The AI-memory super-cycle is durable enough to keep DS (memory) operating profit and group earnings revisions running ABOVE the market's mean-reversion prior through CY2026-2027, rather than reverting like prior memory cycles: sold-out multi-year HBM LTAs, a ~3-year fab-plus-qualification lead, and demand doubling YoY against only ~14% supply growth make a 2026/27 air-pocket genuinely hard, so DRAM contract-price growth stays positive and DS OP keeps beating (1Q26 group OP KRW 57.2tn beat the Street-high by ~33%). The sell-side already prices the cycle (KRW PTs quadrupled from ~W128k to ~W550k in seven months), so the differentiated read is that the PRINT and the revision slope keep outrunning even the raised bar, not that the multiple re-rates.

Samsung narrows the HBM share gap to SK hynix - independent trackers now project its HBM bit/volume share rising toward ~40-41% by 2027 - but does NOT reach yield/qualification PARITY through mid-2027: its HBM4 die yield stays materially below SK hynix's (a ~20-25pt gap persists, not converging to within ~10pts), and its NVIDIA Vera Rubin HBM4 allocation remains a qualified second-source (~20-30% spillover band), not a primary-at-volume socket. So even as Samsung gains share, AI-memory profit-per-bit and the marquee primary sockets keep routing disproportionately to SK hynix, and the sell-side's pricing of full HBM parity remains too optimistic.

Samsung Foundry remains structurally sub-scale and loss-making on a STANDALONE basis through FY26-FY27: 2nm GAA yield stays ~20-30pts behind TSMC, the ecosystem-trust gap (no CUDA-equivalent, hardened IP ports to TSMC first, packaging maturity vs CoWoS) is not closed by yield improvement alone, and it fails to land a marquee EXTERNAL 2nm anchor customer at volume (Exynos self-supply and small ASICs do not count) - so its only credible value contribution is as an enabler of Samsung's HBM4/5 logic base-die integration, never as a standalone profit pool, and the "foundry turning profitable this year" bull case (a single ex-employee source) does not materialise.

The hyperscaler shift to custom ASICs (Google TPU, Tesla, Meta, Microsoft Maia, Amazon Trainium) is NET ADDITIVE to Samsung's HBM/memory demand through the forecast horizon, not erosive: custom silicon targets only the compute pillar while memory and networking are still required, and HBM keeps growing as a share of AI-infrastructure BoM - the read the two highest-signal customer voices (AWS provisioning PM, Microsoft cloud-AI engineer) both land on. So ASIC proliferation broadens, rather than concentrates, the buyer base for Samsung's merchant HBM.

RISK/BEAR CLAIM (tracked literally; a challenge is GOOD for a constructive stance): China's CXMT advances FASTER in commodity/mainstream DRAM than a near-term 'non-event' framing implied - it is now the #4 DRAM maker by volume (~16% of wafer starts), with bit share climbing toward ~10%+ by 2027, selling at ~half price and pressuring low-end contract pricing - BUT remains EUV/leading-edge-lagged and domestic-tilted, without DDR5/HBM-grade parity, through FY27. The structural cap on Samsung's through-cycle margin materialises only if CXMT secures leading-edge (DDR5/HBM-grade, EUV-class) capacity despite export controls; through the FY27 horizon that leading-edge leap is NOT yet evidenced - a ~2028+ commodity-margin risk, not a near-term leading-edge threat.

RISK/BEAR CLAIM (tracked literally; a challenge is GOOD for a constructive stance): Samsung's chaebol governance stays a live, under-priced valuation discount that the near-unanimous sell-side ignores - the Samsung Life Insurance 21.16% affiliate cross-holding, the Lee Jae-yong controller overhang (2017 bribery conviction; inheritance-tax-driven family share sales/pledges), related-party supply, and severe 2022-2025 C-suite churn (incl. co-CEO Han Jong-hee's death) are treated as RED by the System cg_report while the 140-note bull corpus is near-silent on controller risk. The risk persists unless a durable governance de-rating catalyst lands; the corpus's benign "value-up plan + buyback" framing under-weights it.

