FII Zen Horizon Research: No. 1 in AI server rack assembly, real earnings acceleration, but a thin business at a rich valuation

Hunting ten-year five-baggers among great growth stocks — pressing the upside question: "Can it get much bigger?"

• QHow high is its market ceiling? Is it enlarging an existing pie, or creating an entirely new market?5/10

  A

  Net judgment upfront: Question 1 is about the market ceiling. On absolute scale, FII is a top-tier answer; on market type, it is the opposite. The AI compute capex chain where AI server/rack-scale assembly sits is one of this era's largest capital spending waves: the AI server market is projected at about USD 525.0 billion-854.0 billion by 2030, while upstream AI data center total capex is seen by Nvidia/McKinsey at a cumulative USD 3 trillion-7 trillion by 2030. FII fully checks the box for a very large 10-year opportunity. But what Baillie Gifford really wants is a company that creates a new market, whose volume growth can be extrapolated, and whose moat widens with scale. FII is exactly the reverse: it is expanding the existing, mature, and increasingly Nvidia-compressed outsourcing pie of rack-scale assembly/system integration, not creating a new category. Its unit economics, with consolidated gross margin of 6.94% and white-box rack assembly gross margin sliding from ~18-20% toward 6-8%, mean that as volume gets larger, the blade gets thinner. So the honest conclusion is: the height of the ceiling is impeccable; the nature of the ceiling is where Baillie Gifford would deduct points. It is huge but low-quality, extrapolatable but not thickening.

  ---

  1. Absolute scale: the ceiling is indeed extremely high, and FII fully scores here

  FII sits on the main wave of AI compute capital expenditure, with absolute scale among the largest buckets in any growth-stock hunting ground:

  • The AI server market itself: third-party 2030 forecasts range from ABI's about USD 524.0 billion (CAGR ~18% from 2024) to Grand View's about USD 854.0 billion (2025-2030 CAGR ~38.7%). The ranges differ, but the order of magnitude is consistent: this is a market starting from about USD 125.0 billion-143.0 billion in 2024 and multiplying several times in five to six years. In the narrower generative AI server category, MarketsandMarkets sees about USD 448.6 billion by 2030 (CAGR ~34%).

  • Upstream AI data center total capex, FII's real demand pool: Nvidia management sees global data center capex at USD 3 trillion-4 trillion by 2030 (from about USD 600.0 billion in 2025, CAGR ~46%); McKinsey is more aggressive, with cumulative AI infrastructure investment before 2030 of about USD 7 trillion, including about USD 5.2 trillion into AI data centers.

  • Near-term validation, consistent with the report's hard numbers: the 2026 aggregate capex estimate for the world's nine largest cloud vendors is about USD 830.0 billion, with YoY growth raised from 61% to 79%, broadly in line with the report's "North American cloud vendors > USD 800.0 billion, +67%" and slightly higher; NVL72 rack shipments rise from about 29,000 units in 2025 to about 70,000-80,000 units in 2026, with order visibility extending to 2027.

  On Baillie Gifford's preference for a market with 10-year scale and extrapolatable volume growth, FII is a textbook "yes". This is why it deserves AI compute beta, and that must be acknowledged.

  ---

  2. Type: it is enlarging an existing pie, not creating a new market, and Baillie Gifford would deduct here

  Baillie Gifford ranks "creating an entirely new market" above "enlarging an existing pie" because new markets often come with pricing power, structurally high margins, and winner-take-most extrapolation. On this yardstick, FII is clearly the opposite:

  • It works in an existing link: rack-level L10-L12 assembly/system integration is an extension of EMS/ODM, a mature business that has existed for decades. The object has changed from phones and general servers to AI racks. AI is not a new market created by FII; Nvidia created it. FII is the participant enlarging the old "assembly" slice within that new demand chain.

  • The value and pricing power of the new market sit upstream, not in assembly: Nvidia's gross margin is about 74.9%, HBM memory suppliers take another large share, and white-box assemblers are left with single-digit gross margins. Even when rack value is high, with GB200 NVL72 at about USD 2.80 million-3.40 million, GB300/Rubin higher, and next-generation Vera Rubin NVL72 quoted up to about USD 8.80 million per rack, almost all of the incremental value is in Nvidia-specified GPUs/compute trays. Compute trays account for about 90% of total rack cost; FII earns only a thin processing and integration fee on top. This exactly confirms section 8.1 of the report: even when AI servers scale, this remains a "low-margin, high-volume" business whose ceiling is set by outsourcing economics.

  • Number-one share does not equal creating a new market: FII's NVL72 rack outsourcing share is about 40%-48%, ranking first, consistent with the report's 45-48%, with Quanta at about 25%-40% and Wistron at about 8%-10%. But that is the largest slice of an existing pie. Competitors such as Quanta, Wistron, and Inventec can be reallocated share by Nvidia at any time. FII is a same-order outsourcing leader, not an irreplaceable category definer.

  ---

  3. The most damaging point: this pie's ceiling gets thinner as it scales, not thicker in the Baillie Gifford sense

  Baillie Gifford wants not only extrapolatable volume, but compounding where unit economics and the moat improve with scale. AI server assembly moves the other way: volume is extrapolating, profit is thinning.

  • Assembly gross margin is structurally compressed: white-box GB200 NVL72 assembly gross margin was about 18%-20% in 1Q25, but GB300 white-box is already expected to fall to about 6%-8% by 1Q26. The reason is structural: Nvidia keeps increasing AI server design integration, using more Nvidia-specified components, compressing ODM customization space; revenue rises with GPU/memory prices, but processing fees do not rise proportionally. The assembly link is commoditized as volume comes without profit.

  • This means the quality of the ceiling is worsening: the market-volume ceiling is not visible within 10 years, but the profit ceiling FII can capture from it is capped by outsourcing economics and moves lower as Nvidia deepens integration. This conflicts with Baillie Gifford's picture of a great growth stock that can rise fivefold over the next 10 years and become stronger as it grows. The volume story is large, but the profit story does not thicken and may thin.

  • Two fair strengths should still be added: 1. FII's full-stack vertical integration (modules -> substrates -> high-speed switches -> liquid cooling -> whole systems -> data centers) and self-supplied liquid cooling (liquid-cooling gross margin about 15%, higher than full racks) let it eat slightly more of this thin pie and resist compression better than pure assemblers; 2. its near-term volume growth is real acceleration and real delivery: FY25 revenue ¥902.887 billion +48.22%, cloud computing ¥602.679 billion +88.70%, AI server revenue up more than 3 times YoY, Q1'26 net profit ¥10.595 billion +102.55%. The volume expansion is real, not PPT. But these improve "taking a bit more in a low-quality ceiling"; they do not change the nature of the ceiling as assembly outsourcing.

  ---

  4. Final conclusion for Question 1

  Baillie Gifford Q1 dimension	FII's answer	Assessment
  Absolute market size (large over 10 years?)	AI servers seen at ~USD 525.0 billion-854.0 billion in 2030; AI data center capex cumulative at USD 3 trillion-7 trillion	Strong -- top-tier scale
  Can volume growth be extrapolated?	Shipments 2025 ~29,000 units -> 2026 ~70,000-80,000 units; orders visible to 2027	Strong -- near-term path is clear and extrapolatable
  Enlarging an existing pie vs creating a new market	Enlarges the existing outsourcing pie of rack assembly/integration; Nvidia created the new market	Weak -- expands an old pie, not a new category
  Quality of the ceiling (capture and thickening)	Consolidated gross margin 6.94%; white-box assembly margin sliding from 18-20% toward 6-8%; value and pricing power go to Nvidia/memory upstream	Weak -- capped by outsourcing economics and thinning

  One-sentence characterization: FII stands in a huge market with impeccable 10-year absolute scale and extrapolatable volume growth, but it earns from the existing, low-margin, upstream-compressed assembly-outsourcing slice of that market. It is enlarging an existing pie, not creating a new market, and certainly not the compounding ceiling that widens with growth that Baillie Gifford most wants. On Question 1, FII scores high on "space" and low on "nature and capture quality". This is the specific market-ceiling expression of a good track, thin business.

  Jun 5, 2026
• QCan its revenue at least double over the next five years? Is growth mainly driven by volume, price, or new businesses?7/10

  A

  Conclusion upfront: a five-year revenue doubling, from ¥902.9 billion to about ¥1.8 trillion, is likely achievable, but it is a front-loaded pulse from the AI capex supercycle, not steady compound growth. It is also a doubling of low-margin outsourcing revenue: the larger later volumes become, the thinner per-rack assembly margin gets, so revenue doubling does not equal per-share value doubling

  Start with the math: doubling in five years requires only about 14.9% annual compound growth. FII's current run rate, with FY2025 revenue of ¥902.887 billion, up +48.22% YoY, and Q1'26 revenue of ¥251.078 billion, up +56.52% YoY, is several times that threshold. The question has never been whether the first two or three years are fast enough; that is almost certain. The question is whether the firepower can extend into years 3-10. That is exactly what Baillie Gifford cares about most, and exactly where this business deserves scrutiny. Below, the drivers are tested by volume, price, and new business, then the risks are bounded.

  ---

  1. Near-term certainty is very high: sell-side profit paths already imply a doubling within two to three years

  The sell-side revenue path I independently checked brings the doubling point forward: some institutions forecast FY2026 revenue of about ¥1.13 trillion and FY2027 of about ¥1.43 trillion. That means even modest FY2028 growth would let revenue touch ¥1.8 trillion in FY2028-2029, well inside the five-year window to FY2030. Profit forecasts are even more optimistic: Goldman Sachs puts FY2026 net profit at ¥60.6 billion, FY2027 at ¥86.3 billion, and FY2028 raised to ¥115.6 billion, with a ¥100 target price. Orders also support this: AI server order visibility is locked to 2027, and the Rubin platform starts shipping in the second half. This is a relative strength for FII and must be credited: performance is genuinely accelerating, and orders really extend to 2027.

  But one caveat matters: sell-side dispersion is huge. On the same basis, FY2026 net profit ranges from about ¥35.5 billion to CICC's ¥75.8 billion and Goldman's ¥60.6 billion. That dispersion itself says the slope of the doubling is highly dependent on two exogenous variables FII does not control: cloud capex and GPU generation timing, rather than an endogenous, predictable compounding engine.

