Fabrinet

Phase A — Understand the business

Lens 1 · Company Overview

What it actually does. Fabrinet provides "advanced optical packaging and precision optical, electro-mechanical and electronic manufacturing services to OEMs of complex products" — optical communication components/modules/sub-systems, industrial lasers, automotive components, medical devices and sensors . In plain terms: it is a **specialist contract manufacturer (an EMS/ODM hybrid) for things that are hard to build** — high-mix, any-volume, high-complexity optical and electro-mechanical assemblies. It explicitly is *not* a product designer; in most cases it is "the sole outsourced manufacturing partner used by our customers for the products that we manufacture for them" .

Scale (FY2025, FYE 27-Jun-2025). Revenue $3,419.3M, up +18.6% from $2,883.0M in FY2024; net income $332.5M . The most recent quarter (Q3 FY26, ended 27-Mar-2026) ran at a **$1,214.3M** revenue rate, +39.3% YoY — i.e. the business is now annualising well above $4.5B .

Where the money comes from (FY2025 by end market) ``:

Customers (the defining feature). Two customers each contributed ≥10% of FY2025 revenue, together 45.8% (down from 48.5% in FY24) . Named 10%+ customers :

• NVIDIA: 27.6% of FY25 revenue (down from 35.1% in FY24, up from 12.5% in FY23)
• Cisco Systems: 18.2% (up from 13.4%)
• (FY23 comparatives: Lumentum 15.4%, Infinera 12.4% — both now sub-10%)

Critically, the latest 10-Q shows concentration rising again: in the nine months ended 27-Mar-2026, four customers were ≥10% and together were 58.1% of revenue (vs. two = 47.5% a year earlier) ``. More names crossed the 10% line, but the top-of-book grew heavier, not lighter.

Contract structure / payment terms. Master supply agreements with OEMs; high-mix manufacturing-services revenue (no take-or-pay, no SaaS recurring) — Fabrinet earns a thin manufacturing margin on customer-specified product. Customer-concentration also concentrates receivables: NVIDIA was 25.5% and Cisco 13.7% of trade AR at FY25 year-end; "Nokia (incl. Infinera)" was 12.0% . There is also a small **customer warrant** arrangement — $4.1M was recognised as a *reduction to revenue* in FY25 (a customer was granted warrants, effectively a price concession booked against revenue) ; this is almost certainly tied to the NVIDIA relationship and is worth watching as a margin/contra-revenue item.

Plain read: This is a high-quality, mission-critical sub-contractor to the optical-networking and AI-datacenter supply chain. It wins because it can build the hardest optical assemblies at volume and yield that its OEM customers can't or won't do in-house — but it sells labor + process + capex, not IP, so it earns ~12% gross margins, not 60%.

Lens 2 · Supply Chain

Fabrinet sits in the middle of the optical supply chain — it is the contract assembler between component makers and the OEM/hyperscaler buyers. The named map:

Upstream (inputs into Fabrinet):

• Lasers / EML / DFB chips — sourced from III-V semiconductor suppliers; the Q3 FY26 call flagged laser supply shortages as a binding constraint on datacom output ``.
• DSP / ASIC / retimers — Broadcom, Marvell (the datacom transceiver brains); also flagged as shortage items in Q3 ``.
• High-bandwidth memory / specialty memory — flagged as a shortage item ``.
• Optical sub-components — lenses, isolators, PLCs, MUX/DEMUX, photodiodes, packaging substrates.
• Strategic upstream move (subsequent event): On 25-Mar-2026 Fabrinet agreed to acquire a 16.0% equity interest in Raytek Semiconductor for NT$1.02B ($32.4M) — a vertical reach toward semiconductor/photonics supply, accounted under ASC 321 (no significant influence) ``.