TCB-led advanced packaging is inflecting faster than the group: SEMI AP bookings doubled YoY in Q2 FY26 (SEMI bookings +125.9% YoY vs group +97.6%), the key memory player is qualifying ASMPT's fluxless AOR TCB for HBM4 16H alongside a flux-based production tool, and logic C2W tools booked at a leading advanced logic customer. This converts into above-consensus SEMI revenue and estimate revisions through FY27 as the HBM4 ramp and the US$1.6bn-by-2028 TCB TAM (30% CAGR, mgmt via Citi/GS) are absorbed into street models.

Photonics scales from a low base into a durable second advanced-packaging growth leg: revenue grew ~5x YoY in Q1 FY26, Photonics is already the No. 2 AP contributor behind TCB, 1.6T pluggable transceiver tools won bulk orders from key optics suppliers, and CPO demo tools (AMICRA Nano) have shipped. Photonics plus TCB - the two named Q3 FY26 bookings drivers - keep group bookings growing YoY through FY27 even as the 1Q25 low-base effect washes out.

Portfolio pivot plus mix drives a structural margin re-rate: gross margin stepped +542bps HoH to 41.1% in 1H FY26 (Q2 42.4%), the loss-making NEXX unit (FY25 net loss HK$182.8m) is sold to Applied Materials for US$120m with margin relief from 2H26, and management is exploring strategic options for SMT (sale, separate listing, or retention - JPM: "the next big strategic move"). Operating leverage (Q2 revenue +52.1% YoY vs operating profit +300.1%) holds through FY27 as AP mix rises.

Mainstream softness caps the group revision cycle despite the TCB story: AP is only ~30% of revenue (FY25), SMT (~46% of FY25 revenue) is cyclical and largely unrelated to the AI cycle, automotive outside China and industrial "have not yet shown significant recovery" (GS, May 2026; industrial fell 13% to 10% of revenue in FY25), and management volunteered longer lead times for certain materials while guiding Q3 - a constraint that binds SMT (GS: SMT raw-material bottleneck). Through FY27 mainstream drag keeps group revenue growth below what the SEMI bookings inflection implies.

Competitive erosion of the TCB position at memory customers: Hanmi remains dominant at Micron (ASMPT only a tester there, alongside Besi), SK hynix dual-sources (ASMPT ~half of its ~50 TC bonders, per The Elec via broker research), and hybrid bonding - where management concedes its Gen 2 tool is only "on par with the industry incumbent" - could displace TCB at HBM generations beyond 16H. Through FY27, competitor wins and/or an early hybrid bonding shift erode ASMPT's TCB share of the HBM tool base.

The specialty franchises - tantalum capacitors (#1 at ~46% share, ~24% of sales but >35% of operating profit at GM ~17-20pp above the company average) plus chip resistors (#1 at ~40% volume share, GM 42-46% vs MLCC 31.6-34.8%) and >50%-GM sensors - are Yageo's durable, share-led pricing/margin engine, NOT commodity MLCC. This is the leg where a peer-convergence multiple is legitimately earned and it is unopposed in the corpus; the group's profit is a barbell whose tip is KEMET tantalum first, then resistors/sensors, and it holds through cycle because Yageo has real concentrated share and pricing power there.

Supply discipline + JP/KR spillover, not structural pricing power, drives Yageo's record 2026-1H27 earnings: as the single largest commodity-MLCC capacity holder (~23%; #2 only ~15%) running FLAT capex while Murata adds ~20%/2y and Chinese low-end players suspend orders, Yageo captures the overflow from N-Asian peers occupied by high-end AI MLCC. With inventory lean (2025 ~36% below the decade peak, distributor 4.8mo below the healthy band), book-to-bill ~1.3x and - critically - NO double-ordering yet, the near-term upcycle is real and delivers record revenue and margin through 1H27.