  ---

  2. Growth is mainly driven by the twin engines of volume and per-rack value, not endogenous pricing power

  Breaking down the drivers, the doubling mainly comes from shipment explosion plus stepwise rises in per-rack ASP; FII itself has almost no pricing power:

  • Volume, the largest contributor: full-year GB200/GB300 NVL72 rack shipments rise from about 29,000 units in 2025 to about 70,000-80,000 units in 2026 by Morgan Stanley's estimate; TrendForce expects global AI server shipments to grow more than 28% YoY in 2026. FII itself reported Q1'26 AI GPU rack shipments up +3.8 times YoY and ASIC servers up +3.2 times. This is the hardest base for revenue doubling.
  • Price, meaning mix upgrade rather than discretionary price hikes: per-rack value steps up by generation: GB300 full rack about USD 4.00 million; next-generation Vera Rubin (VR200) full rack jumps to about USD 7.80 million, nearly doubling. But be clear: almost all of this ASP surge is captured upstream. In VR200's increase over GB300, memory cost rises 435%, PCB 233%, MLCC 182%, ABF substrate 82%, not assembly processing fees. FII passively benefits as rack value rises; it is not the price setter.
  • The underlying TAM is large enough, so the five-year runway is there: AI server market estimates point from about USD 125.0 billion-143.0 billion in 2024 to about USD 524.0 billion-854.0 billion in 2030, CAGR 18-39%; the top five North American cloud vendors' 2026 capex is indicated at USD 630.0 billion-690.0 billion, more than double 2024, and may exceed USD 1 trillion in 2027. If FII holds share, the industry pie itself is enough to support revenue doubling.

  ---

  3. Share sustainability: the number-one position is real, but reallocatable, not a shutdown bottleneck

  Revenue doubling depends on share retention. After checking this, I am cautiously optimistic: FII's leadership is real and has temporarily held through generation shifts, but its moat is vertical integration depth, not exclusivity.

  • GB200 NVL72 substrate assembly share is about 45-48%, ranking first, with Quanta about 40% and Wistron about 10%; GB300 still has FII at about 40-45%, ranking first; GB-series racks for the full year, on the Hon Hai system's basis, exceed half, about 52%.
  • But peers are taking volume: after Inventec was introduced for GB300, its share rose significantly versus GB200 to about 20%, with Quanta about 25%. This shows share can be reallocated among same-order ODMs. FII is first because of capacity, certification, and full-stack self-supply, including liquid-cooling cold plates, CDUs, and self-developed 800G+ switches; it is not an irreplaceable bottleneck like Nvidia or HBM. Every generation shift, from GB300 to Rubin, is a share reshuffle window and an embedded downside tail in the doubling narrative.

  ---

  4. Three points must be held down honestly: the doubling is a pulse, profit does not move in lockstep, and the cycle is exogenous

  This is the key honest part of Question 2: revenue can double, but it should not be packaged as the kind of 10-year steady compounding growth Baillie Gifford seeks.

  1. Growth is front-loaded, and firepower likely cannot extend through years 3-10. Even the most optimistic views acknowledge that capex growth will decelerate quickly: capex YoY growth falls from about 51% in 2026 to about 13% in 2027 and about 5% in 2028; TrendForce and industry sources broadly expect this AI demand cycle to likely peak in 2027, with possible pullback in prices and procurement thereafter. The doubling is almost entirely completed by the first three years, FY26-28, after which years 4-5 are likely low-speed digestion on a high base. That is the opposite of Baillie Gifford's preference for upside concentrated in years 3-10. This is a one-off AI compute capex supercycle, not a 10-year extrapolatable compounding machine. History is clearer: 2018-2023 revenue hovered at ¥400.0 billion-480.0 billion for six full years, and 2023H1 revenue fell -8.2%, the first half-year decline since listing. The steady state of this business was low growth; the current spike came from AI and may recede with AI.

  2. Revenue doubling does not equal profit or per-share value doubling; unit economics may even worsen. That is the fate of outsourcing: consolidated gross margin is only 6.94% and net margin 3.91%, and the more expensive later racks become, the thinner assembly gross margin gets. ODM full-rack assembly gross margin is about 2.7% on GB300 and falls to about 1.9% for VR200/Rubin. This means revenue elasticity from volume doubling plus ASP doubling is partly offset by falling per-rack processing margin, so profit elasticity is weaker than revenue elasticity. The only hedge is mix-up, including self-developed liquid cooling at about 15% gross margin and higher 800G+/CPO network share, with Q1'26 consolidated gross margin already recovering to 7.4%. But that merely nudges net margin from ~4% up by one or two points; it does not change the low-margin, high-volume nature.

  3. Three exogenous dependencies can break the doubling if any loosens: 1. cloud capex durability: Sequoia estimates about USD 600.0 billion of AI revenue gap must be filled each year to justify this investment wave, and the 2026 gap is widening rather than narrowing; if capex marginally contracts, outsourcing sits at the end of the cycle and the elasticity cuts both ways. 2. share retention: as above, generation shifts can dilute share through Inventec and others. 3. smooth GPU transitions: any demand gap in GB200 -> GB300 -> Rubin disrupts shipment rhythm. The top five customers account for 62.01% of revenue, deeply tied to Nvidia and the four North American clouds, which further amplifies these tail risks.

  ---

  One sentence back to Question 2

  Can revenue at least double over five years? Yes, and it is likely to be achieved early in FY2028-2029. This is one of FII's strongest dimensions in the Baillie Gifford ten questions and should be credited: volume explosion, stepwise per-rack value increase, number-one share, and orders visible to 2027. But it is a pulse from the first three years of the AI capex supercycle, not steady compound growth that can power years 3-10. And it is a doubling of low-margin outsourcing revenue whose per-unit assembly margin declines by generation, so revenue doubling is not profit or per-share value doubling. In the Baillie Gifford framework, the honest characterization is: near-term high certainty, long-term high dependence on exogenous cycles. The "yes" on revenue doubling deserves recognition, but it should not elevate FII into a great growth stock that can rise fivefold over the next 10 years. That requires pricing power and durable compounding, not only a capex tailwind.

  Jun 5, 2026
• QAfter five years, what will take over as the next growth engine? Does this "second curve" exist today?4/10

  A

  Question 3, the second curve: what takes over after five years, and is it already emerging?

  Conclusion upfront: the four candidates visible today -- liquid cooling, high-speed switches/networking (800G/1.6T/CPO), L10-L12 data center turnkey solutions, and ASIC custom servers -- have all emerged and are delivering quickly. But measured honestly by Baillie Gifford's yardstick of being truly independent and able to take over when AI capex fades, most are not a "second curve". They are mix-up or higher unit value on the same AI compute curve. The beta and demand root are the same; FII is merely climbing from selling full systems to selling higher-margin components plus systems on that curve. CPO is the only candidate with a differentiated profit-pool potential, but its timing is still controlled by the same Nvidia/Broadcom roadmap and the same cloud customer orders. The bigger reminder is historical: FII has already changed "curves" once, from the 2018-2020 industrial-internet narrative to today's AI server engine, while the industrial-internet segment had only ¥694 million of FY25 revenue, 0.26% of total. That means FII's next-curve narrative should be discounted and tested for whether it is merely the same beta in new packaging.

  ---

  1. First define the standard: what counts as a true second curve

  Baillie Gifford asks about a second curve because it asks who takes over when today's engine stalls. A qualified second curve needs two things: 1. independent demand source, not the same demand driver as the current engine; 2. counter-cyclical or at least asynchronous behavior when the current engine declines. FII's current engine is "North American cloud AI capex -> Nvidia GPU racks -> FII rack assembly". Each candidate must therefore be judged on whether it is a new demand root or the same demand root with FII earning more per unit.

  2. Candidate by candidate, ranked by independence from weakest to strongest

  1. Liquid cooling (self-developed cold plates/CDU/Manifold) -- unit value uplift, not an independent curve. This is the most concrete delivery: FII vertically integrates cold plates and CDUs, self-development reduces cost by about 25%, liquid-cooling gross margin is about 15-18%, higher than about 6% for full racks, and undelivered liquid-cooling orders for the year have exceeded ¥50.0 billion, with North America about 70% and orders extending to 2027. The market also has real independent growth: China's liquid-cooled data center market has a 2024->2027 CAGR of about 41%, and AI training server cooling TAM moves from USD 1.5 billion in 2024 to USD 12.4 billion in 2027, CAGR about 101%. But honestly, liquid-cooling demand is driven by rising GPU/ASIC power density; it is the other side of AI server shipments. Penetration rising from about 20% in 2025 to more than 50% in 2027 depends precisely on these AI racks. If AI capex fades and racks stop shipping, liquid-cooling orders disappear in sync. So liquid cooling is FII replacing a roughly 6% gross-margin task inside the same rack with a roughly 15% gross-margin task. It is unit economics and margin mix uplift, not a counter-cyclical second curve.

  2. L10-L12 data center turnkey solutions -- value-chain extension, beta fully shared. FII is one of the few companies able to self-develop the full stack from modules -> substrates -> high-speed switches -> liquid cooling -> whole systems -> data centers and provide L10-L12 one-stop rack-scale delivery. That is a real differentiation versus Quanta/Wistron, supporting number-one share and about 45-48% NVL72 share. But L10 to L12 means expanding the delivery scope within the same AI data center project, from server to rack to data-center level. The customers are the same cloud vendors, the demand is the same capex, and it expands FII's wallet share per dollar of AI investment, not a new demand pool. It is a horizontal capability moat, but as a "second curve" it is 100% co-rooted and rises or falls together with the current engine.

  3. ASIC custom servers -- somewhat more diversified customers, but still the same AI compute beta. Q1'26 ASIC server shipments rose +3.2 times YoY versus GPU racks +3.8 times, and ASIC is highly customized with profitability significantly better than general servers. Its partial independence is that ASIC servers map to Google TPU, Amazon/Meta self-developed chips, so customer structure is slightly more diversified and dependency on one Nvidia generation is lower. But it is still AI training/inference compute capex, sharing the same cloud AI capital expenditure root as GPU servers. It moves the eggs from the Nvidia basket to "Nvidia plus self-developed ASIC" baskets. It is a hedge against generation concentration, not an independent second curve.

