Gold Circuit Electronics Ltd. (2368.TW): Deep Dive

Reporting-period anchor: GCE reports quarterly and its fiscal year ends in December. Under Taiwan rules interim results are due within 45 days of quarter-end, so Q2 2026 results were due by about 14 August 2026. They were released, and management presented them at a Fubon Securities investor conference on 28 August 2026. Q3 2026 results are not due until about 14 November 2026, so Q2 2026 is the latest period covered.

1. What the Company Does

Gold Circuit Electronics (GCE, 金像電子) makes printed circuit boards: the flat, layered slabs of glass-fibre laminate and copper that every electronic component plugs into. Its boards are not the thin, cheap ones inside a toaster or a phone charger. They are thick, dense, many-layered boards that sit at the centre of data-centre servers and network switches. Many of them now go into the AI accelerator servers that Amazon, Google, Meta and Microsoft design for their own clouds.

A server motherboard or accelerator baseboard is a city map drawn in copper, stacked dozens of layers high. Every chip on it (processors, AI accelerators, memory, network interfaces, power stages) talks to every other chip through copper traces buried inside the board. As chips get faster, those traces have to carry signals at higher speeds with less loss, which forces more layers, thinner lines, exotic low-loss laminates and tighter manufacturing tolerances. GCE's business is building those boards in volume, on time, at yields that let it earn money. Its own description of its capability is multilayer boards of 50+ layers, high-density interconnect (HDI) and heavy-copper boards, used in servers, networking, AI, automotive and IC test applications (Fugle memo of the 28 Nov 2025 investor conference).

History. Per the company's own milestone page (gce.com.tw), GCE was founded in September 1981 in Taoyuan with NT$10 million of capital to produce PCBs. It obtained US UL certification in 1982, finished developing multilayer boards in 1986, listed on the Taiwan Stock Exchange in March 1998, and was producing 16-layer boards by 1999. Then came the moves that shaped today's footprint:

• Construction of a Suzhou plant in mainland China began in 2000.
• A third Zhongli (Taoyuan) plant was completed in 2001, the same year GCE gained TL-9000 telecom certification.
• In 2010 it took full ownership of the Changshu facility.
• In 2025 a plant in Prachinburi, Thailand started production.

The sources reviewed do not name the founders, so this report does not.

The company's identity changed mainly through its mix. For most of the 2000s and 2010s GCE was known as Taiwan's largest maker of networking PCBs (MoneyDJ company profile), with a big notebook-board business alongside. Servers were about half of revenue in 2020 (poorstock conference summary). By Q2 2026 they were 78%, with networking at 15% and notebooks down to 5%. The pivot was not a single bold decision. It came from being the qualified high-layer supplier when the hyperscalers began designing their own servers, and then their own AI chips.

The problem it solves. A cloud company that designs a custom AI accelerator needs someone who can turn its board design into tens of thousands of identical, defect-free, electrically consistent boards per month, in materials the supply chain barely makes enough of, and who can be trusted with the next generation before it launches. Only a handful of factories worldwide can do that at the layer counts and speeds involved. GCE's offer is that combination of capability, capacity and a track record that has already passed the customer's qualification.

Why it is hard. A 30-50 layer board is laminated from many thin cores and prepreg sheets in several press cycles. Each layer must line up with the others to within microns, or the drilled holes miss their pads. Holes are drilled at extreme aspect ratios, plated with copper, and often back-drilled to cut off the unused stub that would otherwise degrade high-speed signals. A single defect anywhere in that sequence scraps a board whose material alone is expensive, so yield is the business. Section 3 covers the detail.

A concrete example (illustrative of how the work flows). Take the custom accelerator board for a Trainium server. Its design is set by the chip and system owner, here AWS. GCE was confirmed as a Trainium mainboard supplier when Amazon took an equity stake in September 2026 (Digitimes, 30 Sept 2026). A new generation goes through GCE's engineering team months before launch: design-for-manufacturing review, material selection with the laminate maker, test lots and reliability qualification. Once qualified, GCE builds the boards to forecast, ships them to the contract manufacturer that assembles the server, and adjusts output week to week as the customer's rack build-outs move. If the board fails in the field, or the next generation is late, the customer's entire AI cluster schedule slips. That is the stake the customer is managing when it picks a board supplier.