Fabrinet (the node): Process design → supply-chain management → manufacturing → complex PCB assembly → advanced optical packaging → integration → final assembly & test. Facilities are concentrated in Thailand (Chonburi / Pinehurst campus), with smaller operations in the U.S., China and Israel. Long-lived assets are $475.0M in Thailand vs $35.9M U.S. as of Q3 FY26 (up from $338.1M Thailand at FY25 year-end) — i.e. the manufacturing base is overwhelmingly Thai, and the capex build is landing there ``.

Downstream (Fabrinet → end customer):

• NVIDIA (27.6% FY25) — optical interconnect for AI datacenters (DCI, and increasingly direct 800G transceivers / 1.6T modules).
• Cisco (18.2%) — telecom/datacom systems (Cisco acquired Acacia, a long-standing Fabrinet relationship).
• Nokia / Infinera (~12% of AR) — telecom optical (Nokia closed its Infinera acquisition; the two relationships consolidated).
• Coherent, Lumentum historically (now sub-10% but still in the mix as both customers and competitors-by-association).
• A new hyperscale customer — Fabrinet began shipping two 800G scale-out transceiver programs directly to a hyperscaler in Q3 FY26 (datacom diversification away from selling through a single module OEM) ``.

Geography of demand (bill-to, FY25) ``: Asia-Pacific 48.4% (Israel $993M, India $324M), North America 43.4% (U.S. $1,475M), Europe 8.2%. Israel as the #1 single country is the Nokia/optical-component cluster; the U.S. line is NVIDIA/Cisco bill-to.

Chokepoints / single-source dependencies:

1. Laser/DSP/ASIC component supply is the live bottleneck — Q3 FY26 datacom revenue was demand-constrained by inputs, not by orders ("we could have shipped a lot more if we had those components") ``.
2. Customer concentration is itself a supply-chain chokepoint — NVIDIA + Cisco are ~46% of the book; a single program loss is material.
3. Single-country manufacturing (Thailand) — geographic, geopolitical and FX concentration in one country (mitigated by Thailand's tax-privilege status, but a single-point-of-failure for the physical base).

Verdict on this lens: Names present — this is not a generic map. The two binding constraints are upstream component supply (near-term, cyclical) and customer concentration (structural).

Lens 3 · Competitive Advantages (moats)

The moat is real but narrow — it is a process/yield/trust moat, not an IP or pricing moat.

1. Switching costs / qualification lock-in (the strongest moat). Optical transceivers and modules require painstaking qualification; once Fabrinet is qualified as the sole outsourced manufacturer for a given product, moving it is slow, risky and expensive for the OEM. Fabrinet is repeatedly "the sole outsourced manufacturing partner" for the products it builds ``. This is why margins are stable-to-rising even with two giant customers — the customers need Fabrinet's yield and capacity more than Fabrinet needs any single program in the short run.
2. Scale + specialised capacity. Fabrinet can produce "high complexity products in any mix and any volume" — the high-mix/high-complexity end is exactly where in-house OEM manufacturing breaks down. Building 10 (below) extends this to ~$3B incremental revenue capacity ``.
3. IP protection / trust. Grady has explicitly built the firm's reputation around IP protection — critical when you handle competitors' designs (NVIDIA and a hyperscaler in the same building) ``.
4. Capital intensity as a barrier. $3B of revenue capacity requires real capex and operational know-how; a new entrant can't conjure qualified optical-packaging capacity quickly.

Bargaining power (who needs whom): Mixed. In the short/medium run Fabrinet has power (qualified, sole-source, capacity-constrained world). In the long run the customer holds power — these are EMS economics; the OEM can dual-source, in-source, or push price, and the $4.1M customer-warrant contra-revenue is direct evidence that a giant customer (almost certainly NVIDIA) can extract concessions ``. Gross margin of ~12% is the honest tell: this is a strong contract manufacturer, not a moated franchise.

What the moat does NOT protect against: a technology substitution (co-packaged optics) that removes the pluggable transceiver Fabrinet assembles. Process moats are worthless if the process becomes obsolete (see Lens 12/13).