  4. High-speed switches/networking (800G/1.6T/CPO) -- the fastest-growing and most curve-like, but still tied to the same roadmap. This has the strongest second-curve appearance. Hard numbers: FY25 800G-and-above high-speed switch full-year revenue rose more than 13 times YoY. The basis matters: this is full-year; Q3 single-quarter YoY once exceeded 27 times, and both can coexist. The communications and mobile network equipment segment had FY25 revenue of ¥297.851 billion, 32.99% of total, with gross margin 9.28%. That segment margin of 9.28% is higher than cloud computing's 5.73%, so the mix-up thesis holds. Products cover Ethernet/InfiniBand/NVLink Switch. The structural uplift logic is valid.

  Why it is "almost" a curve: in AI clusters, network capex is growing faster than the GPU count itself. In 100,000-card data centers, interconnect including switches reaches about 26.1% of capex, and by 2027 about 20% of Ethernet data-center switch ports will connect to AI servers. So even if rack shipment growth slows, the network layer may keep rising because each card needs more and more expensive networking. But two points must be held down: first, Arista, Broadcom, and Cisco, the pure networking/custom silicon leaders, are also exposed to the same hyperscaler capex cycle. Networking is not demand outside AI; it retreats with the same tide. Second, FII is still an assembly/integration outsourcer in switches, while pricing power is held by upstream players such as Broadcom's Tomahawk switching chips. This repeats its full-system position: high share, low pricing power.

  On CPO (co-packaged optics), this is the only candidate among the four that may jump out of the same beta: FII management publicly calls it "another important growth inflection point after AI servers", CPO all-optical switch prototypes have entered production, and FII is co-developing next-generation CPO ASICs and 1.6T switches with customers. CPO is a genuine technology inflection point, replacing pluggable optics with optical-electrical co-design and potentially reshaping value allocation. But today it is only emerging, and its timing is still locked to the same roadmap: Nvidia Rubin plans CPO in scale-out in 2026 and scale-up only in 2027; Broadcom Tomahawk 6 supports CPO deployment from 2025H2; LightCounting expects mass production only in 2026-2027, with CPO ports reaching about 30% of total 800G/1.6T ports by 2027. Its emergence lands in Baillie Gifford's important years 3-10, which is directionally good. But demand still comes from the same AI data centers, timing is still set by Nvidia/Broadcom, and independence will only be proven if FII obtains irreplaceable share plus pricing power in CPO. Today's evidence is insufficient to call it a formed independent curve.

  3. Back to Baillie Gifford's ruler: sprouts exist, independence is weak, and the narrative needs a discount

  Putting the four together, FY25 facts are clear: cloud computing, mainly AI servers, was +88.70% and 66.75% of revenue, the current curve. Liquid cooling, networking, L10-L12, and ASIC are all parasitic on the same AI capex. They may help FII's consolidated gross margin climb from 6.94% upward, with gross margin rising sequentially since 2025Q2 and profit growth overtaking revenue growth. That is real and should be credited. But it is the same curve becoming steeper and thicker, not a second curve taking over.

  Why the market may overrate this: sell-side and management both tend to package liquid cooling/CPO/networking as a multi-engine second-curve narrative to support the high valuation of ¥78.55 and static PE about 44×. FII's history warns otherwise: the narrative changed once before, and the previous industrial-internet curve never delivered. The 2018-2020 lighthouse factory/industrial internet story left only ¥694 million in FY25 revenue, 0.26% of total. Baillie Gifford leans into upside in years 3-10, but the successor must be independent and cycle-resistant. FII's four candidates are real and fast, but if North American cloud capex marginally contracts, while the report's current upside assumption is 2026 North American cloud capex > USD 800.0 billion, +67%, orders to 2027, liquid cooling, networking, full racks, and ASIC would fall together. None can take over against the main engine.

  One-sentence close: the second-curve sprouts have indeed broken ground and are growing quickly, but they grow on the same AI capex tree. They represent value-content uplift and margin-mix improvement on the existing curve, which is a real strength, not Baillie Gifford's independent, counter-cyclical, successor curve. CPO is the only seed that might become a new tree, but today it is still a sprout, with timing and share controlled by the same Nvidia/Broadcom roadmap. Given that FII's previous "industrial internet" curve never delivered, its next-curve narrative should be discounted and scored by delivery, not PPT.

  Jun 5, 2026
• QWhat is its core competitive advantage? Will this moat widen or narrow over the next three to five years?4/10

  A

  One-sentence conclusion: FII's moat is vertical integration depth, not irreplaceability. It is the "high-share, high-barrier, low-pricing-power" integration link in AI racks. If Nvidia cuts it off, Nvidia switches to another ODM rather than shutting down, while upstream takes most profit (Nvidia gross margin about 75% versus FII consolidated gross margin of only 6.94%). Over the next three to five years, the moat widens through mix-up and deeper full-stack capabilities, while it narrows through reallocatable share and inherently thin bargaining power. Net judgment: absolute width widens slightly, but relative irreplaceability merely not narrowing is already optimistic. It deserves credit as a share-leading good business, but not a bottleneck-moat valuation premium. This is one of FII's most real dimensions, yet still overestimated.

  ---

  1. First acknowledge the strengths: the moat is real, in three layers

  The moat is not a narrative; it has three measurable layers that must be recognized:

  1. Rack-level system integration share is number one, and it still holds in GB300. In Nvidia GB200/GB300 NVL72 rack outsourcing, FII has about 45-48% share, ranking first, with Quanta about 40% and Wistron about 10%. After Inventec's about 10% entry in GB300, FII is still expected at about 45%, while Quanta falls to about 30%. Independent checks match the report's share figures. Per-unit value is about USD 3.00 million for GB200, about USD 4.50 million for GB300, and Rubin may reach USD 5.50 million-6.00 million, so generation value uplift is a structural tailwind. This leadership is real.

  2. Full-stack vertical integration is the strongest and hardest-to-copy layer. FII is among the few vendors that can self-develop modules -> substrates -> high-speed switches -> liquid cooling -> whole systems -> data centers and provide L10-L12 one-stop rack delivery. Nvidia's shift from "selling chips" to "selling full racks" strengthens rather than weakens ODMs with rack-level system design capability. Quanta's point is right: ODM value is not just the board; it includes system design, cost-saving ability, and services, and the number of vendors that can play shrinks. Full-rack assembly is complex, power-intensive, and resource-scheduling heavy, so entry barriers are rising. This is the core logic for moat widening over three to five years.

  3. Self-supplied liquid cooling and networking move the value chain upward. FII self-develops cold plates/CDU/Manifold, with liquid-cooling gross margin about 15%, far above full racks at about 6%. Domestic cold-plate share is about 40%, and CDU share leads; liquid-cooling orders extend to 2027. In communications networking, 800G+ high-speed switch revenue rose more than 13 times YoY. GB300 fully moves to water cooling, making liquid cooling standard rather than optional, and FII's self-supply pulls margin mix upward. This is the second leg of widening and one of the few real levers to lift outsourcing gross margin above 6%.

  ---

  2. Fatal weakness: high share and high barriers, but low bargaining power. It is not a bottleneck; it is a replaceable integration link

  Apply the strengths above to the real Baillie Gifford Q4 question: does the moat make it indispensable and convert into pricing power? The answer is: the moat is real, but it is not a shutdown bottleneck, and value capture is not in FII's hands. Three hard pieces of evidence:

  1. Upstream takes most profit; share does not equal pricing power. Rack value is high, with GB200 about USD 3.00 million, but GB200/GB300 full-rack outsourcing gross margin is only about 6% and net margin 3-3.5%, while FII consolidated gross margin is 6.94% and Nvidia gross margin is about 75%. In the same rack value pie, pricing power belongs to Nvidia in GPUs, memory suppliers in HBM, and Broadcom in networking chips. FII earns processing fees. The wider the moat and larger the volume, each unit still earns a thin processing fee. That is the outsourcing model's ceiling, and number-one share does not change it.

  2. Competitors are same-order peers whose share can be reallocated, not excluded. Quanta, Wistron, Inventec, and FII are vendors in the same share pool that Nvidia/cloud vendors can reallocate. In the GB300 generation, Inventec enters and Quanta's share falls from about 40% to about 30%, proving share is dispatchable by Nvidia, not locked by FII. If Nvidia cuts off FII, the cost is shifting to another ODM and losing some transition efficiency; it is not full-chain shutdown as with HBM/CoWoS/GPU. FII's moat is vertical integration depth, meaning others need time to fill the stack, not irreplaceability, meaning it cannot be done without FII. This is the sentence that matters most in Q4.

  3. R&D is only about 1.4-1.8% of revenue, so the moat rests on manufacturing know-how, certification, and capacity, not patents. These barriers, including full production lines, high North American cloud customization, long certification cycles, and preference for Taiwanese ODMs, are real but erodible by time and capital. They can block a new entrant from taking orders tomorrow, but not a cloud vendor from building ODM direct sourcing three years later.

  ---

  3. Three-to-five-year two-way pressure: widening versus narrowing

  Dimension	Widening (moat deepening) ✓	Narrowing (moat erosion) ✗
  Share	GB300 still holds about 45%; Rubin value rises again; first-mover and capacity advantages continue	Share is dispatchable by Nvidia, and Inventec has entered; cloud vendors actively multi-source to diversify supply chains, so share has a natural ceiling
  Mix	Self-supplied liquid cooling/networking can lift gross margin toward 8%+ (GB300 full water cooling becomes standard; liquid-cooling margin about 15%)	Core full-rack gross margin still falls back to about 6%; whether mix improvement offsets outsourcing deflation is uncertain
  Full-rack trend	Nvidia's shift to selling full racks strengthens ODMs with system design capability; fewer vendors can play	Quanta/Wistron also have similar capabilities; this is head-end concentration, not FII exclusivity
  Second customer	Google is discussing ASIC chip plus server outsourcing with FII, opening a second growth pole beyond Nvidia	self-developed ASIC grows at 44.6% CAGR, far above GPU's 16.1%; cloud self-developed chips also come with self-build/multi-source ODM tendencies, so ASIC-ization is both an opportunity and a threat that bypassing Nvidia may also bypass FII
  Bargaining power	Almost no widening logic; larger volume can spread costs and improve ROE	Outsourcing bargaining power is naturally thin; upstream GPU/HBM/network chips hold pricing, and that does not change in three to five years

  Net judgment: absolute width, defined by share, value, and mix-up, widens moderately over three to five years, and that should be credited. But the nature of the moat, a replaceable thin-bargaining-power integration link, merely not narrowing is already optimistic. Share can be reallocated, cloud ASIC-ization and multi-sourcing are structural headwinds, and outsourcing bargaining power is built into the model. Against Baillie Gifford's standard for a fivefold 10-year great growth stock, the missing piece is exactly the most important one: converting scale advantage into pricing power and into indispensable bottleneck status. FII has number-one share and real integration depth, but it has not converted those into a thick-profit moat. That is the root cause of "real performance acceleration, thin business, expensive valuation" coexisting, and it supports a Watch rather than Buy stance on Q4.