The 46-51% terminal group gross margin and "structural commodity-MLCC pricing power" embedded in the ~9x re-rate are cyclical/peak assumptions priced as permanent, and will NOT prove durable. Yageo is a commodity-MLCC price-FOLLOWER that cannot lead a blanket hike without Murata (confirmed 2026 hikes are select consumer/auto SKUs + a first public list, not a broad DIRECT-customer commodity move that has held), its own-brand MLCC earns GM <20%, and the 46-51% blended terminal requires peak UTR (97-100%) to hold. Through-cycle group GM most likely lands high-30s/low-40s and the 2-3SD multiple compresses as book-to-bill normalizes off ~1.3x.

(RISK claim, tracked literally.) The Rubin tantalum design-out is the single most under-priced risk: Nvidia designs tantalum content OUT of the Rubin/VR200 generation (-60%, swapping to MLCC/aluminium to cut BOM), opening a 2027+ air-pocket in KEMET's tantalum datacenter book - the group's crown jewel (#1 ~46% share, ~24% of sales, >35% of operating profit, AI ~35% of the tantalum book). The corpus has barely priced this because it rests on a single source (TDK competitor, Mar-2026) and needs urgent second confirmation.

(RISK claim, tracked literally.) Yageo is a fully crowded long at a deteriorating risk/reward: all 6 covering houses rate Buy/OW with every one above Street EPS, the multiple sits at a 20-year peak (BofA's own math: ~3.0SD above the 9.5x since-2014 average; now 26-40x fwd / 26x PB-ROE), and the variant (AI + transformation) has BECOME the consensus. Against a 2-3SD multiple the payoff is asymmetric to the downside: the first book-to-bill or ASP roll-over triggers multiple compression far larger than any remaining fundamental upside.

Revenue growth through FY27 is a PRICE outcome, not a volume outcome, and the realised 2027 like-for-like wafer price round lands at or above 7% blended. The corpus shows three realised rounds (about 3% in 2Q25, 6-10% in 1Q26, a 2027 round in discussion at 5-10%), blended ASP up 16% YoY in 2Q25, and advanced-node pricing now 9x mature. With leading-edge capacity fixed through 2027 by decisions already taken in 2024-25, price is the only free variable left in the revenue line, so the claim holds if the disclosed or broker-verified 2027 blended like-for-like increase prints at or above 7% rather than the sub-5% implied by TSMC's stated partner-pricing posture.

Leading-edge (N3/N5) utilisation stays above 100% through FY27, which is the load-bearing assumption under both the pricing and margin legs. Komico puts Taiwan N3 at 110-115% (Oct 2025) and about 110% (Apr 2026), J.P. Morgan independently at 115-120%, and models a 2026 N3 demand-supply gap of roughly 600k wafers or US$15-18bn. Non-AI weakness does not break this while released capacity is refilled by AI, so the claim holds if leading-edge utilisation stays above 100% even as smartphone, PC, auto and consumer units decline in FY26.

Gross margin plateaus in the mid-60s rather than reverting, holding at or above 65% through FY27 despite the guided 3-4pp N2 dilution and a 28% half-on-half depreciation step in 2H26. The printed series is 62.3% (4Q25), 66.2% (1Q26) and 67.7% (2Q26) against a raised structural floor of 56%-plus, and every 2025 bear GM forecast (Bernstein 54.6%, UBS 55.8%, BofA 57.2%) was beaten by 700-1,000bps. The claim requires GM to stay at or above 65% as N2 rises from 3% of sales in 2Q26 to 15-20% in 4Q26, evidencing that pricing offsets the ramp and overseas drag.