2. Business Segments

GCE reports one business, PCBs, but discloses revenue by end-application. The applications differ enough in technology, customer and economics that they are worth taking one by one. The mix figures below are from the 28 August 2026 conference (BigGo summary).

Servers (78% of Q2 2026 revenue; 80% of 9M 2025)

This is the company. It covers general-purpose server motherboards for cloud and enterprise and, increasingly, AI server boards: baseboards and mainboards for custom accelerators (ASICs) designed by cloud providers, plus boards for GPU-platform servers. Management said AI servers alone were "around 40% to 50%" of total revenue in Q2 2026, "not yet at 50%."

The core capability is volume production of very high-layer-count boards in low-loss materials at stable yields, combined with long qualification histories at the largest cloud buyers. Servers were the growth bet a decade ago. Today they are both the margin engine and the growth engine, and management talks about almost nothing else. Main rivals: Victory Giant Technology, WUS Printed Circuit, TTM Technologies, Tripod Technology, Isu Petasys.

Networking (15% of Q2 2026; 10% of 9M 2025)

Boards for data-centre switches and routers. The growth driver is the move from 400G to 800G and then 1.6T switches, which roughly adds 20 layers per board generation (see Section 3). This was GCE's historical franchise and is now the second leg of the AI story: every AI cluster needs a much larger back-end network, and the switch boards that carry it are among the hardest boards in the industry. Networking's share rose again in 2026 after shrinking during the server surge. Main rivals: WUS, TTM, Isu Petasys, Victory Giant.

Notebooks (5% of Q2 2026, down from about 20% historically)

Notebook motherboards, now including HDI boards for "AI PCs." This is a mature, competitive, price-driven line. Management described it as flat. It survives as a capacity filler for older lines, and as a proving ground for the HDI process now being pushed into servers (Section 3). Main rivals: Tripod, Compeq, and mainland Chinese HDI makers.

Other (2%)

Automotive, IC test boards and low-earth-orbit satellite boards. These are niche and strategically optional. None moves the business today.

Geographic footprint. GCE does not disclose revenue by customer destination in the sources reviewed, and its customers' contract manufacturers sit mostly in Taiwan, China, Mexico and Southeast Asia. That makes sales geography less telling than production geography, which is how the company itself frames the issue.

• Taiwan (Zhongli, Taoyuan): production since 1981, and the home of the highest-end boards, running at more than 95% utilisation per analyst coverage.
• China: Suzhou since 2000 and Changshu since 2010, mostly mid-tier and volume products.
• Thailand (Prachinburi): production since 2025.

In May 2025 management said direct shipments to the US were only a single-digit share of sales, so US tariffs on Taiwanese or Chinese boards hit it only indirectly (Moneyweekly). The strategic shift is towards Thailand and Taiwan and away from incremental China capacity, which Section 3 covers plant by plant.

3. Products and Operations

The products that matter

1. AI accelerator baseboards and mainboards (custom ASIC platforms). Coverage of the company consistently names four cloud ASIC platforms: AWS Trainium, Google TPU, Meta MTIA and Microsoft Maia (Vocus deep analysis, 2026). Only the AWS relationship is confirmed by a company action, the Amazon private placement. These boards run to the high 20s to 40-plus layers in M8-class and higher low-loss laminates with HVLP (very-low-profile) copper foil. Many ASIC racks use cableless designs, with all signals running through board traces, which pushes layer count and material grade up again. Each new accelerator generation is a fresh qualification and usually a higher-spec, higher-priced board, which is why the Trainium 2.5 to Trainium 3 transition dominates the conference commentary (Section 9).

2. General-purpose and GPU-platform server boards. Motherboards for x86/Arm cloud servers, and universal baseboards (UBB, 30+ layers) for GPU platforms. In May 2025 management linked second-half demand partly to Nvidia GB200 ramping in Q3 2025 and GB300 starting in small volume.

3. High-speed switch boards (800G / 1.6T). These are 38 to 48 layers, about 20 more than a 400G switch board. They are among the most technically demanding boards in volume production because the switch ASIC's SerDes lanes must cross the board at very high speeds with tightly controlled loss.

4. HDI boards for servers and AI PCs. HDI uses laser-drilled microvias and sequential lamination to pack more connections into less area. GCE's HDI history is in notebooks. At the 28 August 2026 conference management stressed, repeatedly, that it now makes HDI server boards. The emphasis is a tell: next-generation AI boards are moving towards HDI-plus-high-layer hybrids, and investors have doubted whether GCE can match HDI-heavy rivals there (risk 8.4).