  Jun 5, 2026
• QIf its core business is disrupted, does it have the DNA to reinvent itself? How does it handle mistakes and bad news?4/10

  A

  Conclusion upfront: FII's "reinvention" looks more like a lucky, passive turn into a tailwind than an active corporate gene. Across its two historical narrative shifts, the only one that truly became profit, AI servers, relied on the same hardware outsourcing base happening to meet a new buyer; it was not self-disruption. Its handling of bad news is split: it is candid on operating issues such as cash-flow collapse and inventory build, but when bad news touches its own narrative and governance, such as failed industrial-internet promises, institutional selling, and fairness of parent-subsidiary overlap, it uses neutral language or selective silence. For a Baillie Gifford 10-year holding, the lack of evidence for active self-disruption is the main deduction.

  ---

  1. Reinvention DNA: across three phases, only one delivered, and it was passive

  Split the story into industrial-internet narrative -> quiet period -> AI server engine. Evidence of active reinvention is thin.

  1. First shift, 2018-2020 industrial internet/lighthouse factories: active narrative, failed delivery. Around the IPO, Terry Gou loudly declared that Foxconn was no longer an outsourcing factory. The company did invest in R&D, changed leadership, and launched a lighthouse program to help 10 lighthouse factories and 100 digital transformations over 5 years. But this active transformation never became a profit engine. Industrial-internet revenue never scaled: 2024 full-year revenue was only ¥940 million, down -42.92% YoY, and FY2025 fell further to ¥694 million, 0.26% of total revenue. Long-running criticisms of old wine in new bottles and smart factories still employing more than 200,000 workers accompanied the company; the stock fell from its post-listing highs and broke issue price at ¥13.77 as early as 2018-10. This was a transformation of active narrative plus objective failure. It proves willingness to tell a new story, not execution or delivery. The gap between story and delivery matters.

  2. Second shift, 2024-2026 AI server engine: real delivery, but passive. This time the performance is real: cloud computing first exceeded half of revenue in FY2024; FY2025 cloud computing revenue was ¥602.679 billion (+88.70%); AI server revenue rose more than 3 times YoY, pushing total revenue in two years from ¥476.3 billion to ¥902.887 billion (+48.22%) and net profit attributable to parent to ¥35.286 billion (+51.99%). But the driver is exogenous. FII's HPC/AI server products were already launched in 2017, sat unnoticed for years, and only found a second spring when Nvidia plus North American cloud capex rose. A third-party review put it bluntly: the company could never have imagined that after years of laying out smart manufacturing and industrial internet, AI, something seemingly unrelated, would give it a second spring. This is catching a tailwind, not actively disrupting itself.

  Key judgment: this "reinvention" did not change the business model. From general server outsourcing to AI rack outsourcing, FII uses the same low-margin assembly plus high asset turnover outsourcing base, only with more valuable Nvidia racks as the customer/product. Consolidated gross margin is still only 6.94%, and cloud computing segment margin is lower at 5.73%. It did not reinvent itself into a higher-bargaining-power species: not chips, not HBM, not software. It simply received a larger order in the same thin business. Real reinvention should change the nature of the business; FII changed the order source. By Baillie Gifford's active self-disruption standard, this is closer to cyclical elasticity than reinvention capability.

  2. How it handles mistakes and bad news: split candor

  Looking at two types of bad news, management shows a clear gap:

  ▶ Operating bad news: candid, even proactive. The most questionable FY2025 operating cash flow of ¥5.238 billion (-78.01%) was not dressed up. The annual report directly attributed it to strong AI server market demand and substantial inventory build, while disclosing inventory +76.99% YoY and cash paid for goods of ¥857.254 billion (+55.21%). Later, Q1'26 operating cash flow jumped to ¥25.024 billion, about +1826% YoY, as collections arrived. This shows the cash-flow collapse was staged inventory occupation, and management explained the logic in the period. On GB200 yield and GB300 transition timing, chairman Benjamin Cheng also gave relatively specific guidance: GB200 yield improved, GB300 had clear shipment demand from major customers, and customers would not cut orders because of the GB200/GB300 transition. This layer of disclosure is acceptable.

  ▶ Narrative/governance bad news: selective silence. When bad news touches whether its own narrative delivered or whether governance is fair, candor drops sharply:

  • The failed industrial-internet promise is handled silently. This was the core IPO story sold to the market, and the segment has shrunk to 0.26%. Yet the FY2025 annual report did not separately explain or flag the shrinkage of the industrial-internet business. At company level, it uses neutral wording such as focusing on new industrial directions and actively developing a second growth curve, with no review saying the earlier story failed. Quietly replacing a failed old narrative with an attractive new one is exactly the narrative packaging Baillie Gifford is wary of.
  • Institutional exit and retail taking over are not proactively addressed. Fund holdings fell from the historical peak of ¥42.6 billion to ¥12.347 billion in 2026Q1, a cumulative reduction of more than 70%, and the number of funds fell to 318, while shareholder accounts surged from about 476,000 to 720,000 in half a year as retail investors took over. This is strong bad news: professional capital voting with its feet at high valuation. But it appears in third-party deep reports, not as a company communication topic.
  • The fairness of parent-subsidiary same-business overlap receives formulaic responses. The Hon Hai system holds about 84.8%, and Hon Hai itself is also the largest Nvidia GB-series outsourcer. Parent and subsidiary both do AI servers, with orders, capacity, and pricing allocated within the system; related-party purchases were ¥17.670 billion and related services received ¥7.077 billion. On this structural issue that most harms minority shareholders, company responses are often formulaic, saying it has relevant replies, with no substantive disclosure breaking down conflicts of interest.

  3. Complete the scenarios: if the core business is disrupted, does it have the capital to reinvent?

  For Question 5, three disruption scenarios must be tested against its self-rescue gene:

  1. AI capex peaks or slows. This is its most fragile scenario. FII is highly tied to North American cloud capex (2026 cloud capex revised up to > USD 800.0 billion, about +67% YoY). If marginally contracted, the low-margin high-volume outsourcer would suffer amplified downside and a Davis double-kill. It has no second growth pole to switch to: industrial internet has been disproved at 0.26%, and communications equipment segment growth is only +3.46%. History tells us that when it last lost its main engine, in 2021-2023, with 2023H1 revenue down -8.21%, the first half-year decline since listing, it waited by luck for AI rather than creating a new engine itself. That is counter-evidence to a strong reinvention gene.

  2. GPU generation transition gaps, GB200 -> GB300 -> Rubin. Short-term risk is controllable: GB300 can reuse existing infrastructure, ODM test results are optimistic, and transition is expected to be smooth; management also says customers will not cut orders. But Rubin, including more aggressive architecture jumps such as CPO/Kyber, is the real test. The greater the architectural change, the more production lines and yields reset, and transition demand gaps plus share reallocation, with GB300 already adding Inventec at about 10%, can hit results. FII's response remains passive adaptation by following Nvidia, not defining the next generation.

  3. Outsourcing migrates to cloud self-build or other vendors. This is a structural threat. FII's moat is vertical integration depth, from modules to liquid cooling to full systems, not irreplaceability. Quanta, Wistron, and Inventec are same-order competitors whose share can be reallocated; cloud self-developed ASICs and order diversification are long-term trends. It has no shutdown-bottleneck status. Once share is cut, it lacks the capital to reinvent upward into chips/HBM or sideways into software.

  In sum for Question 5: FII has the willingness to change stories, but not the ability to change species. It can actively tell new narratives, industrial internet and AI, but the one that delivered was passively catching a tailwind and did not change the low-margin outsourcing nature. On bad news, it is candid on operating data, cash flow, inventory, and yield, but selectively silent on narrative delivery and governance fairness, including industrial-internet failure, institutional exit, and same-business overlap. For a growth stock meant to be held through years 3-10, this passive reinvention plus split candor means: when the AI wind stops, there is no evidence it can self-disrupt and build another curve like a great growth company. This dimension is honestly a weakness.

  Jun 5, 2026
• QDoes management, especially the founder, have a long-term vision and deep alignment with the company? Is it willing to sacrifice current profit for five to ten years from now?3/10

  A

  Conclusion upfront

  Question 6, management long-termism, alignment, and willingness to sacrifice current profit for 5-10 years later, is almost FII's weakest dimension under the Baillie Gifford LTGG framework, and it is thoroughly weak. Baillie Gifford's ideal is a founder-level figure with personal wealth heavily invested, in the same boat as minority shareholders, able to tolerate an ugly current income statement for a 10-year outcome. FII is the opposite: a structure of absolute control by the controlling-shareholder system, about 84.8%, the company claiming no actual controller, professional managers rotating at the helm, and potential conflicts from parent and subsidiary both making Nvidia AI servers. This is not a founder breathing the same air as shareholders; it is textbook governance discount. The layers below use independently checked facts.

  ---

  1. Equity is highly concentrated in the Hon Hai system, free float is extremely low, and minority shareholders have almost no voice

  Hon Hai Precision (2317.TW) controls FII through multiple entities: controlling shareholder China Galaxy Enterprises Ltd holds 36.73%, Futaikang Industrial (Shenzhen) 21.98%, Foxconn Technology Group 8.24%, and other related entities including Hongfujin Precision Electronics and Argyle Holdings about 17.85%, giving about 84.8% on the full "under common Hon Hai control" basis. The truly free-floating equity in the market is only a bit above 10%. The report's about 84.8% matches this. One honest basis note: if only China Galaxy, Futaikang, and Hongfujin Zhengzhou are counted directly, the narrow basis is about 61.71%. Both numbers are correct; the difference is whether affiliates also under Hon Hai control are included. But whether 61.71% or 84.8%, the conclusion is the same: the controlling party calls the shots, secondary-market float is thin, and minority shareholders can mostly only watch on many related-party proposals. That is not the Baillie Gifford preference for founder and outside minority shareholders being deeply aligned. It is one-sided control by the controlling-shareholder system, not the same boat.

