (ordered by information value)

1. What is your demonstrated volume yield at 14+ layers for GPU-class FC-BGA today, and what is the concrete timeline/capex to close the gap to Ibiden?
2. Give the AI-revenue split between GPU vs ASIC/networking — where are your actual design wins, and with which customers?
3. What customer concentration sits behind "4 US CSPs" — top-1 and top-3 as % of revenue?
4. How much of the record NT$34B 2026 capex is committed/irreversible, and what utilization do you need to earn your cost of capital on it?
5. What is your downside plan if AI-backend capex digests for two quarters in 2027 — how fast can capex/opex flex?
6. How do you protect margin against Ajinomoto's film pricing and the fibreglass shortage — any second-source or long-term-agreement leverage?
7. What is the convertible-bond conversion overhang and total expected dilution through 2027?
8. Post-Tseng, what changed (if anything) in the capacity/allocation strategy under the new chairman and UMC's representative?
9. What is UMC's intended role — passive strategic holder, or a path to deeper integration/further capital injections?
10. Where does glass substrate fit your 3–5-year roadmap — are you investing, and does it strand any 2026-27 organic-ABF capacity?
11. What pricing structure (LTAs? take-or-pay? capacity reservations?) locks in the current sell-out beyond spot POs?
12. How do you see mid-layer ABF pricing in 2027-28 as your capacity + Ibiden's + Nan Ya's all come online together?
13. What is the Thailand PCB expansion economics and its exposure to tariff/geographic-diversification demand?
14. What normalized through-cycle gross margin should investors underwrite — 2022's 36% is not it; is it high-teens?
15. What is your capital-return philosophy at the top of a cycle — why a 0.2% dividend rather than a buyback if the shares are undervalued on your own forward view?

---