RISK CLAIM (confirmations are BEARISH for the core thesis): the advanced-packaging bundle leaks, not as a share-loss event but as a slow substitution already under way. J.P. Morgan models about 70% of CoWoS-L on-substrate steps moving to ASE and SPIL by 2026 and reports TSMC actively steering customers to OSAT 2.5D from 2H26; Google chose Intel EMIB for TPU v8 inference; an Intel litho engineer projects EMIB density parity within two to three years at roughly 30% lower cost. The AWS product-management director named the converting trigger explicitly: a sustained 15-20% gap between demand and primary-supplier capacity, a gap J.P. Morgan independently measures and which has been open for four quarters. The risk materialises as 2027-28 packaging mix and per-unit capture erosion rather than in any single quarter.

RISK CLAIM (confirmations are BEARISH for the core thesis) - REINSTATES the concentration risk on corrected evidence: customer concentration has risen monotonically for four years and is intensifying in identity as well as level. Top-10 customers went from 68% of sales in 2022 to 78% in 2025; Nvidia OVERTOOK Apple as the largest customer at 19% of 2025 sales and is modelled at 21% in 2026, while holding roughly 60% of 2026 CoWoS demand. The prior ledger read the top-1 line falling from 25% to 19% as de-concentration; the debate corpus reads the same numbers as a counterparty ROTATION into a customer that is more concentrated than Apple ever was, because it couples wafer share to packaging share. The corpus consistently sizes individual defections at 1-2% of revenue and never sizes the correlated case.

MediaTek's cloud AI-ASIC franchise, anchored on Google TPU, proves durable and MediaTek's to keep rather than one-deal optionality: cloud-ASIC revenue inflects from the >US$1bn 2026 guide to the ~US$2bn raised at the May 2026 Analyst Day and toward ~20% of revenue in 2027, the v9 'Humufish' win ships as the primary 2nm family, and MediaTek retains a ~40-50% share of Google TPU (Bernstein) as volumes prove out - the ramp escalates rather than reverting to Broadcom, and the ASIC revenue revision cycle stays steeper than the smartphone-cycle base implies.

BEAR/RISK CLAIM (tracks the concentration bear literally; confirmations are the live thesis risk): the AI-ASIC franchise stays a single-customer (Google), execution-binary bet through 2027 - no second ASIC customer reaches scale (Meta MTIA stays 50:50/undecided, xAI and any non-Google socket remain pipeline-only), the v10 RFQ contest keeps the award at genuine risk of moving to Broadcom, and the lone customer voice's ranking of MediaTek as a distant #3 that 'enters on diversification, not competitiveness, and won't move up near future' [Microsoft principal 2025-09-09] is never overturned - so the entire re-rating stays hostage to one binary socket.

Blended gross margin stays structurally below the 48-50% long-term model through 2027 and the model is quietly abandoned rather than reset: GM prints ~46% for FY26 (vs 47.5% FY25) and glides toward ~44-45% because (a) MediaTek cannot fully pass through 2026 China memory/BOM inflation on smartphone ASP (mobile guided down double-digits YoY into 2026) and (b) the scaling ASIC mix (GM ~42-46%) is gross-dilutive - the smartphone-margin erosion the bears won on the fact is structural, not merely cyclical-and-priced.

ASIC scales as a low-value-capture business rather than a genuinely accretive mix: the first datacenter-segment gross-margin read lands materially below the ~46% modeled (toward the consignment ~40% or below, where Google buys the compute wafers and HBM directly), so the 'OM-accretive from 2027' framing proves an operating-leverage artifact - EBIT margin can rise on low incremental opex while per-dollar value capture stays thin, and the next-gen compute-die content does not lift ASIC GM as bulls argue.

The re-rating from a mid-teens-PE smartphone stock to a ~36x AI-ASIC multiple is a bet on a 2028 EPS number the bulls cannot agree on within a factor of two (Citi NT$207 vs GS NT$406), so the valuation is priced for the ASIC to both land and be MediaTek's to keep (claims 2454-th-001 and 2454-th-004) - after a 3-4x run in 13 months with covering-broker PTs 3-4x'd and both structural bears capitulated near the top, the risk/reward is asymmetric to the downside until the out-year model becomes knowable.