"We are now doing AI PC, and we are making HDI server boards... We really can do HDI." - GCE spokesperson, Q2 2026 conference, 28 August 2026 (translated, via BigGo)

5. Heavy-copper, IC test and satellite boards. These are niche, and matter mainly as capability proof.

How a high-layer board is made, and where it goes wrong

1. Engineering (CAM / DFM). The customer's design is converted into production tooling, and the stack-up and laminate choice are agreed with the customer and the laminate supplier.
2. Inner-layer imaging and etching of each core.
3. Lamination. Cores and prepreg are pressed under heat, often in several sequential cycles for very thick or HDI boards. Layer-to-layer registration is the first big yield risk.
4. Drilling. Mechanical drilling of high-aspect-ratio through-holes and laser drilling of microvias for HDI. Back-drilling removes stubs on high-speed nets.
5. Plating. Copper plating of hole walls. Thick boards make it hard to plate the middle of a deep hole evenly.
6. Outer-layer imaging, solder mask and surface finish.
7. Electrical and impedance testing, reliability sampling, then shipment to the server or switch assembler.

The binding constraints are not just floor space. They include lamination press and drill capacity for thick boards, and the supply of premium materials. Low-Dk glass cloth (including T-glass), HVLP copper foil and top-grade CCL were all short through late 2025 and 2026, and industry commentary expects T-glass to stay tight into 2027 (Chinatimes, Feb 2026). GCE is a buyer, not a maker, of these materials. Key laminate suppliers in Taiwan include Elite Material and Taiwan Union Technology.

Plant map and capacity events

Group capex was guided at NT$6-7bn for 2026 in November 2025. It was raised to about NT$17bn at the January 2026 board and to about NT$19bn by August 2026, with management promising more for H2 2026 and 2027 (Section 9 assesses this pattern). For the November 2025 conference, management had mapped peak monthly output stepping up through Q1-Q3 2026 as the Taiwan and Thai capacity came on. July 2026 output ran well above the top of that plan, which says the plants were squeezed harder than planned rather than that capacity arrived early.

4. Customers

Who buys. In substance GCE sells to perhaps half a dozen decision-makers: the hyperscale cloud providers (AWS, Google, Meta, Microsoft per analyst coverage), the networking-system vendors whose 800G/1.6T switches go into those clouds, the GPU-platform ecosystem, and notebook brands. The purchase order, however, usually comes from the original design manufacturer (ODM) or electronics manufacturing services (EMS) firm that assembles the server or switch. This split between who chooses and who pays is the most important fact about GCE's customer base.

Who decides, and how.

• Hyperscaler ASIC programmes. The cloud provider's hardware engineering team (board designers, signal-integrity engineers and supply-chain leads) decides which PCB plants are qualified, effectively approving a vendor list for each board. The ODM then allocates volume among the approved plants, within shares the cloud provider often steers. Criteria, roughly in order: can the plant make this board at all (layer count, material, loss budget), yield and delivery reliability during the ramp, capacity commitment for the next 12-24 months, geographic diversification (non-China capacity increasingly counts), then price. The "sales cycle" is a design-in that starts six to twelve months before a chip generation launches and is effectively re-run every generation.
• Networking vendors. These are similar but slower-moving. Switch platforms live longer than accelerator generations, and qualification is tied to each switch ASIC generation.
• Notebook brands and ODMs. Classic price-and-capacity buying, with quarterly allocation.

Why GCE gets chosen. It has a long qualification record at the cloud providers, going back to the general-server era. It has the high-layer process proven at volume. It is willing to commit capacity ahead of demand. And it now has a Thai plant that answers the customer's question "what if China capacity is restricted?" Management said in May 2025 that its server boards held the "highest certification grade and allocation weight" at customers. Broker estimates put its share of AWS Trainium boards at roughly 35-50% and rising (MoneyDJ, 12 Jan 2026). Those figures are broker figures, not company disclosures.

Switching costs. These are real within a generation and partial across generations. Re-qualifying a different PCB plant for a live board means new test lots, signal-integrity validation and reliability testing, months of work that no customer starts mid-ramp. But every new chip generation reopens the allocation, and customers deliberately keep two or three qualified plants per board. GCE therefore cannot be displaced quickly, but it can be out-allocated at the next generation, which is how share in this industry actually moves.