Lens 4 · Segments

Fabrinet reports a single operating segment but breaks revenue by end market/product and by geography. All figures ``.

By end market — the trend and the cause:

FY2025 full year ``:

• Telecom $1,463M (+28.5%) — the FY25 growth engine; DCI and 400G/800G telecom ramp.
• Datacom $1,156M (+0.5%) — flat, surprisingly, given AI; a generational transition year (400G→800G) plus component supply.
• Automotive $464M (+41.9%) — strong non-optical leg.

Q3 FY26 — the reporting line-up changed, exposing the AI mix ``:

What this shows:

• DCI (+90%) and the brand-new HPC line ($106.7M from zero) are the AI fingerprint — these are the GB200/NVL-class interconnect and compute-system assemblies. They are why total revenue accelerated from +18.6% (FY25) to +39.3% (Q3 FY26).
• Narrowly-defined Datacom grew only ~4% YoY and fell ~6% sequentially — because of component shortages, not weak demand ``. This is the single most important nuance in the whole file: the AI demand is being rationed by upstream supply, so reported datacom understates true demand.
• Telecom (+42%) is no longer the sleepy leg — it is the largest single line and is itself riding AI-adjacent DCI.
• Automotive decelerated (−11% in Q3) — the non-optical diversification leg is cyclical and currently soft.

Geography (FY25): North America $1,481M (43.4%), Asia-Pacific $1,657M (48.4%, led by Israel $993M and India $324M), Europe $282M (8.2%) ``.

---

Phase B — Measure performance

Lens 5 · Earnings Result (latest print: Q3 FY2026, reported 4-May-2026)

Headline (all `` unless marked):

• Revenue $1,214.3M, +39.3% YoY (vs $871.8M); record quarter, above the guided range ``.
• Gross profit $144.3M, gross margin 11.9% (vs 11.7% PY — roughly flat, the EMS reality).
• Operating income $120.0M, op margin 9.9% (vs 9.0% PY — modest operating leverage).
• Net income $125.2M, net margin 10.3% (helped by a $7.0M FX gain and low ~6.7% tax rate).
• Diluted EPS $3.45 (vs $2.25 PY, +53%); basic $3.49 ``.
• 9-month FY26: revenue $3,325.3M (already ≈ all of FY25's $3,419M), diluted EPS $9.22.

What drove it ``:

• Telecom +55% YoY (per company framing), powered by DCI +90% to $197M.
• Non-optical +52%, led by the new HPC line.
• Datacom held back by supply — lasers, memory and ASIC shortages broadened; management said revenue was "well below demand levels."

Guidance / outlook: Q4 FY26 revenue guided $1.25B–$1.29B (midpoint $1.27B, ≈ +31% YoY) ``. Management flagged new datacom customer agreements (two 800G programs shipping direct to a hyperscaler) as a forward growth lever. Tone: confident on demand, candid that supply is the governor near-term.

Balance-sheet / cash-flow flags:

• FY25 OCF $328.4M, capex $121.8M → FCF ~$206.5M (FCF margin 6.0%) ``.
• But FCF turned slightly negative in Q2 FY26 (≈ −$5.4M) as working capital (AR +inventory) and the Building-10 capex outran operating cash . FY25 already showed the working-capital strain: AR built **+$165.7M** and inventory **+$117.8M** (partly offset by AP +$194.2M) . This is the classic hyper-growth EMS cash signature — growth consumes cash. Watch FCF, not just EPS.
• Net cash fortress intact: $306.4M cash + $627.8M short-term investments = $934.2M, zero debt at FY25 year-end ``.

Market reaction / what was priced in: The stock is up ~+160% over the trailing year into the print , trading ~$574 as of 21-Jun-2026 . A beat-and-raise on a name already up 160% means expectations are extremely elevated — the bar is now "keep beating, and prove the absorption of Building 10."