  2. The company says it has no actual controller, so the alignment story breaks at the source

  FII explicitly states in announcements: Hon Hai Precision indirectly holds 100% interest in controlling shareholder China Galaxy, but because Hon Hai Precision itself has no actual controller, the company also has no actual controller. This is fatal to the Baillie Gifford framework: there is no identifiable founder/actual controller with personal net worth staked on this listed company. The Hon Hai system itself, whose founder Terry Gou has stepped down and whose current chairman and general manager is Young Liu, is also a dispersed governance structure with no actual controller. Thus the central Baillie Gifford axis of founder-level long-term vision plus economic alignment simply has no carrier at FII: above is a no-controller outsourcing empire; below is a no-controller listed subsidiary.

  3. The current helmsman is a professional manager plus rotating CEO, not a founder story

  Chairman and general manager Benjamin Cheng was elected on 2023-07-07, is from Taiwan, born in 1966, graduated in electronic engineering from Chung Yuan Christian University, has served as FII general manager since 2017, and has nearly 20 years at Foxconn. He is a typical internally promoted professional manager, not a founder whose personal wealth is tied to the company. Since 2024-06, the company has also formally implemented a rotating CEO system; the first rotating CEO is chief data officer Liu Zongchang, born in 1989, joined only in 2019, age 36, with a one-year term. A rotating system can be neutral or positive governance, but in Baillie Gifford context it further shows that this is a company operated by a professional management team appointed by the controlling shareholder, with replaceable leaders whose tenure and personal holdings do not constitute a founder's long-term sacrifice commitment. Public holdings show management's personal alignment is almost negligible relative to a market cap of about ¥1.56 trillion and price of ¥78.55.

  4. Parent-subsidiary same-business competition: Hon Hai itself is Nvidia GB-series' largest outsourcer, and orders/pricing are allocated inside the system

  This is the most concrete conflict in the alignment dimension. FII is the number-one assembly supplier for Nvidia NVL72 rack-level AI servers, with report share of about 45-48%, but the parent Hon Hai Group as a whole is the largest Nvidia GB-series AI server outsourcer. In 2025 GB200/GB300 rack server ODM shipments, Hon Hai accounts for more than half, up to about 52%; Foxconn takes most Nvidia GB300 orders, with share of about 40% to more than half; and Hon Hai has built the world's largest GB200 plant in Mexico and laid out GB300 lines in Houston. In other words, the same Nvidia AI server pie is made by the listed company FII and by unlisted entities inside the system. How orders, capacity, and pricing are divided between the listed subsidiary and other parent entities is decided by the controlling shareholder, not minority shareholders. A-shares require listed subsidiaries and parents not to compete in the same business. FII was originally the product of carving out high-end business for listing, but on the fattest AI server track, the parent and subsidiary cooperate and support each other while naturally overlapping. Whether good orders stay in the listed company or elsewhere in the system creates a structural possibility of shifting benefits toward the controlling shareholder rather than minority shareholders, exactly the misalignment Baillie Gifford is wary of.

  5. Related-party transactions are large and growing fast; alignment points to the system rather than minority shareholders

  The scale is material and rising, all from official announcements (FII 2026 annual recurring related-party transaction forecast announcement, No. 2026-013):

  • 2025 approved quota, consistent with the report: purchases from related parties about ¥17.670 billion and services received from related parties about ¥7.077 billion;
  • 2025 actual amount, honestly slightly below quota: purchases about ¥16.924 billion, services received about ¥6.527 billion, sales to related parties about ¥6.011 billion, total related-party transactions about ¥30.3 billion for the year;
  • 2026 forecast jumps further: purchase quota raised to about ¥24.647 billion, services received about ¥8.787 billion, total about ¥43.3 billion, with the announcement saying transaction scale is expected to increase in future periods, and services received already account for about 24% of that business category.

  The company says related-party transactions are on equal, mutually beneficial, paid-at-fair-price terms and do not harm the company or especially non-related shareholders and minority shareholders, with pricing preferentially referencing independent third-party market prices/comparable unrelated transactions, or cost-plus when no comparable exists. But the key is: under 84.8% one-sided control plus parent-subsidiary same-business overlap, the larger and faster-growing related-party transactions become, the harder it is to verify fairness. Minority shareholders are left to trust that the controlling shareholder will be fair. That is the opposite of Baillie Gifford's preference for alignment rather than trust.

  ---

  Overall judgment, against Baillie Gifford LTGG preferences

  Baillie Gifford core management requirement	FII reality
  Founder-level long-term vision, personal heavy stake	Absent: no actual controller; helm is professional manager plus rotating CEO; personal alignment is weak
  Economic interests breathe with minority shareholders	Misaligned: controlling-shareholder system about 84.8%, free float about 10%, parent-subsidiary same-business overlap plus related-party transactions may tilt benefits toward the system
  Willingness to sacrifice current profit for 5-10 years later	No evidence: cash-flow and capacity decisions serve the AI capex cycle and system coordination; no founder-style long-term tradeoff narrative is visible

  In one sentence: on Question 6, FII does not offer Baillie Gifford the story of a founder in the same boat with shareholders, willing to suffer today for 10 years later. It offers a governance discount: controlling shareholder absolute control, no actual controller, professional managers, and potential parent-subsidiary conflicts. This does not necessarily mean poor operations; performance is in fact accelerating and ROE is 21.65%. But on deep management-minority alignment and willingness to sacrifice short-term profit for the long term, a dimension Baillie Gifford values heavily, FII is clearly weak and should command a valuation discount, not narrative elevation.

  Jun 5, 2026
• QIf it disappeared tomorrow, how much would customers miss it? Is its growth sustainable and not dependent on harming society or regulation?4/10

  A

  Conclusion upfront: FII has weak indispensability and medium sustainability. It is the share-leading AI compute outsourcer with real switching costs, but at its core it is the first among same-order competitors, not an exclusive bottleneck like Nvidia or HBM. Its growth activity, building AI compute infrastructure, has positive social utility and does not touch red lines such as tobacco, alcohol, gambling, or data abuse. But it is highly dependent on the AI capex cycle, about 34% of capacity in Mexico is exposed to USMCA/tariff risk, R&D is only about 1.4-1.8% of revenue, and parent-subsidiary same-business overlap is a governance flaw. It is not growing by harming society or exploiting regulation, but it is far from steady, self-driven growth that can independently cross cycles. Baillie Gifford wants indispensability where customers lose sleep if the company is cut off. FII cannot give that answer. It gives "customers need several quarters and switching costs to reallocate share".

  ---

  1. If FII disappeared tomorrow, how much would customers miss it? A leader, not an exclusive bottleneck

  Honest characterization: FII is the first among reallocatable same-order competitors, not an exclusive bottleneck like Nvidia/HBM. Customer dependence has two layers:

  1. Number-one share plus full-stack vertical integration creates real short-term switching cost. In Nvidia GB200/GB300 NVL72 rack outsourcing, FII has about 45-48% share, ranking first, consistent with external research; Quanta about 40%, Wistron about 10%, and after Inventec enters GB300, FII still about 45% (server research notes, AI server expert call). It is one of the few companies able to self-develop the full stack from modules -> substrates -> high-speed switches -> liquid-cooling cold plates/CDU/Manifold -> whole systems -> data centers, and provide L10-L12 one-stop rack delivery. Per-rack value is very high: GB200 NVL72 about USD 3.00 million, GB300 about USD 4.50 million. NVL72 is liquid-cooled rack-level system integration, with long certification cycles, highly customized lines, and North American cloud preference for Taiwanese ODMs. Moving FII's share to someone else in the short term requires real costs in capacity ramp, recertification, and yield rebuilding. This should not be dismissed. Hon Hai's own U.S. plant has ramped to about 1,000 racks per week, targeting 2,000 (Technetbook), so scale barriers are real.

  2. But at its core it is a leader among same-order competitors, and share is already being continuously reallocated. The hard indispensability test is whether alternatives exist and how painful switching is. FII's answer is that alternatives sit next door, and Nvidia is actively nurturing them:

  • Nvidia deliberately diversifies and prevents any one vendor from becoming a bottleneck. GB-series OEM/ODM partners include Foxconn, Quanta, Inventec, and Wistron, a multi-supplier system where avoiding dependence on a single supplier is an explicit strategy (GB300 OEM/ODM partners). Even components are de-exclusive: CDU was exclusive to Vertiv in GB200, but GB300 removed exclusivity and opened to Delta, Boyd, Cooler Master, and others (GB300 liquid-cooling supply chain). At the main-chip level, GB300 remains exclusively manufactured by TSMC. The real bottlenecks are TSMC/Nvidia/HBM; at rack assembly, Nvidia wants replaceability and price pressure.

  • Share is recut every generation, and Inventec is the live example. From GB200 to GB300, Inventec's share rises from about 10% to about 20%, Quanta about 25-30%, and FII remains first but not monopolistic (Inventec GB300 share). The fact that a new player can be added and double share within one generation proves this business is reallocatable, the opposite of a shutdown bottleneck.

  • Weak bargaining power directly disproves indispensability. Truly indispensable links take profit. Nvidia gross margin is about 75%, while FII consolidated gross margin is only 6.94% and cloud computing segment margin only 5.73% (FII 2025 annual report). Upstream captures most profit, which shows FII sits in a replaceable integration position with high share, high barriers, and low bargaining power. If it were truly indispensable, Nvidia would not leave it only 7%.

  Summary: customers would miss FII as the fastest, most stable, most full-stack vendor in the short term. But "miss" means spending several quarters and money to find substitutes, not shutdown. The top five customers account for 62.01%, with no single customer above 50% (FII 2025 annual report), so the binding is two-way and dilutable. This is reallocation cost, not indispensability. In the Baillie Gifford sense, the moat answer is weak.

  ---

  2. Is growth sustainable and not dependent on harming society or regulation? Social utility is positive, but sustainability is discounted

  First, on regulatory/social red lines: it passes. FII outsources AI compute infrastructure: servers, switches, liquid cooling, and data centers. It is a picks-and-shovels provider for the digital economy, with positive social utility. It does not involve tobacco, alcohol, gambling, addictive design, data abuse, privacy trafficking, or similar backlash-prone red lines. Its growth is not based on harming users or regulatory arbitrage. This dimension is clean and should be credited.