Lotes' general-server socket franchise drives revenue above the street's modelled ramp through FY27: server-related interconnect is now 55.4% of revenue (Q2 FY26 deck) and effectively all of the growth, management reiterates ~20% general-server shipment growth in 2026 with 1H26 server ODM feedback of +8% YoY implying 30-35% YoY in 2H26, share has risen to 35-40% on Intel and 40%+ on AMD platforms, and the AMD Venice ramp (SP7, 7,536 pins, ASP at least +15% vs prior generation on a 23% pin-count increase) plus Intel Oak Stream sustain per-platform content growth into FY27, with server CPU demand visibility extending to 2028.

Gross margin recovers in 2H FY26 off the Q2 FY26 trough of 45.5%: the AMD Venice ramp (ASP at least +15% with pricing that already incorporates higher material costs), a no-price-down base case on existing and new products, and improving mix lift gross margin sequentially in 2H26 as management has guided, arresting the eight-quarter operating-margin decline (34.6% to 22.0%) even if the FY26 EBIT consensus still takes the mechanical 7-15% cut the H1 shortfall implies.

AI-related products become a second content vector: SOCAMM enters mass production in 3Q26 (Lotes is the first supplier for Vera Rubin), QD liquid-cooling and NPO/OE-socket projects reach mass production in 2H27, and AI-related revenue expands from the guided 2-3% of 2026 revenue toward management's ~10% AI-server revenue target for 2027, on a path to double-digit share over 3-5 years.

RISK - the margin compression is structural, not transitory: raw materials (copper, gold, plastic) at ~65% of COGS combined with Lotes' deliberate no-pass-through/stable-pricing strategy and ramp costs (admin +61% YoY, R&D +36% YoY in Q2 FY26) keep gross margin at or below the 45.5% trough through 2H26, extending the eight-quarter operating-margin erosion and forcing the FY26 EBIT consensus cut toward the deeper end (roughly -15% or worse) with FY27's +4%/+11% out-year EBIT raises falling next.

RISK - the implied 2H FY26 revenue acceleration (~NT$22.4bn, +18.6% HoH, needed to hit FY consensus) misses: component supply constraints (CPU, memory, PCB, passives) that already capped Q2 sales at NT$9.5bn below UBSe/Street persist, ODM server-CPU-socket market volumes are rolling over into 3Q (Q2'26 4.38 -> Jul 4.19 -> Aug 3.9 KK/month), PC units decline 10%+ against prior 5-10% expectations, and the NT$5,430mn inventory build (+59% YoY) proves demand-driven accumulation rather than a 2H pre-build.

MARGIN-QUALITY (revision-cycle). Acter holds a >=25% gross margin and closes a meaningful part of the ~1.6x net-margin gap to UIS through FY27, as the rising Taiwan revenue mix (turnkey-heavy, ~10% owner-designated procurement) and the guided above-average US work lift the blended margin - i.e. 1Q26's 25.4% GM is a new base, not a completion-quarter spike. If this holds, EBIT and EPS revisions run above Fubon's already-above-consensus slope on operating leverage (opex only ~5% of revenue). DISCONFIRMING SIGNALS ALREADY ON THE TAPE: FY25 GM FELL 280bps YoY (21.7% -> 18.9%) in a +37% revenue year, and the out-year story rests on one quarter plus one broker's forecast, not a demonstrated annual trend.