Anything unusual vs. its own history: (1) the appearance of the HPC line from zero — a genuine new revenue vector; (2) datacom demand being supply-rationed — unusual to have the AI leg held back by inputs rather than orders; (3) rising customer concentration (4 names = 58% of 9M revenue) reversing the prior de-concentration.

Lens 6 · Earnings Calls (sentiment trend)

No transcripts on disk (transcripts/ empty) → web-grounded ``.

Q3 FY26 (May 2026) — confident, supply-frustrated. Management's posture: demand is excellent and broadening (DCI +90%, new HPC, two hyperscaler 800G wins), but supply shortages (lasers/memory/ASICs) are the binding constraint — "we could have shipped a lot more if we had those components." Heavy emphasis on Building 10 (2M sq ft, ~$3B revenue capacity, Phase 1 opening ~July 2026) as the answer to future demand ``.

OFC 2026 (March 2026) conference — "sustained growth amid challenges." Tone framed around 1.6T transceiver ramp and capacity pull-forward ``.

Shift over time (the phrase trend):

• Stopped saying: worry about the 400G→800G datacom transition air-pocket that dogged FY24/early-FY25 (datacom was flat).
• Started saying: "supply-constrained," "1.6T," "Building 10," "scale-out," "direct to hyperscaler," "HPC." The vocabulary has migrated from telecom-cycle language to AI-capacity language.
• Consistent through-line: IP protection, operational excellence, "any mix any volume," customer diversification.

Sentiment read: Management tone is bullish and increasingly AI-centric, but notably honest about constraints (they flag supply shortages and absorption risk rather than hand-waving). That candor is a mild positive on trustworthiness (see Lens 9).

Lens 7 · Comps

Peer set: EMS/contract manufacturers (Celestica, Jabil — the right economic comp) and optical-component vendors (Lumentum, Coherent — adjacent but higher-margin product companies). Fabrinet is economically an EMS but is valued like an optical-AI play — which is the entire valuation debate.

Read: FN at ~45× FY26E trades above Celestica (~38×), the one peer with comparable economics — and Celestica itself is already richly valued by historical EMS standards (EMS used to be a 10-20× business). The market is explicitly pricing FN as an AI-optics growth franchise, not a contract manufacturer. The EV/Sales of ~4.3× `` on a 12%-gross-margin business is the single most aggressive number in the file — for context, a 4.3× EV/Sales implies ~36× EV/gross-profit. This is the bear case in one ratio.

Lens 8 · Stock-Price Catalysts (what moves FN >5%)

Mostly ``; pattern over recent years:

• Quarterly earnings prints are the dominant catalyst — beat-and-raise prints (like Q3 FY26) and guidance are what move the stock most ``.
• NVIDIA-linked datapoints — because NVIDIA is ~28% of revenue, FN trades as a high-beta NVIDIA/AI-capex derivative; NVIDIA product-cycle news (GB200/GB300, 800G→1.6T) and hyperscaler capex commentary move it.
• CPO / co-packaged-optics headlines — any signal that NVIDIA/Broadcom CPO accelerates pressures FN (removes pluggables); any signal that pluggables persist supports it ``.
• Datacom transition air-pockets — the 400G→800G handoff produced a flat-datacom stretch that pressured the stock in FY24/early FY25.
• The +160% trailing-year run itself `` means the stock now reacts violently to any crack in the AI-capacity narrative.

What the pattern reveals: the market reacts to (1) the print/guide, (2) the NVIDIA/hyperscaler demand signal, and (3) the CPO existential question. It is not a dividend/defensive name; it is an AI-capex momentum vehicle.

---

Phase C — Judge people & books

Lens 9 · Management

CEO: Seamus Grady (CEO since 2017; also Chairman since 2025) ``.

