Wiwynn Corporation
ODM-direct: designs and manufactures hyperscale servers, storage and full AI racks sold directly to cloud service providers (no channel, no own-brand premium); low-margin, high-volume, working-capital-intensive pass-through of GPU/component costs with value-add in system design, rack integration and liquid cooling
Earnings, margins, COGS & capex
FY2025 was a step-change year: revenue rose 163.7% to NT$950.7B with AI-related products exceeding half of revenue for the first time; PAT rose 124.4% to NT$51.1B (5.4% net margin) and EPS reached NT$275.06 vs NT$126 in 2024 (+117.3%). Momentum carried into Q1 2026 (revenue NT$276.5B, +62.0% YoY; PAT NT$14.1B, +44.1%; EPS NT$75.95) -- all record highs for a first quarter. Percentage margins compress as GPU pass-through content grows, but absolute profit dollars are scaling faster than opex. The May-2026 AGM approved distribution of 2025 earnings as a NT$145/share cash dividend plus a NT$20/share stock dividend (vs NT$74 cash for 2024). The model is working-capital-heavy: nearly 95% of receivables sat with three customer names as of Q3 2025.
Income statement — where each revenue dollar goes
% of revenueOf every $1 of revenue, ~92¢ is cost of goods and ~2¢ operating expense, leaving ~7¢ of operating profit (~5¢ net).
Revenue trend
Margins
declining in % terms on AI rack mix (GPU cost pass-through), gross profit dollars rising sharply
stable-to-slightly-lower; opex leverage partly offsets mix
stable
COGS structure
Dominated by pass-through components: NVIDIA GPUs/NVL72 rack content, ASIC accelerators, memory (HBM/DDR), CPUs, power and liquid-cooling hardware. Component cost inflation (memory in particular) and GPU allocation are the swing factors; Wiwynn carries inventory risk on very expensive racks, making COGS timing and working capital the key financial risks rather than unit economics.
Capex
Capex tripled to roughly NT$18B (announced Mar-2025) for the global footprint build: a ~US$300M Texas plant (now ramping, per the Q4 2025 release) for tariff-resilient near-customer assembly, a new Kaohsiung (Lujhu) factory, expanded Malaysia (Johor, including the YTL AI data-center engagement) and Mexico capacity, and the Czech Republic serving Europe; 2026 expansion continues across the Americas and Asia with a further European site planned. Capex intensity stays low (~1-2% of revenue) because ODM assembly is far less capital-intensive than fabs -- the real capital demand is working capital.
Latest earnings
Formal consensus beat/miss not disclosed; sell-side stance is Strong Buy (18 analysts, zero sells) with average target ~TWD 6,914, ~37% above the 2026-07 price (StockAnalysis/MarketScreener)
Company expects shipment volume to keep achieving double-digit growth in 2026 with total profit unaffected by partial business-model changes; President William Lin (DigiTimes, Jun-2026) said he sees no AI bubble for at least four years and that expansion still lags order strength. April 2026 monthly revenue was NT$82.7B, +29.7% YoY. Next earnings: Q2 2026 results expected around early Aug-2026 (date not confirmed)
- FY2025 EPS
- NT$275.06 (vs NT$126 FY2024, +117.3%)
- FY2025 PAT
- NT$51.1B, +124.4% YoY
- AI-related products share of FY2025 revenue
- >50%
- Dividend (2025 earnings, approved May-2026 AGM)
- NT$145/share cash + NT$20/share stock (2024: NT$74 cash)
- Trailing / forward P/E
- ~17.5x / ~13.3x (StockAnalysis, 2026-07-10)
Growth drivers
- Hyperscaler AI capex supercycle — rack-scale GB200/GB300 NVL72 system shipments to US CSPs (GB300 deliveries began H2 2025; Vera Rubin NVL72 infrastructure showcased at GTC 2026)
- Custom ASIC AI servers diversifying beyond NVIDIA-based racks
- Liquid cooling and power-density leadership (OCP ecosystem, direct-to-chip designs) as air cooling hits limits
- Geographic capacity expansion near customers (US/Texas, Malaysia, Mexico, Czech Republic, Europe planned) winning tariff-sensitive orders
- Customer diversification beyond Meta/Microsoft toward additional CSPs and sovereign/regional AI builders (YTL in Malaysia: initial 100-200 GB200 NVL72 racks plus co-development of a 600MW Johor AI data-center complex)
Bull & bear
Wiwynn is the cleanest listed pure-play on hyperscaler AI rack deployment: revenue tripled in 2025, shipments are guided to keep growing double-digit on that base, ASIC servers add a second engine, and at ~13x forward earnings with a ~0.55 PEG the market is pricing a cycle, not a multi-year infrastructure buildout.