Concentration. This is high and rising. AWS's Trainium programme is the single most important account, now formalised by Amazon taking an equity stake. In Q4 2025 a single customer's product transition was enough for management to flag a sequential decline in the quarter, with the AI share of revenue falling from about 50% to 30-40% (Section 9). The concentration is mainly a reflection of quality, because you only get this concentrated by winning the hardest boards. It is also the business's main exposure (risk 8.1).

Contract structure. There are no long-term take-or-pay contracts in evidence. Business runs on rolling forecasts and purchase orders, with allocation set per generation. The Amazon private placement (Section 10) is the closest thing to a long-term commitment. It is a 0.3-0.4% equity stake bought at a discount and locked up for three years, which signals intent but does not guarantee volume. Revenue is predictable a quarter or two out and much less so at generation boundaries.

5. Competitive Landscape and Moat

Industry structure. The global PCB industry is large and fragmented, dominated by Taiwanese and mainland Chinese makers. Chinese and Taiwanese firms held 13 of the top 20 places in recent rankings, with Zhen Ding, Unimicron and Dongshan near the top by size (Sina). Most of that volume is phone, PC and consumer boards. The slice that matters here, very-high-layer boards for AI servers and high-speed switches, is a much tighter oligopoly. On the company's own milestone page GCE ranked 13th globally on the Prismark-100 in 2025, yet in its niche it is one of perhaps five or six serious suppliers.

How GCE wins and loses.

• Against Victory Giant, GCE wins where the customer is a cloud ASIC programme that values non-China capacity and GCE's long qualification history. It loses where the board leans heavily on HDI and the Nvidia ecosystem, which Victory Giant has captured. The HDI-in-servers push in Section 3 is GCE moving onto Victory Giant's ground.
• Against WUS, the fight is in switches. WUS has the deeper networking pedigree in China. GCE counters with Taiwan and Thai capacity for customers de-risking from China.
• Against TTM, GCE wins on cost and Asian ramp speed. TTM wins where a buyer needs US-made boards (defence, some government work).
• Against Tripod and the Taiwan HDI makers, GCE is more specialised at the high-layer end. They are broader and steadier.

Barriers to entry are high but not permanent. A new entrant needs:

• A large up-front bill for presses, drills and plating lines sized for thick boards.
• Several years of yield learning on 30-50 layer stack-ups.
• Allocation of scarce low-loss materials, which suppliers ration to proven customers in a shortage.
• Most of all, qualification at a cloud provider, which only comes generation by generation.

The best-funded challengers are mainland Chinese firms that already have scale. Victory Giant's rise shows the barrier can be climbed in a few years by an incumbent PCB maker, though it remains very hard for a true newcomer.

Market share. GCE is widely cited as the largest server PCB supplier, at about 25% of global server PCBs per analyst coverage (Vocus). Its share of individual ASIC programmes is higher: a US broker in December 2025 projected up to 70% of Trainium 3 boards and a rise above 50% in Meta's high-end specs (CMnews, 5 Dec 2025). These are broker projections, not company figures. Share is concentrated because each generation rewards whoever ramped the last one cleanly.

Structural shifts.

1. Everyone is expanding at once. Victory Giant, WUS and GCE all announced multi-year capacity programmes in 2025-2026, setting up a test of pricing in 2027-2028.
2. Geography matters: Thailand has become the default non-China location for Taiwanese and some Chinese PCB makers.
3. Technology convergence: AI boards are mixing HDI with high-layer construction, blurring the line between GCE's specialty and the HDI specialists'.
4. Upstream materials have become the scarce resource, which shifts bargaining power towards CCL, glass-cloth and copper-foil makers.

Moat: Narrow - regulatory or qualification barriers, scale / cost advantage

GCE's position rests on per-generation qualification at a few cloud buyers (Section 4) and on proven volume-yield on the hardest boards. Amazon taking an equity stake is evidence that a key customer values continuity of supply. The advantage has not yet been tested through an AI-specific downturn or against a determined price attack in its core ASIC niche, so it does not qualify as wide. Qualification resets at every chip generation, and peers are building comparable capacity. The thing most likely to erode it is a technology shift to HDI-heavy or hybrid AI board designs in which a rival with deeper HDI process depth wins the next generation's allocation.