1. Track record: Ex-EVP & COO of Sanmina (13 years; ran worldwide ops, 10 facilities / 5 countries); earlier supply-chain leadership at Lucent (Ireland) and Manufacturers' Services Ltd (now Celestica). A career operations/supply-chain executive — precisely the right archetype for a high-mix contract manufacturer . Under his tenure revenue roughly tripled to $3.4B (FY25) with steady margins .
2. Tenure & skin in the game: ~9 years as CEO. Insider-transactions data not on disk (insider-transactions.csv absent) → insider ownership n/a; do not fabricate. The CEO+Chairman consolidation (2025) is worth a governance note (see red flags).
3. Capital-allocation history: Disciplined and shareholder-friendly. FY25: buybacks $125.7M (up from $39.5M), zero debt, $934M net cash . Reinvestment is now stepping up hard — capex $121.8M in FY25 (from $47.5M) for Building 10. **No dividend.** ROE is high: FY25 net income $332.5M on ~$1.98B equity ≈ **~17-18% ROE** ; ROIC is effectively higher given net-cash and minimal debt.
4. Red flags (management-level): (a) CEO/Chairman role consolidation weakens board independence; (b) the $4.1M customer warrant booked as contra-revenue is a related-incentive structure with a key customer worth monitoring; (c) the Cayman domicile + Thailand-tax-privilege structure is legitimate but adds opacity. None are disqualifying.
5. Founder vs. professional manager: Professional operator (not founder). For a contract manufacturer at this stage, that is the right archetype — the job is execution, yield, capex discipline and customer trust, all of which Grady has delivered.

Assessment: A credible, disciplined operator with a clean balance-sheet record and the exact operational pedigree this business needs. Management quality is a positive in the thesis. The main watch-items are governance optics (CEO=Chairman) and the capex-absorption risk now in his hands.

Lens 10 · Forensic Red Flags

Accounting-risk scan (income statement, balance sheet, cash flow):

• Revenue recognition: Manufacturing-services revenue recognised over time/at shipment per ASC 606; contract assets $14.5M . The **customer warrant ($4.1M contra-revenue in FY25)** is the one non-standard item — it *reduces* reported revenue and is a soft margin/incentive concession to a major customer; track whether it grows .
• Cash flow vs. earnings divergence (the real flag): FY25 net income $332.5M but OCF $328.4M — OCF slightly below net income, because AR (+$165.7M) and inventory (+$117.8M) outran revenue . Q2 FY26 FCF went **slightly negative (≈ −$5.4M)** . This is growth-driven working-capital absorption, not manipulation — but it means EPS is currently higher-quality on paper than in cash. In a demand-down quarter this reverses favourably; in continued hyper-growth it keeps consuming cash.
• Receivables concentration: trade AR $758.9M (FY25), with NVIDIA 25.5% + Cisco 13.7% + Nokia 12.0% of AR `` — payment-default exposure is concentrated in three strong-but-cyclical names.
• Inventory: $581.0M (up from $463.2M) — rising with the ramp; watch for a write-down risk if a datacom generation transitions abruptly (the 400G→800G handoff history shows this is a real risk).
• SBC: $33.0M in FY25 (≈1.0% of revenue) — modest and not flattering non-GAAP egregiously (unlike many tech names). A point in Fabrinet's favour.
• Goodwill/intangibles: negligible (intangibles $2.2M) — no goodwill-impairment landmine. Clean.
• Leases / related parties / contingencies: small operating leases ($5.8M ROU); no material related-party flags surfaced.

Regulatory findings (required sub-section):

• SEC Litigation Releases: none found naming Fabrinet (EDGAR EFTS LR search, 2021-06-21→2026-06-21) ``.
• SEC AAERs: none found (same search window) ``.
• 10-K Item 3 (Legal Proceedings): the FY25 10-K discloses customary litigation; no material proceeding flagged as reasonably likely to have a material adverse effect ``.
• Non-SEC (FTC/DOJ/FDA/CFPB) web check: no material enforcement actions, consent decrees, fines or penalties surfaced for "Fabrinet" ``.
• Conclusion: No material regulatory or legal findings — verified via SEC EDGAR EFTS (LR, AAER), web search, and 10-K Item 3 as of 2026-06-21.