- FY2025 revenue +163.7% to NT$950.7B and EPS +117.3% to NT$275.06; Q1 2026 still +62% YoY with record Q1 profit -- execution at scale is proven, not promised
- Valuation is undemanding for the growth: ~17.5x trailing, ~13.3x forward P/E, 0.89x sales, PEG ~0.55; consensus Strong Buy (18 analysts, zero sells) with ~37% implied upside to the ~TWD 6,914 average target (2026-07)
- ASIC AI server programs diversify away from NVIDIA-allocation risk and deepen hyperscaler lock-in (custom programs are stickier than merchant GPU racks)
- Liquid-cooled rack-scale integration (GB300 NVL72, Vera Rubin NVL72 next) is exactly where value-add per rack is rising; Wiwynn is in the first wave with only a handful of qualified peers
- Global footprint (Texas ramping, Malaysia, Mexico, Czech, Europe planned) converts tariff chaos into a competitive moat and opens sovereign-AI customers like YTL (600MW Johor complex, won from Supermicro)
- Management sees demand outstripping its own aggressive expansion for years (President William Lin: no AI bubble for at least four years); a NT$145 cash dividend on 2025 earnings shows profitability is real cash, not just accounting
A 7-8% gross margin assembler whose fate rests on the uninterrupted capex of about three customers, carrying billions in GPU inventory on a newly net-debt balance sheet, is priced for continuation of the steepest hardware cycle in history -- any hyperscaler digestion year, insourcing move, or NVIDIA transition air pocket hits revenue and working capital simultaneously.
- Customer concentration is existential, not cosmetic: nearly 95% of receivables with three names and ~98% of 2024 revenue from three customer groups; Microsoft or Meta trimming one product cycle can erase a year of growth
- Gross margin fell from 8.3% (FY2025) to 7.6% (Q1 2026), and Q4 2025 printed just 7.2%, as AI mix rose -- growth is structurally margin-dilutive in % terms, so the P/E deserves an ODM discount, not a growth premium
- Working capital is the silent risk: net debt ~TWD 36B, a US$500M capital injection into the operating subsidiary, a shelf for up to 18M new shares (potential dilution), and NVL72 rack inventory that costs millions per unit; an order pushout strands very expensive inventory
- Hyperscalers actively dual-source rack integration (Foxconn, Quanta, Celestica, Dell for ASIC programs) and could insource more system design; ODM moats are shallow by construction
- AI capex is cyclical even in secular uptrends -- a 2027 digestion year after the GB300 wave, or a shift of spend from training racks to power/land, would compress both revenue and the multiple simultaneously; April 2026 monthly revenue growth already slowed to +29.7% YoY vs Q1's +62%
- Tariff-regime changes force costly production reshuffling across the Mexico/Czech/Malaysia/Texas footprint; policy risk on Taiwan-linked hardware remains unhedgeable
What it is worth
Earnings multiple on TWSE-listed shares (trailing ~17.5x, forward ~13.3x P/E; ~0.89x sales, PEG ~0.55 per StockAnalysis, 2026-07-10); cross-checked against sell-side consensus (Strong Buy, 18 analysts, average target ~TWD 6,914, range TWD 6,000-8,500)
TWD ~3,000-3,800: a 2027 hyperscaler digestion year or a major customer insourcing/dual-sourcing shift cuts EPS growth to flat-negative; multiple compresses to ~10x on cyclical ODM framing with working-capital strain
TWD ~5,000-5,800: current-price zone; double-digit growth with continued % margin drift, multiple stays ~13-15x forward as cycle-durability is debated
TWD ~6,900-7,500: consensus-target zone; assumes 2026 double-digit shipment growth lands, ASIC ramp adds a second engine, and operating margin holds ~6%+ -- ~15-16x on rising forward EPS
The debate is whether a ~3-customer, 7-8%-gross-margin ODM deserves a growth multiple. Bulls anchor on consensus EPS growth (PEG ~0.55) making ~13x forward cheap; bears argue ODM cyclicality caps the multiple regardless of near-term growth, and note the approved 18M-share capital-raising shelf plus the NT$20/share stock dividend add share-count growth. USD investors also carry TWD exposure. Market cap ~US$32B at TWD ~937-943B (2026-07-09/10).