6. Industry

What drives demand. High-end PCB demand is now set mainly by data-centre capital spending: the number of AI accelerators and servers the hyperscalers deploy, multiplied by how much board each unit needs. The second factor is the more interesting one. Each accelerator generation raises board layer count, material grade and area (more chips per tray, cableless designs), so board value per server grows faster than server units. The same applies in networking: AI clusters need far larger back-end switch fabrics, and each speed step from 400G to 800G to 1.6T adds layers and requires better laminate. General servers, PCs, phones and autos still drive most global PCB volume but little of the incremental value.

Size and growth. Estimates vary widely with definitions. One industry estimate cited in May 2026 put the global PCB market at about US$114bn in 2026, up about 23% (CMnews). A narrower Mordor Intelligence estimate sizes data-centre and AI-server PCBs at about US$9.7bn in 2026 (Mordor). Industry commentary puts AI servers' share of PCB demand at over 25% in 2026, up from about 15% in 2025. The direction is clear: AI and networking are taking a fast-rising share of a slow-growing total.

Position in the supply chain. The chain runs: glass-fibre cloth and copper foil, then copper-clad laminate (CCL), then PCB fabrication (GCE), then server/switch assembly at an ODM or EMS, then the hyperscaler's data centre. Upstream is concentrated and capacity-constrained in the premium grades, with T-glass and Low-Dk cloth in particular controlled by few producers, mainly Japanese. Downstream buyers are extremely concentrated. PCB makers sit between two concentrated groups, which caps their bargaining power except during shortages.

Geography and substitution. Production has long concentrated in mainland China and Taiwan. US tariffs and export controls, together with customer risk policies, are pushing new capacity to Thailand, Vietnam and Malaysia. Thailand alone has drawn dozens of PCB investments since 2023. This is less "import substitution" than "China-plus-one". US-domestic PCB capacity remains small and mostly defence-oriented.

Regulation. No licence is needed to make PCBs. The gates are customer qualifications (UL, TL-9000 for telecom, IATF 16949 for auto) plus environmental permits for plating and etching chemicals. These permits are increasingly a real constraint on building new plants in Taiwan and China. US tariffs and trade policy shape where capacity gets built.

Cyclicality. PCBs are a classic capital-intensive cyclical. Demand swings follow downstream inventory cycles, and capacity arrives in lumps 12-24 months after a decision. The 2023 downturn, when server and PC inventories were run down, is the most recent reminder. AI has made the high end less tied to the consumer cycle and much more tied to a handful of buyers' capex budgets, which can change faster than consumer demand.

Tailwinds: accelerator and switch generation upgrades that raise board value per system; hyperscalers designing their own ASICs, which multiplies the number of board programmes; and China-plus-one capacity requirements. Headwinds: concentration of demand in a few capex budgets; materials shortages and cost inflation; a synchronised industry-wide capacity build that may land in 2027-2028; and the possibility that advanced packaging or co-packaged optics move some interconnect off the board in the longer term.

7. Growth Triggers

Ranked by potential impact on the business, most material first.

• Trainium 3 ramp and a deeper AWS relationship. Management and its board actions point to the next AWS accelerator generation as the main volume and spec step-up. Trainium 2.5 was expected to ramp from late Q4 2025, with Trainium 3 in H2 2026 (Q3 2025 conference, 28 Nov 2025; repeated Jan 2026 and 18 Mar 2026). On 29 September 2026 the board approved Amazon as a strategic investor through a private placement, with the company stating that high-end AI board capacity "still cannot fully meet customer demand" (SMM). Repeated across four events.
• AI server share to keep rising in H2 2026. Management guided Q3 2026 AI server and networking demand above Q2, with AI server growth "even more pronounced", and described Q3 as peak season. Coverage of the May 2026 conferences reported an AI-related revenue share target of about 70% for 2026 (Q2 2026 conference, 28 Aug 2026; Q1 2026 conferences, 26 May to 5 Jun 2026).
• Thailand scale-up. Phase 2 of plant 1 was pulled forward to Q2 2026 and targets several times phase 1's output in H2 2026. Plant 2, on existing land, is targeted for H2 2027 (board, 27 Jan 2026; Q1 2026 conferences; Q2 2026 conference, 28 Aug 2026). Repeated across four events.
• Taiwan capacity, near and long term. The leased Yangmei site is already producing. A new Taiwan plant on the site bought from CMC Magnetics is expected late 2028 to early 2029. A further NT$7.9bn land-and-plant project begins in Q4 2026, funded by convertible bonds (board, 27 Jan 2026; Q2 2026 conference; board, 29 Sept 2026).
• China capacity rebuilt for existing customers. Suzhou plant 2 targets H2 2027 and a new Suzhou plant 3 H1 2028. Changshu is being demolished and rebuilt for 2028 production (board, 27 Jan 2026; Q2 2026 conference, 28 Aug 2026).
• 800G to 1.6T switch boards. Management cited strong 800G switch demand as the second growth leg alongside ASIC servers, with 1.6T boards following (Q3 2025 conference, 28 Nov 2025; BofA conference, 18 Mar 2026). Repeated.
• HDI server boards and AI PC HDI. GCE is now producing HDI server boards and AI PC HDI boards (Q2 2026 conference, 28 Aug 2026).
• More capex in H2 2026 and 2027. Beyond the roughly NT$19bn already announced for 2026, management said further increases would come in H2 2026 and continue into 2027, calling it future "good news" (Q2 2026 conference, 28 Aug 2026).

