At a glance. FY2025 revenue NT$950.7bn (~US$30.2bn) and EPS NT$275.06 were records. AI crossed 50% of sales; inside AI, ASIC was about 90%. Q4 gross margin fell to 7.2% on a GB200 GPU wave that then ended. New NVIDIA and AMD GPU platforms, and new ASIC platforms, are guided to 2H 2026. The call never said 800 V.
What changed this quarter
AI became half the company. 2025 split roughly 50/50 AI versus general compute. Inside AI, ASIC carried almost all of it. GPU was the small slice — and the slice that printed in Q4 as GB200, which management said had rolled off by the time of the call.
The 2026 clock is a platform change. Growth is guided to the second half, when new ASIC and new GPU platforms start to ship. AI mix is guided a little over 50%; exact GPU/ASIC weights were refused because they move month to month. NVIDIA is still expected to be the larger GPU contributor; AMD is in, on a double-wide design.
Cash policy flipped with the BOM. 2025 dividends go stock NT$20 plus cash NT$145 a share, not cash-only, so more cash stays in the company. A mandate of up to 18 million new shares was approved for flexibility. GPU consignment is still being talked, with a hope of an answer before the next generation ships.
The numbers
| Q4 2025 | Value | Change |
|---|---|---|
| Revenue | NT$292.4bn (~US$9.3bn) | +9.6% QoQ, +152.9% YoY |
| Gross profit | NT$21.1bn, margin 7.2% | vs 8.8% in Q3, 9.4% a year ago |
| Operating profit | NT$16.5bn, margin 5.6% | vs 7.3% in Q3 |
| Net income | NT$13.8bn, margin 4.7% | −10.5% QoQ, +95.7% YoY |
| EPS | NT$74.21 (~US$2.36) | vs NT$82.92 in Q3, NT$37.92 a year ago |
| FY 2025 | Value | Change |
|---|---|---|
| Revenue | NT$950.7bn (~US$30.2bn) | +163.7% YoY |
| Gross profit | NT$78.5bn, margin 8.3% | vs 10.4% in 2024 |
| Operating profit | NT$63.9bn, margin 6.7% | +127.5% YoY |
| Net income | NT$51.1bn, margin 5.4% | +124.4% YoY |
| EPS | NT$275.06 (~US$8.73) | vs NT$126.57 in 2024 |
2025 capex about NT$13bn. 2026 capex guided materially higher, including Malaysia land and a new Taiwan HQ. Bank borrowings jumped in Q4 with the working-capital book.
NVIDIA 800V read-through
Wiwynn is a rack ODM for CSPs. It is not on NVIDIA's published 800 VDC list. This call never named 800 V, HVDC or Kyber.
The 800V read is the cabinet and the factory, not a dated HVDC SKU:
- GB200 already shipped in Q4 2025 and then left the mix — a 2025 GPU rack, air- or liquid-cooled, still on today's voltage.
- Next NVIDIA (and AMD) GPU platforms are guided to 2H 2026. That is the Rubin-class clock other names are on, still before Kyber.
- Level-12 full-rack work is tenders and sovereign AI, not the default hyperscaler SKU. Management said it would pick customers and countries carefully.
- Air- and water-cooling capacity are both in. Factory power is planned to 2028 across Taiwan, Malaysia, Mexico and Texas (shipping since end-2025). That is MW at the meter for test and production, not NVIDIA's 800 VDC bus.
A 25 May 2026 AGM, after this note, did mention HVDC, CPO and cooling as ecosystem work. Do not read that sentence back into this call. Q2 2026 (7 August) was a board results release, not a 法說會.
Segments that moved
| Line | 2025 | 2026 guide |
|---|---|---|
| AI servers | ~50% of sales; ASIC ~90% of AI | A little over 50%; ASIC still the bulk; new GPU + ASIC in 2H |
| General compute | ~50% | Units +20–30%; dollars also up. 2025 was front-loaded; 2026 is described as even through the year |
| GPU (NVIDIA vs AMD) | Small AI slice; Q4 GB200 wave | NVIDIA still larger; AMD double-wide, first-cut ASP talked as approaching 2× GB200 (~US$5m/rack in that comparison), similar gross-margin rate |
Memory is about 30% of general-server cost after the latest prices, and a much smaller share of AI servers because the GPU/ASIC dollar dominates. Price is passed through; the hit is to gross-margin percent, not to absolute profit or operating-margin dollars, management said.