SWOT
Strengths
- Pure-play exposure to hyperscale AI infrastructure with direct (no-middleman) relationships to Meta and Microsoft, plus other US CSP engagements
- Among the first ODMs shipping NVIDIA GB300 NVL72 rack-scale systems (one of a handful of NVIDIA-certified rack ODMs); deep liquid-cooling and OCP design credibility
- Profitable at scale — 6-7% operating margin is at the high end for server ODMs, with EPS more than doubling in 2025
- Wistron parentage gives manufacturing depth while Wiwynn keeps the focused, CSP-only business model
Weaknesses
- Extreme customer concentration — nearly 95% of receivables with three customer names (Q3 2025); three customer groups were ~49%/28%/21% of 2024 revenue -- loss or insourcing by one customer is a material hit
- Structurally thin gross margins (7-8%) with GPU cost pass-through diluting percentage profitability
- Working-capital intensity has flipped the balance sheet to modest net debt; inventory on GPU racks is expensive to carry and risky if orders slip
- No pricing power over NVIDIA allocation or memory costs; margins are hostage to component cycles
- Taiwan-listed only (no US ADR), limiting some US institutional access and adding TWD currency considerations
Opportunities
- Custom ASIC AI server ramp (hyperscaler in-house accelerators need ODM rack integration just as GPUs do)
- Sovereign/regional AI datacenter buildouts (YTL Malaysia 600MW Johor complex as template, including full data-center buildout scope) widening the customer base
- Liquid cooling attach and rack-power innovation raising ODM value-add per rack
- US manufacturing expansion (Texas plant ramping) positions it to win tariff-driven onshoring orders
- Inference buildout wave could prove larger and longer than training capex
Threats
- AI capex digestion or pause at Meta/Microsoft would hit revenue with little cushion
- Tariff and trade-policy volatility (Texas/Mexico/Czech/Malaysia footprint hedging against tariff scenarios)
- Competition from Foxconn, Quanta, Supermicro, Dell and Celestica for the same hyperscaler rack programs
- NVIDIA architecture transitions (product gaps between Blackwell and Vera Rubin platforms) create shipment air pockets
- Component shortages — power semis, HBM/memory, cooling parts (chairman has warned publicly of AI server component shortages) -- and datacenter power constraints throttle deliveries
Moats, dependencies & bottlenecks
Moats
Qualified direct-ODM relationships with Meta and Microsoft (multi-generation design wins) moderate-strong Qualification cycles and co-design create switching friction, but hyperscalers deliberately dual-source; the moat is a seat at a small table, not exclusivity
Rack-scale liquid-cooling integration expertise (GB200/GB300 NVL72, OCP designs) First-wave capability with few qualified peers today; erodes as Foxconn/Quanta/Celestica scale the same capability
Malaysia, Czech Republic, Mexico, Taiwan) Tariff-era geography is genuinely hard to replicate quickly and is winning orders now
Scale matters in ODM but confers no pricing power against hyperscaler procurement
Dependencies
Three customer groups were ~49%/28%/21% of 2024 revenue; nearly 95% of receivables concentrated in three names (Q3 2025)
supplier/platform GB300 NVL72 rack content defines the current cycle; allocation and architecture-transition timing (Blackwell Ultra to Vera Rubin) gate shipments
Samsung, Micron MU; power semis) Component shortages and memory price inflation squeeze thin margins and stretch lead times; chairman has publicly warned of AI server component shortages
Shared heritage and some capacity interplay; also a related-party consideration
Footprint decisions (Texas plant, Czech/Malaysia capacity) are explicitly framed by management as tariff countermeasures
Power, not demand, increasingly gates rack deployment schedules
Advantages