8. Key Risks

8.1 One programme, one customer, one transition at a time. The largest single driver is AWS's Trainium board family (Section 4), and allocation is reset at each generation. Mechanism: if AWS slips a generation, re-tiers suppliers, or shifts part of the next design to a rival with HDI depth, GCE's highest-spec volume falls sharply even if total AI spending keeps growing. Q4 2025 was a small preview, when a CSP product transition alone pulled the AI share from about 50% down to 30-40%. Calibration: a high probability of periodic air pockets at generation changes, and a low-to-moderate probability of a material share loss. The Amazon stake lowers but does not remove the second risk.

8.2 Everyone builds at once. GCE has gone from a NT$6-7bn capex plan to about NT$19bn and rising for 2026 (Section 9), with new plants landing from H2 2026 through early 2029. Victory Giant, WUS and others are expanding on similar timetables. Mechanism: if AI capex growth slows in 2027-2028 just as this capacity lands, high-layer boards go from allocation to price competition. Depreciation from the new plants is then fixed while pricing is not. Analyst coverage flags exactly this (Moneyweekly). Calibration: moderate probability, high impact. This is the classic way PCB upcycles end.

8.3 Premium-material shortages. GCE cannot build boards without T-glass/Low-Dk cloth, HVLP copper foil and top-grade CCL, all tight through 2026 and possibly 2027 (Section 3). Mechanism: materials allocation, not floor space, caps output in a ramp; cost increases may not pass through fully if customers resist; and a rival with better laminate-maker relationships ramps first. A December 2025 broker report singled out T-glass and Q-glass lead times and pricing as a direct threat to costs and competitiveness. Calibration: high probability of friction in 2026-2027, moderate impact.

8.4 Technology transition risk, specifically HDI. Next-generation AI boards increasingly combine high-layer construction with HDI microvia structures. That favours makers whose process depth lies in HDI, notably Victory Giant and the Taiwan HDI specialists (Section 5). Management's need to tell investors repeatedly that "we really can do HDI" (Section 3) shows the market's doubt. Mechanism: lower yields on hybrid boards during qualification, or losing the design-in at the next generation. Calibration: moderate probability, high impact, because it would hit exactly the highest-value boards.

8.5 Execution across four regions at once. GCE is simultaneously ramping Thai phase 2, starting Thai plant 2, running a leased Taiwan site, building two new Taiwan plants, expanding Suzhou and rebuilding Changshu. Mechanism: yield learning curves at new sites, staffing new plants with experienced engineers, and management bandwidth. A new site running high-spec boards at low yield is worse than no site. Calibration: moderate probability of at least one site slipping, moderate impact. Thailand phase 1 already took about a quarter longer than first guided before it turned profitable (Section 9).

8.6 Dilution from funding the build. The NT$7.9bn plant is funded by unsecured convertible bonds, and convertible bonds were issued earlier in 2026 as well. The Amazon placement was priced at 80% of the reference price, the maximum discount Taiwan's private-placement rules allow. Mechanism: conversion of bonds into new shares, and possibly further discounted placements to customers if more "strategic" stakes follow. Calibration: high probability of some dilution, low-to-moderate impact.