Forensic verdict: The books are clean (no goodwill, low SBC, no enforcement history). The one genuine analytical flag is cash-vs-earnings quality during the ramp — FCF is thin and went briefly negative because hyper-growth + Building-10 capex is consuming working capital. That is a quality-of-growth issue, not an integrity issue.

---

Phase D — Project & stress-test

Lens 11 · Forward Projection (EPS, next three fiscal years)

Fabrinet's fiscal year ends late June. Building bottom-up from 9M-FY26 actuals + Q4 guidance; all inputs labeled, outputs ``.

FY2026 (ending ~June 2026) — nearly locked:

• Revenue = 9M actual $3,325.3M + Q4 guide midpoint $1,270M = ~$4,595M (+34.4% vs FY25) ``.
• EPS = 9M diluted $9.22 + Q4 $3.55-3.75 → **$12.8-13.0** ``.

FY2027 — base/bull/bear:

• Drivers: 1.6T transceiver ramp + Building-10 Phase-1/2 capacity unlock (~$3B incremental capacity over time; Phase 1 ~July 2026 adds ~$350-400M) ``; continued DCI + HPC growth; partly offset by component-supply rationing and automotive softness. Margins ~flat (EMS ceiling ~12% GM, ~10% OM).
• Base: revenue ~$5.4B (+18%), EPS ~$15.0 ``.
• Bull: revenue ~$5.9B (+28%) as supply unclogs and 1.6T inflects, EPS ~$16.8 ``.
• Bear: revenue ~$4.9B (+7%) if CPO begins biting datacom and/or a datacom-generation air-pocket hits, EPS ~$13.0 ``.

FY2028 — base: revenue ~$6.2B, EPS ~$17.5 `` — with materially widening error bars because the CPO-vs-pluggables question resolves over this window.

Valuation cross-check at ~$574: ~45× FY26E and ~38× FY27E base EPS ``. For an 18-20%-EPS-grower with 12% gross margins and a live technology-substitution risk, that is a full-to-rich multiple — the price already capitalises the Building-10 ramp succeeding.

(Per --watchlist rules, no forecast.ts create logged in this loop. The natural Brier line to log later, conversationally: "FN FY27 non-GAAP EPS ≥ $15.00", p≈0.55, resolves ~2027-06-30.)

Lens 12 · Bull vs Bear

Bull case. Fabrinet is the outsourced manufacturing toll-collector on the AI optical-interconnect build-out — the unglamorous picks-and-shovels name. As compute scales out (GB200→GB300→Rubin), the number of high-speed optical connections per cluster explodes, and the world's hardest-to-build 800G/1.6T transceivers, DCI modules and HPC assemblies need a qualified, high-yield, IP-trusted manufacturer. Fabrinet is that manufacturer, sole-sourced on most programs, with $3B of new capacity (Building 10) coming online into a supply-constrained market it literally cannot keep up with today ("could have shipped a lot more"). The moat is qualification lock-in; the balance sheet is fortress (net cash $934M, zero debt); management is a disciplined operator buying back stock. Growth is accelerating (+18.6% FY25 → +39% Q3 FY26), datacom is being artificially suppressed by component supply (so reported numbers understate true demand), and the new direct-to-hyperscaler 800G wins prove it can diversify beyond selling through a single module OEM. Earnings surprises skew up as supply unclogs. Secular tailwind + capacity + sole-source = a multi-year compounder.

Bear case (2-3 things that could permanently impair).