TSMC-ADVANCED-ACCESS (structural / business-quality). Acter is - or becomes through FY27 - a qualified vendor in TSMC's leading-edge (N3/N2) advanced-process cleanroom chain, not merely an OSAT / low-end packaging / mature-node contractor. This is the bull leg that would justify the AI-fab quality premium (~33x trailing vs UIS ~25x) the sole broker leans on. The single best independent voice in the corpus (a competitor ex-CEO, T4) directly contests it, calling Acter "primarily OSAT and low-end packaging/testing," a read that is CONSISTENT with Acter's own disclosed order book (memory, packaging/testing, PCB, petrochem).

BACKLOG / CYCLE-DURABILITY (cyclical). The >NT$55bn order backlog (+146% YoY at 1Q26) carries multi-year revenue visibility that sustains 30%+ revenue growth into FY26-27, and the incremental growth leg (US + China) is durable enough to bridge past the management-flagged 2028-2030 Taiwan capex peak rather than air-pocketing when the domestic wave crests. Book-to-bill stays above 1.0x through 2H26.

US RAMP (catalyst). The US entry converts from optionality to booked revenue - Texas turnkey MOUs become formal orders after the 2H26 hazardous-operations license clears, and Phoenix stands up in FY27 for wafer-fab and packaging orders at above-group gross margin and much larger project scale than Taiwan - producing a discrete backlog step-up and a margin-accretive mix shift. This is a single-source, single-broker leg with zero independent or expert corroboration in the corpus.

CYCLE-PEAK + MARGIN-MIX DE-RATE (risk - bear case tracked literally). Acter de-rates from its AI-fab premium multiple (~33x trailing vs the cheaper, larger, higher-margin, more TSMC-anchored UIS ~25x) as the growth premium unwinds: the 1Q26 25.4% GM fades to a structurally capped base (60-70% pass-through cost, China SOE 30-50% owner-designation drag), the mgmt-flagged 2028-2030 Taiwan capex peak pulls forward, and China working-capital / geopolitical drag bites - while the 30%-of-revenue single-client concentration (up from ~13% FY24) creates air-pocket risk on that project's completion. As the smallest, thinnest-margin listed peer on the highest multiple, Acter leads the cohort down first.

AI-server passives remain a small, back-end-loaded optionality rather than a FY27 earnings driver: the AI-server + nearline basket stays ~10-12% of company profit (Citi 2026-06-29 sizes it just under 10% FY27, ~12% FY28), TDK's own IR dates the AI-DC passives inflection to FY28-FY29 and concedes TDK is NOT winning the GPU accelerator-board inductor socket (an Asian maker took share; MS IR 2026-06-17), so through FY27 the AI leg does not scale into the earnings the re-rate implies. The one genuine under-modelled leg is TDK-EPCOS top-share AI-server-PSU aluminium electrolytics (~JPY100bn alu+film, aluminium +10%+), not the crowded GPU-board trade - and even CLSA calls TDK "very marginal in AI servers ... more active only ~2029" (exposure 6762-em-004).

TDK does not monetise the AI-MLCC supercycle: it is auto-indexed (74% of TDK MLCC revenue is automotive; exposure 6762-em-001), pricing is benign and gently declining like-for-like with management explicitly NOT putting through price increases (MS IR 2026-06-17), and it sits outside the >80% AI-grade MLCC supply held by Murata + SEMCO (JPM MLCC AMA 2026-06-16) - a self-professed laggard "catching up" via high-voltage. The AI-MLCC demand inflection is real but is a Murata / SEMCO / Yageo story, so TDK's MLCC line does not deliver the price-up / mix uplift the sector narrative implies through FY27.

FY27 is a rechargeable-battery profit trough, not a mix-driven re-acceleration, and the ~90%-of-OP smartphone-battery concentration is the material near-term headwind the AI re-rate looks past: management guides FY27 rechargeable batteries alone to a YoY profit DECLINE on ~-7% volume (MS 2026-04-28) with smartphone units guided ~-10% on the memory shortage (Citi 2026-06-29), while the offsetting mix / tech (metal-case doubling, silicon-anode, medium / BBU-UPS batteries) is dated to mass-produce F3/27 and ramp F3/28 - i.e. it lands in FY28, not FY27. Energy is ~90% of consolidated OP and ~85% of it is IT smartphone / PC (exposure 6762-em-003), so the trough hits the whole company.