8.7 Geography cuts both ways. A meaningful part of capacity remains in Suzhou and Changshu, and those sites are being expanded and rebuilt (Section 3). Mechanism: US restrictions on Chinese-made content in US hyperscaler hardware, or customer policies requiring non-China capacity for advanced AI boards, would strand the China investment or force faster relocation. Concentrating the highest-end boards in Taoyuan also carries Taiwan risk. Calibration: low-to-moderate probability, high impact.

9. Walk the Talk

GCE does not hold formal quarterly earnings calls. It presents results and outlook at broker-hosted investor conferences. The six results events used here:

1. Q1 2025: Goldman Sachs TechNet Taiwan (20 May 2025) and Citi Taiwan Tech Conference (23 May 2025)
2. Q2 2025: Yuanta Securities investor conference (20 Aug 2025)
3. Q3 2025: Daiwa/Fubon investor event (28 Nov 2025), with the presentation filed on MOPS
4. Q4 2025 / FY2025: Thailand plant investor tours (13 and 15 Jan 2026), the capex board meeting (27 Jan 2026) and the BofA Asia Tech Conference (18 Mar 2026)
5. Q1 2026: UBS (26 May 2026), Goldman Sachs (2 Jun 2026) and Citi (5 Jun 2026) conferences
6. Q2 2026: Fubon Securities investor conference (28 Aug 2026), 36 days before this report; JPMorgan Asia Tech Tour on 19 Aug 2026

Where the record is informative.

The Q4 2025 call was honest about the air pocket. At the November 2025 conference, with the stock near highs and the AI story at its loudest, management told investors that the next quarter would be softer because a key ASIC customer was between generations and year-end shipments were being rearranged. Q4 then came in down sequentially, as said, and the following two quarters set records. Management that pre-announces a dip, rather than letting the market find it, is managing expectations conservatively on the downside.

The capex guide was not a guide. Capex went from NT$6-7bn in November to NT$17bn in January, nearly tripling within two months, and kept rising after that. This is not a miss in the usual sense, because spending more is a choice and was driven by customer demand. It does show that management's capacity planning tracks customer pull month by month rather than a long-range plan, so stated capex figures should be read as floors.

"In the second quarter, the AI share is currently around 40% to 50%... not yet at 50%." - GCE spokesperson, Q2 2026 conference, 28 Aug 2026 (translated, via BigGo)

The AI share target reads as optimistic. Coverage of the May 2026 conferences reported an AI-related share heading to 70% in 2026 (CMnews, 18 May 2026). At the half-year mark management itself put AI servers at under half of revenue. Part of the gap is probably definitional, since "AI-related" may include 800G AI-cluster switch boards while "AI servers" does not. Even so, closing it within one half-year needs an exceptional H2. This is the item to watch at the Q3 2026 results.

Suzhou's timeline quietly moved. "Suzhou expansion in Q3 2026" in March became "Suzhou plant 2 in H2 2027, plant 3 in H1 2028" by August. It may be that the Q3 2026 step was a smaller in-plant upgrade and the new plants are additional, but the company has not reconciled the two in the sources reviewed.

Assessment. On near-term operations this management is credible. It flags soft quarters before they happen and delivers on plant start-ups, roughly on time or early. Its capex and long-range mix statements are much looser, shifting with customer pull and drifting optimistic. The pattern is a conservative-on-quarters, expansive-on-ambition team whose capital plans should be treated as floors that move.

10. Ownership, Governance and Shareholder Friendliness

Part 1 - Ownership and control. According to the most recent annual report as summarised by ifa.ai and related data services, the largest shareholder is director 楊長基 (Yang Chang-Chi) with about 19.65%. Next are Taiwan's New Labor Pension Fund at about 6.31% and 李瑞清 (Li Rui-Qing) at about 5.62%. Chairman 楊承澤 (Yang Cheng-Tse) holds about 3.6%. Directors as a group hold about 24%, and foreign investors about 23%. These sources do not document any family relationship among the board members who share the Yang surname, so none is assumed here. GCE has one class of share, with no dual-class, golden-share or pyramid structures in evidence. Amazon will hold about 0.3-0.4% after the September 2026 private placement, locked up for three years. What the structure means for a minority holder: the largest individual holder has about one-fifth of the votes. That is influential, but well short of control in a widely held stock with a large institutional and foreign float, so a minority holder's votes count alongside the board's.