  But growth sustainability must be discounted. Its growth is cycle-driven and geopolitically exposed, not self-driven steady state:

  1. Growth equals AI capex cycle, not steady compounding. FY2025 revenue was ¥902.9 billion (+48%), net profit attributable to parent ¥35.286 billion (+52%), and Q1'26 net profit ¥10.595 billion (+102.55%). Performance is truly accelerating. But longitudinally, 2018-2023 revenue hovered at ¥400.0 billion-480.0 billion for six years, and 2023H1 once saw its first half-year decline. It was AI servers from 2024 that pushed revenue in two years from ¥476.3 billion to ¥902.9 billion (Securities Times). This is a capex-driven surge, not cycle-crossing compound growth. Visibility depends on North American cloud 2026 capex > USD 800.0 billion (+67%), GB-series shipments from about 29,000 units in 2025 to about 70,000-85,000 units in 2026, and orders visible to 2027 (Technetbook). If cloud capital budgets marginally contract or a GPU generation transition such as GB300 -> Rubin creates a demand gap, ODM earnings elasticity is amplified both ways. Sustainability depends on others' capex rhythm; FII does not control the growth switch.

  2. About 34% of Mexico capacity sits on USMCA/tariff exposure, a real geopolitical risk with a clear timetable. Manufacturing locations are about 56% mainland China, 34% Mexico, and 9% Vietnam, while overseas assets account for 71.71% (FII 2025 annual report); FII has the world's largest GB200 full-rack base in Mexico (TMTPost). This capacity benefits from USMCA origin rules and exemption from reciprocal tariffs, but that benefit faces structural tightening:

  • The first six-year USMCA joint review starts on 2026-07-01. It will reshape origin rules and sector tariffs; electronics faces increasing scrutiny because of nearshoring expansion, and products with Chinese content will be a focus, directly hitting Hon Hai's supply chain structure of 56% mainland China plus Mexico final assembly (BSI: USMCA Review 2026, Tetakawi: Mexico Tariffs 2026).
  • Most trade analysis expects USMCA to continue with modifications, adding restrictions on China-linked manufacturing, stricter origin requirements, and likely annual-review uncertainty (Mexico Business: annual USMCA reviews). Mexico in 2025-12 already imposed 10-50% import tariffs on 1,463 tariff lines including electronics, with USMCA goods temporarily exempt, showing Mexico's tariff environment is tightening (Tetakawi).

  So roughly one-third of FII's capacity tariff-cost advantage depends on an agreement being renegotiated in 2026-07 and explicitly targeting Chinese content. That is a concrete pillar of growth sustainability that external policy can remove, not a variable FII controls.

  3. R&D intensity and governance make growth weakly endogenous and less fully capturable. R&D is only about 1.4-1.8% of revenue (Tencent News). For a company calling itself technology/industrial internet, this quantifies that it is outsourcing under a technology label. Growth comes from taking orders and expanding capacity, not compounding proprietary technical barriers. That is the thinnest part of sustainability: outsourcing moats do not automatically deepen over time. Add the Hon Hai system's about 84.8% ownership, parent and subsidiary both doing AI server outsourcing, and 2025 approved purchases from related parties of about ¥17.670 billion (FII 2025 annual report). Part of growth is allocated inside the system. For minority shareholders, who captures this growth and whether pricing is fair itself carries a governance discount. This is not harming society, but it discounts external shareholder capture.

  Summary: growth is not built on harming society or regulatory arbitrage, which passes. But growth is cycle-driven, externally exposed, weakly endogenous, and capture-discounted. Sustainability is medium-low, not steady sustainable.

  ---

  3. Overall judgment for Baillie Gifford Question 7

  Dimension	Honest conclusion
  Indispensability	Weak. Number-one share (about 45-48%), full-stack integration, and real short-term switching costs. But at its core it is a leader among same-order competitors, share is reallocated by generation (Inventec GB300 doubled), Nvidia deliberately uses multiple suppliers, bargaining power is weak (gross margin only 6.94%), and it is not a shutdown bottleneck.
  Does growth depend on harming society/regulation?	No, clean. AI compute infrastructure has positive social utility and avoids tobacco, alcohol, gambling, and data-abuse red lines.
  Growth sustainability	Medium-low. It depends on the AI capex cycle rather than steady state; about 34% Mexico capacity is exposed to the 2026-07 USMCA/tariff renegotiation targeting Chinese content; R&D is only about 1.4-1.8%, so endogeneity is weak; parent-subsidiary same-business overlap discounts capture.

  One-sentence close: FII passes the baseline test of not making money by doing harm, but fails the indispensability moat test. It is the largest, most full-stack, and most painful-to-replace AI outsourcing vendor, but ultimately a leader whose share can be reallocated. Growth depends on others' capex rhythm and has one foot on a geopolitical tariff trapdoor. Against Baillie Gifford's yardstick for a great growth stock that can rise fivefold in 10 years, this is strong but not scarce, accelerating but not steady-state, not the kind of indispensability it seeks.

  Jun 5, 2026
• QWhat are the unit economics of this business, including gross margin and incremental returns? Do they improve or deteriorate as scale increases? Where does the money earned go?4/10

  A

  Conclusion upfront: FII's unit economics combine two pictures: typical low-margin outsourcing and mid-to-high ROE supported by turnover. The ceiling for incremental returns is set by outsourcing economics: consolidated gross margin is only about 7%, most profit is captured by Nvidia upstream, and working capital is heavy. But high asset turnover still delivers ROE of 21.65%, a post-listing high, while gross margin has risen sequentially since 2025Q2 as mix shifts toward liquid cooling and networking. Core judgment: this is more "low-margin, high-volume" than true unit-economic improvement. Mix-up is real but limited; a few percentage points of gross-margin improvement cannot change the fact that it earns processing fees. The money earned mainly goes into capacity expansion and inventory. FY2025 free cash flow was negative, and the gap was filled by external financing. Baillie Gifford LTGG prefers high unit economics that keep improving with scale. FII only touches "slightly improving with scale" and fails on "high".

  ---

  1. Weak side, the main contradiction from Baillie Gifford's perspective: incremental economics are compressed by outsourcing model and upstream bargaining power

  1. Absolute gross margin is low, and this is an industry structural ceiling, not only FII's problem. FY2025 consolidated gross margin was about 6.94% by company annual-report basis; third parties often disclose 6.97%, +0.25pct YoY. The two are basically equivalent, differing by rounding and statistical boundary. Net margin was 3.91%. The key is that this is not FII mismanagement. Independent checks show industry gross margin for GB200 NVL72 full-rack assembly is about 8%, and net margin is about 4% of selling price; GB300 mass production also returns only to 6-8%. In other words, rack-level AI server outsourcing unit economics top out around this level. FII's 7% is what this business can earn, not a capability discount. That proves the ceiling is set by outsourcing economics, not company effort.

  2. Upstream takes most profit, and bargaining power differs by an order of magnitude. Independent checks of Nvidia's official disclosure show FY26 data-center GAAP gross margin 73.4%, Non-GAAP 73.6%, with company target in the mid-70% range. The report's "about 75%" is close, slightly high; more precise is about 73-74%. A more direct value-capture comparison: a single Nvidia GPU sells for about USD 30,000-40,000 with very high gross margin, while the full rack system of about USD 2.00 million-3.00 million leaves ODM partners only "healthy revenue". On the industry basis above, a roughly USD 3.30 million NVL72 gives FII only about USD 130,000 net profit in this link, about 4%. In the same rack, chips take more than 70% gross margin while assembly earns a 4% processing fee. That is the fundamental difference between a shutdown bottleneck and reallocatable share.

  3. R&D intensity is below what the "technology company" narrative requires, even below the report's range. Independently checked: FY2025 R&D expense was ¥11.151 billion, +4.89% YoY, about 1.24% of revenue. This is actually below the report's "about 1.4-1.8%" range, because the +48% revenue base diluted it while R&D rose only +4.89%. R&D growth is far behind revenue growth. This conflicts with the growth-stock picture of technology lifting unit economics over time and better matches outsourcing driven by scale and processing fees. Revenue doubled in two years while R&D ratio fell, which is hard evidence that unit economics did not materially upgrade.

  4. Working capital is heavy, growth consumes cash, FY2025 free cash flow was negative, and external financing filled the gap. FY2025 operating cash flow was only ¥5.238 billion, -78.01% YoY, mainly due to large AI server inventory build; inventory rose from ¥85.266 billion to ¥150.913 billion (+76.99%). Add investing cash outflow of about ¥29.746 billion for capacity expansion plus term deposits, and free cash flow was negative. An independently checked fact not emphasized by the report, but crucial to unit economics: the gap was filled by financing cash inflow of ¥47.397 billion, turning from prior-year outflow to inflow. A meaningful part of growth working capital came from external debt rather than internally generated cash. That is the typical cost of low-margin, high-turnover outsourcing: incremental profit looks fine, but large working capital and capex must be committed first, so incremental free-cash returns are much lower than incremental accounting profit.

  ---

  2. Strong side, which must be acknowledged honestly: low net margin times high turnover equals mid-to-high ROE, and mix is moving up

  1. ROE of 21.65% is not low and is a post-listing high, the correct picture of outsourcing. FY2025 ROE reached 21.65%, +5.80pct YoY, the highest since listing; total assets were ¥456.224 billion (+43.68%). Net margin is only 3.91%, yet ROE reaches 21.65% because of high asset turnover plus moderate leverage. This is the standard outcome of a low-margin, high-turnover business. Baillie Gifford asks whether unit economics are good. The honest answer: measured by net margin, they are poor; measured by ROE, they are upper-middle. Both are true, and the whole debate is whether turnover can remain this high. That turnover depends on full orders and fast inventory clearing during an AI capex upswing. If the cycle weakens, turnover declines and ROE falls with reverse amplification.