1. Co-packaged optics (CPO) is the structural axe. If/when NVIDIA + Broadcom move interconnect from pluggable transceivers (which Fabrinet assembles) to co-packaged silicon photonics (integrated at the switch/GPU, which Fabrinet largely does not), Fabrinet's core datacom/DCI TAM shrinks. NVIDIA's Spectrum-X CPO switch is reportedly entering production late 2026 (3.2T silicon photonics); CPO penetration is forecast to go from <1% to >35% of optical networking by ~2030 ``. This is the one risk that doesn't just dent earnings — it can re-rate the entire thesis.
2. Valuation has front-run execution. ~45× FY26E and ~4.3× EV/Sales on 12% gross margins for a contract manufacturer is priced for flawless multi-year execution. A single datacom air-pocket, a Building-10 absorption miss, or a CPO scare compresses the multiple hard — a de-rate to a still-generous 25× FY27E base EPS ($15) implies ~$375, ~35% downside from $574 ``.
3. Customer concentration + EMS economics. Four customers = 58% of 9M-FY26 revenue; NVIDIA alone ~28%. The $4.1M customer warrant contra-revenue proves the big customer can extract price. In a downturn, concentrated AR and the working-capital cash drain (FCF already went briefly negative) bite simultaneously.

Pre-mortem (18 months out, thesis broke — what happened?): CPO ramped faster than bulls assumed on NVIDIA's next-gen switches, datacom/DCI orders softened just as Building 10 capacity came online, fixed-cost absorption inverted (capacity glut → margin compression), a datacom-generation transition produced an inventory write-down, and the ~45× multiple collapsed toward EMS norms (~15-20×). Revenue still grew, but the rate decelerated and the multiple did the damage.

Are multiples too high? For the business it is (12%-GM EMS), yes, clearly. For the growth it's printing right now (+39%), it's defensible only if the AI optical-interconnect ramp runs multiple more years AND pluggables out-survive CPO at the high end. The multiple is a bet on duration and on CPO being slow.

Contrarian view (what the market refuses to see): Bulls treat FN as an "AI optics franchise" and pay a franchise multiple; the contrarian read is that it is a brilliant contract manufacturer at a cyclical/technological peak, being priced as if EMS economics and CPO substitution don't exist. Conversely, the over-bearish contrarian read is that datacom is currently demand-constrained by supply, not by lack of orders — so the next two quarters could keep beating even as bears scream "peak," because Building 10 unlocks shipments that demand already wants. Both can be true: keep beating near-term, re-rate lower medium-term.

Lens 13 · Devil's Advocate (short-seller)

I am short FN at ~$574. Here is how this breaks:

• What structurally breaks the business model: CPO removes the product. Fabrinet's value-add is assembling pluggable optical transceivers and modules. The entire hyperscale/NVIDIA roadmap is pushing toward co-packaged / integrated silicon photonics, where the optics move onto the switch/GPU package and the discrete pluggable transceiver — Fabrinet's bread and butter — gets disintermediated. NVIDIA's own CPO switches are entering production. Fabrinet is "part of NVIDIA's silicon-photonics ecosystem," but as an assembler, not the IP owner — and CPO concentrates value in the silicon-photonics/packaging IP (Broadcom, TSMC, NVIDIA), not the transceiver assembler.
• Revenue concentration / what happens if it shifts: NVIDIA ~28%, top-4 = 58% of revenue. If NVIDIA in-sources, dual-sources, or shifts its interconnect architecture to CPO, a quarter of the business is at risk and Fabrinet has zero pricing power to defend it (it's already conceding warrants). Cisco (18%) is a slow-growth telecom anchor, not a hyper-growth offset.
• Why the moat is weaker than bulls think: A process/qualification moat protects against competitors copying you; it does not protect against your customer changing the technology so the process is obsolete. EMS moats are real until the product category moves — ask anyone who made the last generation of anything. 12% gross margins are the market telling you the moat is thin.
• Most dangerous competitor bulls underestimate: Not another EMS — it's vertical integration + CPO (Broadcom/TSMC/NVIDIA packaging) and the silicon-photonics startups, plus Celestica moving up the AI-hardware value chain. The threat is architectural, not a head-to-head bid.
• Worst capital-allocation / incentive concerns: Plowing $130M+ into Building 10 (2M sq ft) right as CPO threatens the demand it's built for is the classic late-cycle capacity bet — if datacom architecture shifts, that's stranded fixed cost and an absorption nightmare. The CEO=Chairman consolidation and the customer-warrant contra-revenue are governance/incentive yellow flags.
• Assumptions that must hold for $574: (1) pluggable transceivers stay dominant at 800G/1.6T for years (CPO is slow); (2) Building 10 fills and absorbs; (3) NVIDIA/Cisco concentration doesn't break; (4) component supply unclogs without the demand evaporating by the time capacity lands; (5) the market keeps paying ~45× for a 12%-GM manufacturer.
• Valuation if growth disappoints 20-30%: FY27 EPS to ~$11-12 instead of ~$15, multiple compresses to ~20-25× → ~$240-300, ~50-58% downside ``.
• Single scenario that permanently impairs: CPO crosses the chasm on NVIDIA's 2027/2028 switching generation faster than consensus, structurally shrinking the pluggable-transceiver TAM. Plausibility: moderate and rising — production CPO switches in late 2026 make this a live 2027-2028 risk, not a 2030 abstraction.

