At a glance. The notebook company is a server company by mix: servers above 80% of sales, AI 75–80% of servers. Q2 revenue NT$1.04tn (~US$32.9bn) and EPS NT$7.43 were records. CFO 楊俊烈 called 2027–2028 very exciting and raised 2026 capex from NT$30bn to NT$40bn. The call never said 800 V.
What changed this quarter
The AI line is now the company. Servers have been more than 80% of sales since Q1. AI's share of that server line held 75–80% in the first half and is guided to 80% in the second. Full-year AI-server revenue is still guided to double.
Capacity was dated past this year. End-2026 AI-server capacity is still guided to double versus end-2025. The NT$19.7bn AUO Huaya plant in Guishan is the 2028 backend-assembly step, next to R&D HQ and the existing server lines. Management's test-power sentence: the building already has high-voltage supply from its panel days, which it says fits outgoing test of AI servers. That is plant power, not NVIDIA 800 VDC.
Consignment started, in a small way. From Q3 2026 some high-ASP parts on GPU and ASIC programmes move from buy-and-sell to consignment, project by project. Mix is still described as low. The point is cash and the gross-margin percent on expensive chips, not a change in who builds the rack.
The numbers
| Q2 2026 | Value | Change |
|---|---|---|
| Revenue | NT$1,036.6bn (~US$32.9bn) | +28.1% QoQ, +105.6% YoY |
| Gross profit | NT$52.1bn (~US$1.65bn), margin 5.02% | +34.6% QoQ; margin +24 bp QoQ, −202 bp YoY |
| Operating profit | NT$34.2bn (~US$1.09bn), margin 3.30% | +48.7% QoQ |
| Net income | NT$28.7bn (~US$910m), margin 2.76% | +35.2% QoQ, +69.9% YoY |
| EPS | NT$7.43 (~US$0.24) | vs NT$5.50 in Q1, NT$4.37 a year ago |
First half: revenue NT$1,845.8bn (+86.5% YoY), net income NT$49.8bn, EPS NT$12.93. Q2 capex NT$13.6bn; first-half capex about NT$21.8bn, more than 80% Thailand and California. Operating-expense ratio 1.72%. Inventories NT$790.2bn; receivables NT$765.8bn — both up with the book, with turnover days described as shorter.
Weighted-average shares used for EPS: 3,854.5 million.
NVIDIA 800V read-through
Quanta is a rack ODM. It is not on NVIDIA's published 800 VDC list, and this call never named 800 V, HVDC or Kyber.
The 800V read is the factory, not a SKU. A 1 MW-class cabinet needs a plant that can feed and test it. Huaya's "high-voltage supply" is a reused panel fab's incoming power, sold as a test advantage for AI-server outgoing inspection. Do not upgrade that sentence into NVIDIA's single-ended 800 V rail.
What the call did date is 2026–2028 capacity: double by year-end 2026, Huaya in 2028, US subsidiary QMN topped up by US$973m for California assembly and test. Platform transition (old GPU to new GPU) was described as smooth, with one to two months of handover and no material gap in shipments. That is a 2026–2027 GB / Rubin clock, the same clock Foxconn is on — still one generation before Kyber.
NeoCloud GPU and ASIC racks were in small volume. Large CSPs remain the main AI-server book.
Segments that moved
| Line | Q2 / H1 2026 | Notes |
|---|---|---|
| Servers | >80% of sales | AI 75–80% of server revenue; guided to 80% in H2 |
| AI servers | double for FY2026 (guide) | Q3 still +QoQ at a double-digit rate, slower than Q2 |
| General servers | double-digit growth | CPU demand; Q3 described as strong |
| Notebooks | 11.5m units in H1, −5% YoY in Q2 | Q3 −20%+ QoQ; FY double-digit unit decline |
| Automotive | double-digit QoQ and YoY | Mix compressed to low single digits by AI servers |
Guidance
- FY 2026 AI-server revenue: double. General servers: double-digit growth. Notebooks: double-digit unit decline.
- Q3 AI-server revenue still up sequentially, below the Q2 rate. Notebooks down more than 20% QoQ; H2 units below H1.
- Full-year profit rates described as able to hold first-half levels, with mix and operating leverage doing the work — the target is operating profit and margin dollars, not gross-margin percent.
- Capex NT$40bn for 2026, from NT$30bn. End-2026 AI capacity still on the double versus end-2025 path.
- Huaya close is still under contract; production contribution 2028. Taiwan AI backend capacity is then guided to double versus today.
- Germany QCG: €19.5m for smart-driving lines. US QMN: US$973m for California AI assembly and test.
Management Q&A
Q: How much of the book is moving to consignment, and does it apply to every customer?
A: Not every customer. High-ASP key components on some GPU and ASIC programmes move from buy-and-sell to consignment from the third quarter, decided project by project. The second-half consignment mix is still described as low.
Q: Peers have raised 2027/2028 views — how does Quanta see the server cycle?
A: Management said it is very exciting about 2027 and 2028. The concrete example given was buying the AUO plant specifically for 2028 capacity. Server order visibility is talked about as more than a year, and as stretching to 2028.
Q: Does the old-to-new GPU platform transition cut shipments?
A: It was described as a smooth ramp. There may be one to two months of handover. In principle there should not be a large pause or gap in shipments.
Q: If component prices are rising in the second half, why can profit rates hold the first half?
A: Customers and booking points are diverse. The operating target is operating margin and margin dollars, not gross-margin percent, which moves with the revenue base and mix. Operating leverage is also cited as support for second-half profits.
Q: How competitive is the NeoCloud book, and what is the capacity plan?
A: Only a handful of ODMs can serve the large CSPs; NeoCloud specs were described as less rigid than the big SaaS names, which management reads as slightly better for the ODM. 2026 capacity is still aimed at a double; 2028 Taiwan capacity is hoped to double again once Huaya is in.
Q: Which products go on consignment, and what does a higher mix do to next year's revenue?
A: Both ASIC and GPU programmes, on selected expensive parts. A higher consignment mix next year would take some reported revenue out; management said next-generation ASP is high enough that the revenue trend still goes up after that offset.
Q: What do rising memory and CPU costs do to the lines and to second-half pulls?
A: On notebooks, brands already pulled into the first half because component prices were rising; second-half units are guided below the first half, with Q3 down more than 20%. On servers, high-ASP parts are exactly why some programmes are moving to consignment, so that cash and the gross-margin percent are not carrying the chip.
Disclaimer
English notes on Quanta Computer's investor conference of 13 August 2026, covering Q2 2026. 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.