Part 2 - Governance (flags only). Data services record zero pledged shares by directors (ifa.ai), so there is no pledge overhang. The notable item is the September 2026 private placement. It issued new shares to a customer at 80% of the reference price (TechNews, 29 Sept 2026), the deepest discount Taiwan rules allow. The dilution is tiny, but it sets a precedent of discounted equity for strategic customers. No material related-party transactions with major holders, auditor changes or qualified opinions were found in the sources reviewed. Details of board independence and of how executive pay links to performance could not be verified within the search budget.

Part 3 - Capital returns (last three financial years).

Dividends. Cash dividends per share, by year of payment:

• NT$3.5 in 2024, from FY2023 earnings (ex-date 27 Jun 2024)
• NT$6.0 in 2025, from FY2024 earnings (ex-date 24 Jun 2025)
• NT$9.8 in 2026, from FY2025 earnings (ex-date 23 Jun 2026, paid 22 Jul 2026, per the ex-dividend notice)

One data service lists the FY2025 payout as NT$10.0. The dividend roughly tripled over the period, growing each year with earnings, despite capex tripling at the same time.

Buybacks and dilution. MoatMap's database records no buybacks in the last ~90 days (since 5 July 2026). External searches of exchange announcements and financial news for 2023-2026 found no treasury-share repurchase programme. That is an absence of evidence across the search, not a verified zero. The share count is moving the other way: convertible bonds were issued in 2026 and more are planned for the NT$7.9bn plant, and the September 2026 placement adds 1,856,308 shares. The net three-year change in shares outstanding could not be verified. The direction is modest growth through conversions and placements.

Verdict: Returns Capital - a sharply rising cash dividend paid alongside a record capex build, with only small dilution from financing so far.

11. Insider Activities

MoatMap's database (TW venue, current to 2 October 2026) records a single insider transaction in the last 12 months. On 26 August 2026 director 楊長青 (Yang Chang-Ching) reported 100,000 shares, about 0.02% of shares outstanding, classified as "Other", with no price or value (TWSE insider shareholding change filing, 2026-08-26). "Other" in Taiwan's filings typically covers gifts, transfers into or out of trusts, and similar non-market movements rather than open-market trades. The specific reason is not disclosed in the data. There were no open-market purchases or sales by directors, officers or holders of 10% or more in the window. Read: neutral. One small non-market transfer by one director carries no directional signal, and the absence of open-market selling into the stock's 2026 rally is mildly reassuring in its own right.

12. Scenarios

Bull case. Trainium 3 ramps cleanly in late 2026 and GCE takes the dominant share the brokers project. The Amazon stake turns out to be the first step of a capacity partnership rather than a gesture. 800G-to-1.6T switch upgrades fill the networking lines, and GCE's HDI server boards win a place on a second hyperscaler's next accelerator, which settles the HDI debate. Materials remain tight, which helps GCE: its scale and history earn it allocation that smaller rivals cannot get. Thailand plant 2 and the Suzhou and Taiwan plants arrive in 2027-2028 into demand that is still unmet. By 2028 GCE is the default non-China supplier of high-layer AI boards to three or four cloud platforms, with a plant network sized for it and the generational resets favouring the incumbent.

Base case. Management delivers roughly what it has guided: H2 2026 is strong, Trainium 3 ramps, and the AI share rises, though more slowly than the 70% headline. Generation changes produce one or two soft quarters a year, flagged in advance as in Q4 2025. Thai phase 2 and the Yangmei site carry growth through 2027. The Suzhou and Taiwan plants arrive on the later of their timelines, and capex keeps being revised up as customers pull. Competition from Victory Giant and WUS keeps pricing honest without breaking it. GCE ends 2028 as a much larger, more AI-concentrated version of today, with a heavier fixed-cost base and a somewhat larger share count from convertible bonds.

Bear case. Risks 8.1 and 8.4 compound first: the next AWS generation, or another hyperscaler's, shifts to a hybrid HDI design and a rival wins a larger share of it. 8.2 follows, as the industry-wide capacity GCE and its peers committed in 2026 lands in 2027-2028 just as hyperscaler capex growth slows, turning allocation into price competition. 8.3 makes it worse: material costs stay high even as board prices soften. 8.5 shows up as poor yields at new Thai and Suzhou lines during the downturn, when customers are least forgiving. 8.6 adds convertible-bond dilution at a bad time. GCE stays a qualified supplier but becomes one of several, with an expensive plant network built for a demand curve that flattened early.