At a glance. The clearest statement of Foxconn's posture toward the AI rack: it does not want the assembly fee, it wants the contents. Asked how it would offset the margin dilution of rack-level shipments, management listed busbars, CDUs, quick disconnects, manifolds and cold plates — the same parts list an 800 VDC rack requires.
What changed this quarter
GB200 NVL72 was confirmed as the mainstream specification, with production started, small shipments by year-end, volume from 2025Q1, and a target of holding 40%+ share.
AI servers passed 40% of all server revenue, more than doubling year on year over nine months, with 2025 guided above 50%.
Geography was split by product risk. AI servers concentrate in North America — Wisconsin, Texas, Mexico — because of national-security sensitivity, while smartphones expand in India from a single-digit share toward 15–20% within one to two years.
The numbers
| Q3 2024 | Value | Change |
|---|---|---|
| Revenue | NT$1.85tn (~US$59bn) | +20% YoY, +20% QoQ |
| Gross margin | 6.19% | −47bps YoY |
| Operating margin | 2.95% | −4bps |
| Net margin | 2.66% | −14bps |
| Net income | NT$49.3bn (~US$1.56bn) | +14% YoY, +41% QoQ |
| EPS | NT$3.55 (~US$0.11) | from NT$3.11 |
Revenue, gross profit, operating profit and net income were all records for the period. Nine-month revenue reached NT$4.7tn (+10%), with gross profit, operating profit and net income up 8%, 16% and 20%. The operating expense ratio fell below 3.5% on scale.
Gross margin compression is structural, not operational: shipping whole racks raises revenue faster than profit. Management's answer is to own more of the rack.
Segment mix: consumer 45%, cloud and networking 32%, computing 17%, components 6%.
NVIDIA 800V read-through
The architecture is never named, but the component list gives it away. Foxconn's in-house targets are:
- Busbars — the exact item NVIDIA's whitepaper says drops about 200 kg per rack when moving off 415 VAC.
- CDU, UQD (universal quick disconnect), manifold, and cold plate — all four liquid-cooling parts, submitted for customer validation.
- NVLink switches, InfiniBand, and Ethernet networking hardware.
Management also confirmed that AI rack assembly margins run high single digit to double digit, better than branded servers at mid single digit, so the mix is accretive in absolute profit even while diluting the percentage.
On silicon, automotive SiC was completing customer validation with a Hsinchu SiC module plant in trial production. NVIDIA's whitepaper credits SiC and GaN maturity, driven by the EV move from 400 V to 800 V, as a precondition for 800 VDC in data centers.
Segments that moved
Cloud and networking was the fastest grower, with AI server revenue more than doubling over nine months and general servers up over 20%.
Consumer electronics was 45% of sales, with a strong seasonal Q4 expected but a cautious year-on-year view given geopolitics and macro uncertainty.
EV continued expanding: Model B nearing production readiness, the battery cell plant pushed to 2025Q1 after typhoon damage, and CDMS discussions with two Japanese automakers expected to sign within months. The 5% market share target slipped in timing but not in ambition.
Guidance
- Q4 2024: significant sequential and annual growth. Consumer up strongly on seasonality; cloud flat against a high Q3 base; computing down after a strong back-to-school quarter; components up significantly.
- Full year 2024: significant growth maintained, with computing raised to significant growth and consumer trimmed on macro uncertainty.
- 2025: AI servers the main driver with visibility of at least two to three years; cloud and consumer revenue contributions expected to balance out; general servers still growing on the data processing and storage that AI infrastructure requires.
Management Q&A
Q: What is AI server's revenue contribution this year, when does GB200 go to volume, and how should next year be read?
A: Over the first nine months, AI-server revenue more than doubled year on year and was about 45% of total server sales, with a target above 50% next year as rack systems become a larger part of the book. GB200 production has started, with a small shipment targeted by year-end and the ramp from Q1 2025, growing quarter by quarter. That AI-server strength is what is expected to lift the cloud-and-networking segment so that it balances consumer electronics next year.
Q: What is the progress of in-house thermal solutions into GB200, and what is the edge versus thermal specialists?
A: Four liquid-cooling parts are targeted in-house — UQD, cold plate, CDU and manifold — and were shown at Hon Hai Tech Day; whether customers adopt them is still their decision. The claimed advantage is Foxconn's dominant position in rack systems, so it can offer structural parts, compute boards, NVLink switches, liquid-cooling components and the assembled system as one Foxconn solution.
Q: Can vertical integration offset AI-server margin dilution, and what else besides cooling is made in-house?
A: Vertical integration is described as the lever for AI-server profitability versus peers. Beyond liquid cooling, Foxconn has already shown self-made busbars plus networking hardware inside the rack, including NVLink switches, InfiniBand and Ethernet.
Q: How should the 2026 AI-server market be viewed, and will CSPs keep investing if they have not yet earned a return?
A: 2025 is not seen as the problem; 2026 is where the market still lacks confidence, but Foxconn called the cycle early. Its estimate from a year earlier put AI TAM above US$1tn in seven to eight years, CAGR above 40%, with hardware about half; major customers now cite US$500bn to US$1tn. Microsoft's returns on deployed applications such as OpenAI and Copilot are the indicator it will watch.
Q: What growth is expected in general servers next year, and will AI servers cannibalise that budget?
A: Foxconn still holds about 40% of general-server allocations, with good growth this year that has not been crowded out by AI servers, after a recovery that started in Q4 last year. Next year, AI infrastructure still needs data processing and storage, so general-server demand is expected to remain and to keep growing.
Q: How is smartphone demand now, what is the view on next year's iPhone, and what is the India capacity plan?
A: Premium-smartphone demand is described as structurally stable over a three-to-four-year view, with a large refresh typically only when form factor or software/hardware changes in an obvious way; Foxconn has not seen a sharply negative read on this customer. India is still a single-digit share of capacity, aimed at 15–20% within one to two years, set by customer request rather than a Foxconn-only target.
Q: How is revenue from the single largest customer (Apple) changing this year and next?
A: Before this year, that customer was more than 50% of revenue. Next year, as AI servers scale, sales from CSPs, NCPs and other AI accounts rise, so that customer's share falls even if the dollar amount does not — cloud-and-networking was already 32% in Q3, up about 9 points year on year, and is expected to move close to consumer electronics.
Q: What is the geopolitical-risk plan in practice?
A: Risk differs by product: consumer electronics such as phones and PCs have seen limited tariff impact, while national-security items including government PCs and AI servers carry more. AI-server capacity for next year is therefore being built mainly in North America — Wisconsin, Texas and Mexico — and smartphones are being diversified toward India. China server exposure is below 10% and is expected to fall further as GB200, which is not shipped into China, grows.
Q: How should the EV outlook be read after the new US administration?
A: Localised production is still the trend with or without IRA subsidies, and without subsidies cost structure matters more, which management said favours the CDMS model. The US EV leader has underperformed and the cycle has slowed a little, which Foxconn called extra time to prepare; more traditional automakers have come to talk, and some projects may start to land in the next few months. The longer-term EV view is unchanged, with a cited US$4tn market as a long-cycle driver.
Disclaimer
English notes on Hon Hai Precision Industry's investor conference of 15 November 2024, covering Q3 2024. Figures are as presented by management and have not been independently verified. USD equivalents are approximate at about NT$31.5 to the dollar. For reference only; not investment advice.