HDD magnetic heads and suspensions are the cleanest, best-corroborated multi-year growth line and anchor the base-case compounding: heads are guided ~+40% YoY FY27 and suspensions ~+20% (MS 2026-04-28) on US captive-maker (Seagate / WD) resumption from Jan-Mar 2026 (exposure 6762-cust-001 / 6762-cust-002), suspension tightening on competitor withdrawal, and HAMR value-add from 2027 - with management flagging the guide as deliberately conservative (upside skew; MS IR 2026-06-17). This is a recovery / share story, not secular growth, but through FY27 it delivers real profit growth independent of the AI-passives premium.

The ~2x, six-house PT re-rate is multiple expansion on cut near-term numbers, not rising near-term earnings: the biggest moves are discount-rate cuts and base-year rolls (Citi risk premium 4.5%->3.0%, base year FY28->FY29; JPM PT 3,200->5,500 while CUTTING FY27 EPS -9%; MS marks P/B 2.9x vs a 1.5x historical average), the houses disagree ~1.8x on value (GS / MS ~3,000-3,200 vs JPM / UBS 4,950-5,500) on the same numbers, and the AI leg backing the top FY29-based targets is ~10-12% of profit and dated FY28+. Risk / reward has therefore deteriorated regardless of business quality until the AI-profit mix clears ~12% and FY28-29 estimates converge.

BULL-CASE RISK CLAIM (tracks the Section H bull literally; confirmations are BAD for the skeptical core and challenge th-001 / th-005): AI-server passives inflect early and big - the aluminium-electrolytic + power-inductor + high-voltage-MLCC basket clears ~12-15% of company profit by FY28 WITH a NAMED vertical-power-delivery / board-embedded inductor design-win (the Section H bull gating KPI) - converting the AI leg from ~10-12% optionality into a genuine earnings driver, so the FY29-based 25-30x multiple gets earnings underneath it and TDK is re-rated as a real AI-passives beneficiary rather than a lower-vol MLCC proxy.

The AI-driven SoC-tester up-cycle is structural rather than a cyclical peak rolling into FY3/28: through FY27 Advantest keeps printing beat-and-raise, with the company/consensus SoC-tester TAM guide (raised to $8.7-9.5bn for FY27) and FY27-FY28 operating-profit estimates revised UP as AI-accelerator plus HBM test intensity compounds and new CPU/CPO insertions add demand legs, so the earnings-revision slope stays positive and more durable than the bears' single forward-timing 'peak into FY3/28' thesis (owned mainly by GS).

Advantest retains its high-end SoC/GPU tester dominance through FY27: dual-sourcing at hyperscalers (Amazon/Broadcom/Google-TPU) stays low single-to-double-digit qualification share rather than converting into durable Teradyne share, because the ~3-year device-power-supply lead (XHC32/XPS256), >10,000-unit V93000 installed base and one-stop test-cell keep new leading-edge GPU/ASIC sockets landing at Advantest; reported SoC-tester share holds roughly flat even as the market grows ~+30% YoY.

BEAR / RISK CLAIM (tracked literally; confirmations are BAD for a long): Teradyne's HBM4 tester (Magnum-class, eval-approved) erodes Advantest's memory-test dominance - HBM4 base-die and stack test flips from roughly 50/50 toward Teradyne at one or more of SK Hynix/Samsung/Micron, and Advantest's memory-tester revenue undershoots the $2.2-2.7bn FY27 guide as dual-sourcing takes hold in the segment where Advantest is most exposed.

