Daily AI and semiconductor supply chain intelligence

Zero One Investment Research Daily Intelligence Brief, November 12, 2026

NVIDIA's board weighs Huang's future as Foxconn's custom-chip servers outgrow GPUs

Top Market Signals

NVIDIA's board meets on removing Huang

NVIDIA's (NVDA US) board met on the evening of 11 Nov US time to discuss whether to remove chief executive Jensen Huang, Reuters reported via CNA, and Bloomberg said the board is working with an outside firm to identify replacements. Under the company's bylaws, a majority of the 11 directors, so six votes, can remove him with or without cause. Ten of the 11 are independent. The board had not announced a decision, Liberty Times reported on 12 Nov.

The board publicly backed Huang and his team on 8 Nov US time, three days before the meeting. Since then The New York Times has reported a second relationship, with Kimberly Powell, the vice president who runs NVIDIA's healthcare business. Semafor reported that some directors favor CFO Colette Kress as interim chief executive while a permanent successor is found, TechNews said. Huang, 63, co-founded the company and owns about 3.6% of it.

Taiwan's suppliers have not seen orders move. Foxconn (2317 TT), NVIDIA's largest server manufacturing partner, said at its 12 Nov investor conference that it has seen no clear change in AI customer demand, and that customers already committed to data center builds may speed them up. Our 9 Nov Blade reported that House committee chairs asked Huang to a closed-door briefing on 18 Nov, the day after NVIDIA's 3Q results. Who sits in the CEO chair at that briefing, and at the results call, is the next fact to watch.

Linked stocks: NVDA US, 2317 TT, 2330 TT

Foxconn's custom chip servers outgrow GPU servers

Foxconn (2317 TT) said shipments of AI servers built on custom chips (ASICs, application-specific integrated circuits designed for one cloud customer) rose 150% YoY in 3Q, against about 70% growth for GPU servers. Chairman Young Liu expects both to grow in 2027, with ASIC servers growing faster, and expects AI rack shipments to rise at a high double-digit rate next year.

Liu said a custom AI chip now takes about nine months from specification to finished part, and inference and AI agents are lifting demand for them. Many ASIC projects run on consignment, where the customer supplies the chips, so their cost never enters Foxconn's revenue. ASIC servers will therefore count for less of revenue than of shipments, but they tie up less working capital. That matters: Foxconn's free cash flow was negative NT$287.4bn over the first three quarters, NT$125.1bn worse than a year earlier, as capex rose 43% to NT$161.7bn.

Cloud and networking products passed 50% of revenue this year, and Liu expects more than 60% within three to five years. 3Q net profit rose 34% YoY to a record, on revenue up 47%. Liu warned that chips, memory and some key components are short, and that 2027 demand should still exceed supply even as new capacity opens, so the limit on how fast racks ship is parts supply.

Linked stocks: 2317 TT, NVDA US, 3661 TT

US rule would close the overseas rental route for Chinese AI firms

The US Commerce Department is finalizing a rule to stop Chinese AI companies from using advanced NVIDIA (NVDA US) chips remotely through data centers outside China, The Information reported, as carried by Investing.com and TechNews. The rule would extend controls to 7 countries, including Malaysia, Singapore and Thailand, and could be published as early as this month. Tencent, Alibaba and ByteDance already have agreements with overseas data centers to train models on NVIDIA chips.

Chinese buyers are pre-booking overseas capacity ahead of the rule. A large Chinese cloud provider has discussed booking overseas NVIDIA capacity in advance, other companies have approached data center operators to reserve it, and smaller AI labs are talking to large Southeast Asian compute rental firms, Reuters reported via TechNews. They fear the Huang affair adds momentum to the Remote Access Security Act, which passed the House 369 to 22 in January and would let Commerce regulate cloud access to restricted chips. It awaits a Senate vote.

The rule is still in inter-agency talks, and Axios reported that Commerce Secretary Howard Lutnick had repeatedly held back earlier crackdown measures. For Taiwan server makers the exposure runs through their neocloud customers: ASUS (2357 TT) said 40-50% of its neocloud customers by region are in Asia Pacific, and declined to say whether the rule would affect its server business. The country list and start date in the final text decide how much Southeast Asian capacity loses Chinese renters.

Linked stocks: NVDA US, 2357 TT, BABA US, 0700 HK

ASUS guides 4Q server revenue up 280%

ASUS (2357 TT) expects 4Q server revenue to rise 100% QoQ and 280% YoY, it said at its 12 Nov investor conference, and kept its target of more than 150% server growth for 2026. For 2027 it targets high double-digit server growth and is working toward triple digits. Part of 3Q server revenue slipped into 4Q because of customer deployment timing, and 3Q server revenue still doubled YoY.