Guidance
- FY 2026 revenue up on 2025, with the step in the second half on the platform change.
- Q1 2026 up year on year; sequentially hard against a record Q4, but "not bad".
- AI mix a little over 50%. General-server units +20–30%.
- Gross-margin percent may only be a bit better than Q4, not back to old levels, because memory is up even if GPU mix is down in H1.
- Operating target: EPS and absolute profit, not the operating-margin ratio, which GPU BOMs mathematically dilute.
- 2026 capex well above 2025's ~NT$13bn, and rising over the next few years as customers keep building halls — still small versus the customer's own spend, and stretched over time.
Management Q&A
Q: How did 2025 split across ASIC, NVIDIA GPU and general servers, and where is 2026 growth — first half or second?
A: 2025 was about half AI, half general. Inside AI, ASIC was about 90%. 2026 AI mix should be a little over 50%; exact line weights move too much month to month to pin down. The second half is stronger because new ASIC and GPU programmes start to ship.
Q: How should 2026 gross margin and operating margin be read, especially if H1 has less GPU?
A: Margin percent is not only GPU mix — memory and other components are up. Even with a lighter GPU mix in the first half, gross margin may only be slightly better than Q4 2025, not back to old levels. The absolute profit and operating-margin dollars are described as much less affected; the company said it is confident operating margin can stay stable in that sense.
Q: What share of cost is memory in general servers versus AI servers?
A: In general servers, after the latest prices, about 30%. In AI servers the GPU/ASIC dollar is so large that memory is a smaller share. Prices can still move every month or quarter; near-term supply was described as covered, not a shortage.
Q: How much do general servers grow in 2026?
A: Both dollars and units are guided above 2025. 2025 shipments were front-loaded; 2026 is described as steady through the year. On a follow-up, unit growth of 20–30% was called a reasonable range.
Q: On GPU AI servers over the next two years, which is larger — NVIDIA or AMD?
A: NVIDIA is still ahead in the market, so its share is expected to stay larger than AMD's. AMD is trying to come back; Wiwynn said it is already working with it.
Q: Can Wiwynn add a new ASIC programme in the next two years?
A: Talks with each customer are ongoing. Details were not disclosed.
Q: Can GPU programmes move from buy-and-sell to consignment, and by when?
A: It has been in discussion since GPU programmes began. The hope is a firm answer before the next-generation product ships. It was described as not easy.
Q: How is Level-12 (full-rack) progressing, and what about power?
A: Level-12 demand is mainly tenders or sovereign AI; the company said it would keep watching and choose customers and countries carefully. Factory power is planned to 2028 across the global sites, in close discussion with customers so production can be met.
Q: What is 2026 capex?
A: Materially above 2025's about NT$13bn. Some of that is execution of projects already announced, including Malaysia land. As customers keep building data centers, capex is expected to keep rising over the next few years, still small versus the customer's own spend and stretched over a long tail.
Q: How should operating-margin ratio be read for the next few years?
A: GPU and other high-ASP parts raise the material cost base, so the percentage is diluted by arithmetic. The operating target is EPS and absolute profit, not defending the ratio.
Q: Does memory inflation change CSP pull on general servers, and can cost be passed through?
A: Management said CSPs' data-center build is not turned off by memory prices, and many already have long-term memory contracts. Wiwynn's quote separates material cost from conversion profit, so material moves can be passed through.
Q: How does an AMD platform compare with GB200 on ASP and gross margin?
A: AMD's design was described as double-wide. A first-cut ASP was talked as approaching twice GB200 — about US$5 million per rack in that comparison. Gross-margin rate was expected to look similar to GB200.
Disclaimer
English notes on Wiwynn's investor conference of 26 February 2026, covering Q4 and FY 2025. Figures and statements are as presented by management, are unaudited, and have not been independently verified. USD equivalents are approximate at about NT$31.5 to the dollar. For reference only; not investment advice.