  2. Gross margin has risen sequentially since 2025Q2, and mix-up is real. The checked quarterly path: gross margin climbed quarter by quarter after 2025Q2; Q3 rose +0.5pct QoQ to about 7%, and Q4 net profit was ¥12.799 billion. By 2026Q1 it rose to about 7.35%, consistent with the report's 7.4%. Drivers are mix-up: cloud computing, mainly AI servers, has gross margin only 5.73%, +0.74pct YoY; communications network equipment has 9.28%; self-developed liquid cooling has gross margin about 15%, with some solutions at 15-18%, well above full racks at about 6.5%. Add 800G switch Q3 single-quarter revenue up more than 27 times YoY. Higher-margin liquid cooling and high-speed networking are pulling consolidated gross margin upward. That is real evidence of marginal unit-economic improvement and should not be erased.

  3. Q1'26 operating cash flow reversed dramatically, proving working-capital occupation was staged and can flow back. 2026Q1 operating cash flow was ¥25.024 billion, +1826.20% YoY, versus only ¥1.299 billion in the prior-year period. FY2025 inventory prebuild turned into sales collections in Q1. The balance sheet is not tight: cash was ¥110.025 billion (+51.90%), broadly matching short-term borrowings of about ¥104.2 billion and close to net cash. So the cash-flow collapse was a timing issue, not structural bleeding. But heavy working-capital occupation itself is a long-term feature of outsourcing; each volume ramp repeats the cycle of consuming cash first and returning it later.

  ---

  3. Overall judgment: low-margin high-volume or real unit-economic improvement? Mainly the former, with a touch of the latter

  Putting both sides together, FII's scale expansion is still closer to low-margin high-volume. Consolidated gross margin moves from 6.97% toward 7.4%, R&D ratio falls to 1.24%, and incremental returns require large working capital and external financing. These all point to unchanged unit-economic nature. Mix-up through 15% liquid cooling and 9.28% networking is real marginal improvement, but limited in magnitude. Under the 6-8% industry ceiling for AI server outsourcing, a few percentage points of consolidated gross-margin improvement cannot change the business attribute of earning processing fees while upstream takes more than 70% gross margin. The money earned is not mainly reinvested into high-return capabilities that lift unit economics, since R&D grew +4.89% versus revenue +48%; it mainly goes into capacity expansion and AI server inventory. That capital allocation serves larger volume, not higher unit economics.

  The direct answer under Baillie Gifford LTGG: it seeks high and continuously improving unit economics. FII barely touches "slightly improving with scale" through real mix-up, but clearly fails on high, with net margin 3.91%, negative incremental free cash flow, and weak bargaining power. ROE of 21.65% is its strongest rebuttal, but that is high turnover compensating for low margin, dependent on the AI capex upswing, not a moat of excellent unit economics. In one sentence: this is a high-turnover outsourcing business whose ROE is not ugly, but whose unit economics are not great. Scale mainly multiplies thin profit by larger volume rather than making thin profit thick.

  Jun 5, 2026
• QWhat conditions must all be true for it to rise fivefold in 10 years? Are these conditions realistic? What expectations are embedded in today's share price?2/10

  A

  Conclusion first: for FII to rise fivefold in 10 years, from ¥78.55 to about ¥393 and from about ¥1.56 trillion market cap to about ¥7.8 trillion, or about 17.5% annualized, four low-probability events must hold simultaneously and continuously for 10 years: uninterrupted AI capex supercycle, share maintained or improved, structural gross-margin uplift, and no major valuation derating. These conditions are not independent and even drag on each other. This is not a mispriced growth stock the market has not discovered; it is a low-margin outsourcing leader with real accelerating earnings, but whose valuation has fully priced AI and whose downside risk has already been rehearsed in real trading, from the 2025-10 high of ¥83.42 to the 2026-04 low of ¥47.35, a maximum drawdown of more than 43%. Today's price embeds not a fivefold expectation, but an optimistic extrapolation that FY26 net profit grows another 45-70% and high growth can persist for many years. That expectation has already reached, even slightly exceeded, the upper end of sell-side fair-value ranges. Treating fivefold as the base case is not honest.

  Basis note: this answer uses the unified anchor of ¥78.55 and market cap about ¥1.56 trillion as of 2026-06-04. It should be noted that the independently checked latest transaction price had fallen to about ¥73.40 and market cap about ¥1.48 trillion (Sina/investing.com, 2026-05-29), about 6% below the anchor. This does not change the conclusion; it only makes the high-valuation starting point more concrete.

  ---

  1. First do the math: what does fivefold require?

  Fivefold from ¥78.55 means ¥392.75, taking market cap from about ¥1.56 trillion to about ¥7.8 trillion, or about 17.5% annualized over 10 years. The number itself is not absurd for an exceptional growth company. The problem is the starting point: FII is already the fifth-largest A-share company by market cap. In 2025-10, market cap once reached ¥1.6 trillion and surpassed Kweichow Moutai (Sina Finance, 2025-10-30). For an already trillion-yuan company at historical valuation highs to rise fivefold to nearly ¥8 trillion, close to today's A-share ceiling, both profit growth and valuation must avoid becoming a drag. Yet both have already been stretched.

  Split fivefold into the two engines of profit and multiple:

  • If valuation multiple is maintained, with static PE ~44× and forward PE ~26-30× unchanged, net profit must grow from FY25's ¥35.286 billion to about ¥176.5 billion, roughly 5 times in 10 years, about 17.5% annualized. For an outsourcing business with gross margin only 6.94% and net margin 3.91%, this implies revenue also needs to multiply to about ¥4.5 trillion at current net margin, roughly rebuilding several current FIIs.
  • If valuation expansion shares the burden, that means betting against mean reversion: a stock already above the historical 85%+ percentile (eniu.com, historical average about 18×, range about 8×~60×) must keep its multiple rising, against the gravity that outsourcing ultimately returns toward peer-center 18-20×.

  Both paths require optimistic assumptions to hold for 10 consecutive years. Now test them one by one.

  ---

  2. Conditions needed for fivefold, pressed honestly one by one

  Condition 1: the AI capex supercycle continues uninterrupted for 10 years, and growth cannot decelerate too fast. This is the only condition that is clearly true and strong today. North American four-cloud 2026 capex has been revised up to > USD 800.0 billion, about +67% YoY; Bank of America expects 2027 to exceed USD 1 trillion, but growth slows to about +25%, with 70%+ going to AI compute infrastructure (GuBaike citing BofA, 2026-05). Order visibility extends to 2027. But two honest points matter: 1. this is a capex cycle, not steady growth. The move from +67% to +25% already shows the first derivative is down; linearly extrapolating a supercycle for 10 years is the weakest assumption in the bull case. 2. There has already been discussion of Microsoft shrinking capex, although Goldman reassured that the USD 80.0 billion guide was unchanged (Zhitong Finance). Marginal signals are flashing. If capex returns from supercycle to high but not growing, ODM earnings elasticity cuts downward with amplification.

  Condition 2: share is maintained, even improved. FII has about 45-48% share in GB200/GB300 NVL72 full-rack outsourcing, ranking first, with Quanta about 40% and Wistron about 10% (server research notes). This is a real barrier through full-stack self-development and liquid-cooling self-supply. But share is reallocatable: GB300 has introduced Inventec to take share, while Quanta and Wistron remain same-order competitors, not replaced players. FII's moat is vertical integration depth, not a shutdown bottleneck like Nvidia/HBM. It has high share, high barriers, and low bargaining power. Maintaining 48% is already a fully loaded leader case; "improvement" has limited room, while erosion risk persists. This condition's best case is holding steady, and holding steady is necessary but not sufficient for fivefold.

  Condition 3: gross margin structurally moves up enough to digest the valuation. This is the key leap from "low-margin high-volume" to "high valuation is justified", and it is the hardest leap. Reality: consolidated gross margin is only 6.94%, cloud computing segment gross margin is even lower at 5.73%, and upstream Nvidia, with gross margin about 75%, takes most profit. Bulls hope liquid cooling, with about 15% gross margin, plus 800G/1.6T networking and CPO can pull consolidated gross margin toward 8%+. There are early signs: gross margin has risen sequentially since 2025Q2, and Q1'26 reached about 7.4% (Sina). But moving from 7% to a level that supports technology-stock valuation is structural reinvention, not cyclical improvement. As long as full-rack assembly remains the revenue core, outsourcing economics set the ceiling. Direction is right; slope is far from enough.

  Condition 4: valuation does not materially derate, and the market has already disproved this once with real money. This is the most damaging counterexample in the fivefold narrative, and it is not hypothetical; it already happened. After FII reached the historical high of ¥83.42 in 2025-10, it fell within 5 months to a low of 47.35 yuan, with a maximum drawdown of more than 43%. Institutional holding value fell from ¥42.577 billion in 2025-09, held by 659 funds, to ¥12.347 billion in 2026Q1, held by 318 funds, down more than 70% cumulatively, while shareholder accounts rose from about 476,000 to about 720,000: institutions exited and retail took over (TMTPost deep dive, 2026). The stock later rebounded to today's ¥73-78. In other words, a Davis double-kill is not a tail risk; it is this stock's actual script over the past half year. Earnings were still accelerating, with Q1'26 net profit +102%, yet the stock could halve and recover. For a high-position ownership structure that has just shown a 43% fall, betting that valuation will not materially derate for the next 10 years while adding a fivefold rise is betting against its own history.

  The four conditions interact badly: if capex marginally weakens, condition 1 breaks; share value comes under pressure, condition 2; high-margin mix rollout delays, condition 3; valuation immediately reverts toward outsourcing peers, condition 4. The four kill in the same direction. Fivefold requires all four to be true at the same time for 10 consecutive years; reality is more likely one breaks first, then dominoes.

  ---

  3. What does today's price actually imply?

  Not fivefold, but an optimistic extrapolation that FY26 net profit grows another 45-70% and high growth persists for years, already priced near the upper end of the fair range.