Co-CEO Samson Hu named two headwinds. Funding for AI server buyers is getting tighter, so weaker second and third tier cloud providers and neoclouds may merge or exit, and local rules on data center construction are getting stricter after the US midterms. He said both push the industry toward buyers with strong finances and secured power. ASUS sells to exactly that mix of smaller clouds, neoclouds, AI labs, enterprises and sovereign AI projects, and its general CPU servers grew 80% YoY in 3Q alongside AI servers.

The PC side is going the other way. 3Q PC revenue grew more than 25% even as component prices rose, but ASUS expects 4Q PC revenue to fall 15% QoQ (up 15% YoY) and components to fall 15% YoY, with several quarters of headwind into 1H27. 3Q revenue rose 36% YoY to NT$258.79bn and net profit rose 105%. If funding stays tight, ASUS's 2027 server target depends on which of its neocloud customers keep paying.

Linked stocks: 2357 TT, NVDA US, CRWV US

TSMC brings 3nm to Japan as power and substrates join the shortage

TSMC (2330 TT) will put 3nm production into the second phase of its Kumamoto site in Japan, senior vice president and deputy co-COO Kevin Zhang said at a Nikkei forum in Tokyo on 11 Nov. The first Kumamoto fab is fully running. Japan is only about 4% of TSMC's revenue today, and TSMC also plans a joint venture with Sony (6758 JT) to make image sensors for physical AI, meaning robots and machines that act in the real world.

Zhang said demand for TSMC's advanced processes is "endless" (translated from Chinese) and has grown faster than exponentially over the past nine months as AI agents spread, with customers moving from 5nm to 2nm. He also said conversations with customers show that power supply and substrates are now limits on AI data centers, not only logic chips, so TSMC has to plan its own capacity around the whole supply chain.

That matches what Taiwan chip designers reported this week. Our 11 Nov Blade reported Alchip lost about US$100m of 3Q revenue to shortages of wafers, substrates and packaging. When TSMC starts naming other suppliers' capacity in its own planning, more wafers alone stop being enough to raise AI chip output.

Linked stocks: 2330 TT, TSM US, 6758 JT, 3661 TT

Infineon expects data center power to reach a fifth of sales

Infineon (IFX GR) expects revenue in fiscal 2027, the year to September, to grow more than 20% to about EUR20bn, with power supply revenue for data centers near EUR4bn, about a fifth of the total.

The September quarter already showed the shift. Revenue rose 19.4% YoY and 13% QoQ to EUR4.71bn, and segment margin rose to 23.8% from 18.2% a year earlier. Full fiscal 2026 revenue grew 12% to EUR16.35bn. CEO Jochen Hanebeck named data center power, energy infrastructure, electric vehicles and software-defined cars as the drivers. The shares rose 6.26% on 11 Nov.

TSMC's Kevin Zhang said the same day that power supply is now a limit on AI data centers alongside chips. Infineon sells the chips that convert and deliver that power inside each rack, so its data center line grows with every rack built, whichever GPU or custom chip sits inside. The check is whether the EUR4bn data center figure holds through the fiscal year as AI buyers face tighter funding.

Linked stocks: IFX GR, 2308 TT, NVDA US

先聲 First Word: Exclusives from Chinese-Language Sources

Taiwan toolmakers win over 20% of TSMC packaging buys

Taiwanese makers now supply more than 20% of the equipment TSMC buys for advanced packaging, TSMC senior vice president Y.P. Hou said, against under 1% of its front-end wafer tools, according to Commercial Times. Grand Plastic Technology, Scientech, All Ring, Gallant Micro Machining and C Sun sell wet cleaning, dispensing, bonding and baking tools; Hou said their weak spot is software and data links between machines. (11 Nov 2026) Source: 侯永清：台積先進封裝 採購比重逾兩成

Winbond's chairman doubts TSMC will make 3D DRAM

Winbond (2344 TT) chairman Arthur Chiao said he thinks it unlikely that TSMC will enter 3D DRAM for NVIDIA, because it would compete with its own partners Samsung, SK hynix and Micron, which all have their own memory technology. He added that DRAM know-how takes years to build and TSMC usually stays quiet until a technology is ready. (12 Nov 2026) Source: 台積電進軍3D DRAM傳聞滿天飛！華邦電焦佑鈞為何不信？

Samsung faces a shortfall in EUV mask blanks for 2027

Samsung Electronics (005930 KS) reportedly faces a shortfall in its 2027 purchases of EUV blank masks, the coated plates that chip patterns are written onto, because its use should rise about 25% a year for three years and main supplier Hoya cannot expand fast enough. Mask blanks join substrates and power on the list of parts that can slow advanced chip output. (12 Nov 2026) Source: 傳三星EUV空白光罩採購拉警報 日廠Hoya擴產仍難解供應壓力