Advantest's margins stay structurally elevated well above management's ~25% OPM / ~53% GM through-cycle MTP3 model through FY27: favourable SoC/HBM mix and operating leverage hold gross margin in the low-to-mid 60s and operating margin in the low-to-mid 40s, and blended ASP holds flat-to-+5% YoY rather than eroding, so the current 44% OPM / 64% GM proves more than a sold-out-capacity peak.

BEAR / RISK CLAIM (tracked literally; confirmations are BAD for a long): the stock is fully priced after a >3x re-rate (forward P/E ~45x, EV/EBITDA ~32x; UBS lifted its multiple 20x->27x), the bull case is near-unanimous with only GS left at Neutral and no bear bench to convert, so the re-rating is exhausted - absent continued upward EPS revisions the multiple de-rates, and a single demand guide-down (Theme 1) exposes a high-multiple stock with no valuation cushion.

Co-packaged-optics and adjacent new test legs (CPO testers, die-level / known-good-die probers, SiConic EDA<->ATE) become a real incremental Advantest growth leg from FY28: Advantest converts its silicon-photonics-test positioning into a named CPO-test design win and first disclosed revenue, rather than the leg accruing to Teradyne (customer-flagged fiber-alignment edge) or staying a sell-side narrative with no revenue.

AI-server MLCC is a genuinely inflecting leading KPI: management-guided AI-server MLCC sales grow ~+80% YoY in BOTH FY27 and FY28 (a ~3.2x two-year expansion off the +30-35% FY26 base), and 6976 holds its ~20-25% share of AI-server logic-board MLCCs (~2x its company average; #2/#3 with Murata and SEMCO per exposure 6976-em-013) as GB300 ramps into Apr-Jun 2026 and Rubin adds +50-60% in 2027. The slice is small (~3% of sales) but the growth is real and structural (SemiAnalysis: AI-server MLCC demand 4.8x CY25-CY27, lead times 26-40wk), not a 2020-21 over-order.

The AI-server + miniaturisation mix converts to visible margin from FY27 H2: as the small-case high-capacitance parts ramp (1005/47uF, 1608/100uF from 2H FY26) and capex sits below depreciation (JPY41.1bn vs JPY49.1bn; post-over-build digestion, exposure 6976-capex-013), operating leverage lifts consolidated gross and operating margin above the JPY30bn / 7.8% FY27 OP guide and onto the bull path (JPM EBIT margin to 17.5% by FY28), beating the ~JPY35.7bn consensus. The +80% AI growth shows up in the numbers, not just the mix.

The MLCC same-product price-decline drag eases into a profit tailwind: at ~95% utilization the like-for-like price decline slows from the recurring ~JPY19-20bn/yr OP drag (-JPY20.2bn FY25 to -JPY18.9bn FY27 plan; exposure 6976-elast-001) toward stabilization in FY27, and supply/demand tightness re-opens the door to actual S/D-driven MLCC price hikes from FY28 (Citi's tell), materially reducing the pricing headwind to margin.

The upside is largely priced and the earnings-revision cycle is near exhaustion: with the stock at ~JPY8,145 trading ABOVE five of six covering brokers' price targets (only JPM's JPY9,500 sits higher; Citi JPY2,900 to MS JPY4,700 to GS JPY7,100) and the FY27 guide having landed ~16% below consensus, the easy multiple-expansion leg is done - absent delivered margin proof, forward estimates and the multiple compress rather than expand and the shares de-rate toward the cluster of bear targets.

A material share of the earnings "recovery" is weak-yen FX, not operational MLCC-margin improvement, and the FX risk is one-sided: at ~JPY0.9bn OP per JPY1 USD/JPY (rising, exposure 6976-elast-001) with spot ~JPY150-155 versus the medium-term plan's JPY140 assumption, roughly a third to a half of the guided FY27 JPY30bn OP is yen rather than operations (BofA: "guidance lift driven by FX and fixed-cost cut; MLCC margin gains still lagging"), so yen appreciation toward JPY140 would erase much of the OP ramp.