Lens 14 · Management Questions (ordered by information value)

1. Co-packaged optics is the existential question for the pluggable-transceiver business. What is your concrete role and revenue opportunity in a CPO world — are you positioned to manufacture CPO/silicon-photonics assemblies, or does CPO disintermediate you? Quantify the TAM you keep vs. lose.
2. Building 10 adds ~$3B of revenue capacity. What is your line of sight to filling it, what is the expected utilization curve through CY2027, and what happens to gross margin if datacom demand softens while that fixed cost is online?
3. Datacom shipments were "well below demand" on component shortages (lasers, memory, ASICs). When do you expect that to clear, and how much latent datacom revenue is being deferred quarter-to-quarter today?
4. Your top four customers are ~58% of revenue and NVIDIA alone ~28%. What is your realistic path to reducing single-customer dependence below 20%, and how durable are the two new direct-to-hyperscaler 800G programs?
5. The $4.1M customer warrant booked as contra-revenue — what triggers it, how large can it become, and does it signal pricing concessions to your largest customer that will pressure gross margin?
6. Gross margin has held ~12% through a 39% revenue ramp. What structurally caps it, and is there any mix shift (1.6T, HPC, direct-to-hyperscaler) that can move it materially higher — or is 12% the ceiling?
7. FCF turned briefly negative as working capital and capex outran operating cash. At what revenue growth rate does the business become reliably FCF-positive again, and what is your target FCF conversion at scale?
8. How do you think about your manufacturing footprint being ~overwhelmingly in Thailand? What is the geopolitical/FX/single-point-of-failure mitigation, and would you put 1.6T/CPO capacity outside Thailand?
9. The Raytek Semiconductor 16% stake — what is the strategic logic? Is this the start of upstream vertical integration into photonics/semiconductors, and should we expect more?
10. What is your capital-allocation priority order over the next three years — capex (Building 10), buybacks, M&A, or a dividend — and at what cash level do you start returning more?
11. HPC went from $0 to $107M in a quarter. What is this line, who is the customer, what is its durability, and is it as sticky as your optical programs?
12. Inventory is up ~25% YoY into a datacom generation transition. What is your write-down exposure if 800G→1.6T or a CPO shift happens faster than planned?
13. You consolidated CEO and Chairman roles in 2025. How is board independence preserved, and what is your succession plan given the operational specificity of this role?
14. Telecom (+42%) is now your largest line. How much of that is genuinely AI/DCI-driven vs. a normal telecom-capex up-cycle that could roll over?
15. Where do you most disagree with the bull case on your own stock — what do investors most overestimate about Fabrinet's durability?

---