- Pure-play AI infrastructure exposure with proven 160%+ growth execution
- Direct ODM model strips out brand-vendor margin stack -- structurally cheapest route for hyperscalers
- First-wave GB300 NVL72 and liquid-cooling shipper
- Tariff-resilient multi-region footprint including ramping US (Texas) capacity
- Profitable and dividend-paying through the ramp (rare among AI-infra pure plays)
Weaknesses
- ~3-customer revenue base; minimal bargaining power on either side of its P&L
- 7-8% gross margin leaves no room for execution error or component inflation
- Net-debt balance sheet in a rising working-capital cycle
- No US listing or ADR; TWD-denominated for US investors
- Value-add per rack can be competed away by larger EMS rivals (Foxconn) willing to take lower margin
Bottlenecks
- GPU and ASIC accelerator allocation from NVIDIA and hyperscaler silicon programs
- Working-capital financing capacity for multi-million-dollar racks in inventory (US$500M subsidiary injection; 18M-share capital-raising shelf approved)
- Liquid-cooling component and power-shelf supply
- Skilled rack-integration capacity in US/Europe sites (new plants ramp slower than orders)
- Customer datacenter power and shell readiness pacing deliveries
Top signals & trends
Top signals
Demand still exceeds even aggressive capacity expansion, per management
Management conviction backed by continued multi-region capacity adds
Mix-driven; watch whether dollar gross profit growth keeps outpacing opex
Working-capital strain is the canary for ODMs in a downcycle; the shelf adds potential dilution
18 analysts, zero sells; conviction intact after the 2025 tripling
Cost and complexity now, moat and share gains if executed
Trends
strongly positive · Direct beneficiary; management sees 4+ years of runway
Raises ODM value-add and favors early liquid-cooling integrators
ASIC racks still need ODM design-manufacture; a second engine beyond NVIDIA platforms
Costs rise but Wiwynn's multi-region footprint converts it into share gains
Thin-margin pass-through model absorbs timing risk on cost spikes
Gates near-term deliveries but extends the buildout's duration
Ecosystem & competitor graph
Suppliers feed the company; customers pull from it. Line thickness shows the strength of each tie (supply-chain dependency, customer earnings contribution). Hover to isolate a tie.
GPU/rack platform (GB200/GB300 NVL72, Vera Rubin NVL72 next); the defining supply dependency
Server CPUs
Server CPUs and Instinct accelerators; also owns the former ZT Systems design arm
Korean-listed; memory cost inflation is a margin swing factor
US memory alternative
Power/cooling infrastructure ecosystem around rack deployments
Power supplies and cooling components
Anchor CSP customer (servers, storage, AI rack programs)
Anchor CSP customer (Azure fleet, custom accelerator racks)
CSP engagement per press coverage; customer names are not individually disclosed by the company
CSP engagement per press coverage; not company-confirmed
GB200 NVL72 order (initial 100-200 racks) plus co-development of a 600MW Johor AI data-center complex -- first major non-US-CSP diversification win, taken over from Supermicro
Largest EMS globally; aggressive in GB200/GB300 rack assembly for the same CSPs
Fellow Taiwan ODM-direct leader in hyperscale servers; closest structural comp
US-based AI server systems vendor competing for GPU rack deployments; lost the YTL Malaysia order to Wiwynn in late 2024
Branded AI server share gains at CoreWeave/xAI-type buyers; less overlap in pure hyperscaler ODM-direct
Key rack integrator for hyperscaler ASIC/networking programs (e.g. Google ecosystem)
AI systems (Cray) mostly in enterprise/sovereign HPC rather than CSP ODM-direct
Taiwan ODM in server boards/systems; smaller AI rack presence
Parent handles some AI server (GPU baseboard) business itself -- partial sibling overlap