  • Static PE ~44×, based on FY25 EPS ¥1.78, stands at a post-listing historical high, in the 85%+ percentile over the past 5 years, and at a clear premium to Taiwanese peers (Quanta ~20.3×, Wistron ~17.7×, Wiwynn ~18.7×) and U.S. peers (SMCI/Dell ~19×). PE-TTM ~38-40× (TMTPost basis about 39.57×).
  • The basis that makes it "look not expensive" depends entirely on forward numbers and optimistic extrapolation. Forward PE ~26-30× is calculated using FY26E net profit around the broker consensus of ¥51.5 billion, with Goldman at ¥52.161 billion, Tianfeng raised to ¥50.0 billion, and China Securities even ¥67.6 billion, implying same-year PE only about 16× (Tencent News citing brokers), or YoY +45-70%, hence PEG<1. But PEG<1 is only valid if this +45-70% growth can continue, exactly what conditions 1-3 require. Today's price has already priced in AI fully; it implies high growth does not slow, not that fivefold upside remains.
  • The more direct evidence: current price has reached or slightly exceeded the upper end of broker fair ranges. Several brokers use FY26 net profit of ¥51.5 billion and 24× PE to derive fair market cap of about ¥1.236 trillion and fair price about ¥62.24, with target tiers of ¥62/70/78 (Eastmoney citing brokers). One broker range maps to ¥62-78. Thus ¥78.55 already sits at the ceiling of the most optimistic tier. Consensus sell-side average target may be about ¥85-92, with high target ¥101, but one independent analyst puts the low end at ¥56.4 (investing.com consensus). The wide dispersion itself says this is not a mispriced cheap stock.

  Put together: buying FII today is not buying an overlooked fivefold opportunity cheaply; it is paying a premium near the upper end of the fair range for whether optimistic extrapolation can be delivered. This is the opposite of Baillie Gifford LTGG's search for great growth stocks the market has not yet understood. The market has noticed, and has priced the optimistic case at the upper end.

  ---

  4. Honest answer to Question 9

  • Is a fivefold rise in 10 years realistic? As a base case, no. It requires four mutually dragging optimistic conditions to hold together for 10 consecutive years, while the starting point is already trillion-yuan market cap, historical valuation highs, and recent major institutional exit. As an extremely optimistic tail case, with AI capex becoming a true 10-year supercycle, share rising, gross margin structurally moving to 8%+, and valuation somehow not reverting, it cannot be excluded mathematically at 100%. But probability is low, and it requires betting against valuation mean reversion, which history just proved can happen (-43%).
  • Strengths that should be credited: performance is genuinely accelerating: FY25 revenue +48% and net profit +52%, Q1'26 net profit +102%, three consecutive quarters above ¥10.0 billion single-quarter profit, number-one share, real full-stack self-developed barriers, orders visible to 2027, and Q1'26 cash flow sharply repaired. This is not an Avoid-grade bad business; its beta is among the strongest in AI compute.
  • But it must not be elevated to fit a growth narrative: low gross margin 6.94%, weak bargaining power with profit going to Nvidia, valuation at historical highs and at a premium to peers, governance discount from about 84.8% Hon Hai system ownership, parent-subsidiary same-business overlap, no actual controller, and fragile ownership with institutions exiting and retail taking over. This is an outsourcing leader with accelerating earnings, not a great growth stock that can rise fivefold over the next 10 years.
  • Today's embedded expectation: the "not expensive" appearance of forward PE ~26-30× and PEG<1 rests on optimistic extrapolation of FY26 net profit +45-70%. AI is fully priced, and price has reached the upper end of broker fair ranges, with fair price about ¥62 and fair buy price about ¥60. The real margin of safety is around ¥60 or below, where forward PE returns to about 20×, near the upper end of Taiwanese peers and broker center. At ¥78.55, or even the already lower ¥73, paying for "fivefold" offers poor odds and poor probability. The rational posture is Watch, waiting for valuation to revert toward outsourcing peers or for earnings to further digest the valuation, not buying it as the next tenbagger/fivebagger.

  Jun 5, 2026
• QWhy has the market not realized all this yet? Is it too hard to understand, too disdained, or too long-duration? What would become the narrative inflection point?3/10

  A

  Conclusion upfront: FII is almost the negative case study for this Baillie Gifford question. It is not that the market has failed to recognize its greatness; it is that the market has already recognized it fully, even excessively. The stock rose more than 4 times in a year, static PE reached a post-listing historical high of ~44× (past 5-year percentile about 85%+, average about 18×), about 25 sell-side firms are bullish, with average target price about ¥92.47 and Citi high target ¥101. A-shares have already assigned a full AI compute growth premium. So in the Baillie Gifford framework, "too hard to understand" from insufficient coverage basically does not apply, and "too disdained" is reversed. The real disagreement is a valuation-method debate plus AI capex cycle-pricing debate, with one dangerous signal: institutions are systematically exiting while retail investors take the high-price handoff. This is exactly the crowded great narrative Baillie Gifford would beware, not the ignored greatness it seeks.

  1. Why "too hard to understand" does not apply: this is one of the most transparent and crowded A-share research targets. FII is an A-share AI compute heavyweight star. In 2025-10, market cap once exceeded ¥1.6 trillion and surpassed Kweichow Moutai to rank fifth in A-shares. Sell-side coverage is abundant; long and short logic is already known and written into every note. Bulls argue: number-one NVL72 rack outsourcing share of about 45-48%, full-stack self-developed barrier, performance doubling, and order visibility to 2027. Bears argue: low-margin outsourcing, weak bargaining power, most profit going to Nvidia, valuation at historical highs. When a story has been priced repeatedly by so many eyes, and even the dispute that it is "outsourcing in the name of technology" has become standard financial-media narrative, "the market has not realized it" has no foundation. The premise of Baillie Gifford Q10 fails directly for FII.

  2. The real disagreement is valuation methodology: value it as outsourcing, or as AI infrastructure? This is FII's core contradiction and the essence of a market that has realized the story but cannot settle the debate.

  • Bear case, reverting to outsourcing multiples: from electronic outsourcing history, FII's reasonable PE range is only 10-12×, corresponding to fair market cap of ¥500.0 billion-600.0 billion, implying 60%+ downside from current price. The subtext is that more than 60% of today's ¥1.6 trillion market cap is thematic premium. The basis is hard: 2025 consolidated gross margin is only about 7%, net margin attributable to parent below 4%, earnings quality below the 10%-15% industry average for consumer-electronics outsourcing, while valuation is nearly 4 times the industry average of about 10×.
  • Bull case, digesting valuation through AI high growth: the market is re-rating FII from an electronic-manufacturing framework of PE 15-20× toward an AI-infrastructure framework of PE 25-35×. On FY2026E net profit, forward PE 20-25× and PEG<1 do not look absurd. One broker, using FY2026 net profit of about ¥51.5 billion, calculates fair market cap of ¥1.24-1.55 trillion, about ¥62-78, meaning the current price has touched or slightly exceeded that camp's upper range.

  One stock has two internally coherent fair-value logics, "below ¥60" and "above ¥100", more than a twofold spread. This is not a blind spot; it is a paradigm disagreement over valuation anchor, outsourcing versus AI compute. The market sees it clearly. The disagreement cannot be solved by researching it again; fundamentals must disprove one side.

  3. Dangerous ownership evidence: institutions are selling and retail is buying, a sign of over-recognition. This is the opposite of Baillie Gifford's "ignored" setup and a sign of expectation exhaustion. Public fund holdings fell from ¥42.577 billion in 2025-09 to ¥23.363 billion in 2025-12 and ¥12.347 billion in 2026Q1, down more than 70% from the peak. Meanwhile shareholder accounts surged from about 476,000 to about 720,000, adding about 250,000 retail investors in half a year. Professional money systematically reduced exposure at historical highs, while retail investors used real money to absorb institutional selling. This is a typical structure of smart money taking profits and sentiment money crowding in; the trillion-yuan valuation lacks long-term institutional support and depends heavily on emotional retail investors, making the float fragile. In Baillie Gifford language: this is not a great growth stock the market has not yet discovered, but a crowded name that has been fully priced and whose consensus is starting to loosen.

  4. What will become the real narrative inflection point, deciding which valuation regime wins? Since the disagreement is about valuation paradigm, the inflection point is not whether the market discovers it, but which side's logic is verified or disproved by fundamentals. Candidate signals, also the main lines to monitor:

  1. North American cloud capex inflection, the most important downside risk: FII's fate is highly tied to North American cloud AI capex (TrendForce expects 2026 capex for the global top 8 CSPs to exceed USD 600.0 billion, about +40% YoY). If cloud capital budgets marginally contract or growth slows, the base of AI-infrastructure valuation shakes, and valuation will rapidly revert toward outsourcing multiples. This is the most likely trigger for the bear case of reverting to ¥500.0 billion-600.0 billion.
  2. Whether gross-margin mix-up is delivered, the upside catalyst and strongest bull argument: this is the swing factor. Goldman has explicitly judged that FII profitability is entering a new cycle of structural improvement and gives a ¥92.9 target. Q1'26 consolidated gross margin has risen to about 7.35%; the company's 2026 gross-margin target is 7.2-7.5%, high-power server revenue share 40%+, and liquid cooling (self-developed liquid-cooling gross margin about 15%, far above full racks) plus 800G+ networking mix are rising. If gross margin keeps climbing toward 8%+, the bear foundation of "outsourcing equals low quality" weakens, and the valuation debate tilts toward bulls.
  3. Whether valuation reverts to outsourcing multiples, the derating inflection: whether static/forward PE moves back toward Taiwanese peers (Quanta ~20.3×, Wistron ~17.7×, Wiwynn ~18.7×, Inventec ~29.6×) and U.S. peers (SMCI/Dell ~19×) is the direct readout of the bear case.
  4. Demand gap during GPU generation transitions, a cycle risk: if GB200 -> GB300 -> Rubin transition creates demand volatility, ODM earnings elasticity amplifies both ways and provides downside evidence for the valuation debate.
  5. Hon Hai related-party transactions / same-business competition controversy, the governance inflection: the Hon Hai system holds about 84.8%, and parent and subsidiary both do AI servers, with same-business overlap. Approved 2025 purchases from related parties were about ¥17.670 billion. If a related-party transaction dispute causes the market to reprice governance discount, it will compress the AI premium.

  Back to the honest landing point for this Baillie Gifford question: Q10 asks why the market has not yet realized its greatness. FII's reality is that the market has realized it and may have paid too much. It does not fit the LTGG picture of an ignored great growth stock. It looks more like a high-position AI cycle stock that is fully priced, consensus-crowded, and seeing institutions leave. The point is not to wait for market discovery, but to wait for fundamentals to decide the valuation debate. That balance depends on the inflection signals above. For a Baillie Gifford framework seeking a stock that can rise fivefold over 10 years and is not yet understood by the market, FII is almost the negative case: narrative is not lacking, pricing is not lacking; what is lacking is margin of safety and unexhausted expectations.

  Jun 5, 2026