Key takeaways
- Tesla, Nvidia reportedly put ₩3.7T into SEMCO substrate lines
- Micron's Taiwan union rejects offer after 24-hour strike
- TSMC explores rare-earth-free motors
- Jensen Huang dismisses 2029 memory glut fears
- 32 export-restricted B300 GPUs spotted on China Telecom-linked network
- Samsung's Taylor fab targets 20K 2nm wafers a month by 2H27
- Some SEMCO MLCC lead times stretch to 61 weeks
- Musk confirms Tesla-SpaceX merger talks
Watch the video briefing (in Korean)
More briefing videos on our YouTube channel.
1. Tesla and Nvidia reportedly to invest KRW3.7 trillion in SEMCO FC-BGA lines
Citing the Maeil Business Newspaper, Reuters reported that Tesla and Nvidia will together invest KRW3.7 trillion (about US$2.8 billion) to build FC-BGA substrate lines at Samsung Electro-Mechanics facilities including its Sejong plant. Nvidia's share is reported at KRW2.6 trillion and Tesla's at KRW1.1 trillion, with output supplied under long-term contracts.
The substrates are said to be destined for Tesla's self-driving and robot processors and Nvidia's AI chips. The structure and terms come from industry sources and have not been confirmed in a formal company filing.
Why it matters: Customers paying for a component maker's capacity is a sign of how scarce high-end AI substrates have become. Without them, finished chips cannot be packaged into usable products.
- FC-BGA: A high-performance package substrate onto which a chip is mounted face-down, used for large processors such as CPUs and GPUs.
2. Micron Taoyuan union ends 24-hour strike, rejects offer ahead of Oct 27 talks
The 24-hour warning strike by the union at Micron's Taoyuan site in Taiwan ended on the morning of October 23, closing out what was described as the first strike in Taiwan's semiconductor industry.
In a union vote, 1,633 members, roughly 70% of the membership, opposed the company's proposal. The union wants 15% of operating profit paid out to employees as quarterly bonuses. Micron has said it will introduce quarterly profit-linked bonuses for staff worldwide from fiscal 2027 but has not specified the share. Talks set for October 23 were pushed to October 27 by the typhoon.
Why it matters: Labor tension over profit sharing is rising at a company whose earnings have surged on AI memory demand. In fabs that run around the clock, labor stability translates directly into output risk.
3. TSMC explores rare-earth-free motors for factory automation equipment
TSMC has surveyed its equipment suppliers about using motors without rare earths in factory automation tools, Reuters reported, citing three industry executives briefed on the plans. The aim is to prepare for the possibility that China cuts off rare-earth supplies.
The discussions are at an early stage and have not led to a broad rollout of such equipment. Which tools and suppliers are involved, and on what timeline, has not been decided.
Why it matters: Supply-chain risk is now reaching small components inside fab equipment, not just materials and tools. Qualifying new parts for chip fabs takes a long time, which explains the early move.
- Rare-earth motor: A motor that uses strong permanent magnets made with elements such as neodymium, prized for being compact and powerful.
4. Jensen Huang sees no 2029 memory glut, open-minded on CXMT
Speaking to CNBC during Nvidia's GTC Berlin event on October 22, CEO Jensen Huang pushed back on the view that memory supply will outstrip demand by 2029. He argued AI workloads keep consuming more memory and said HBM will remain in use for years.
Asked about working with or buying from Chinese memory maker CXMT, he said Nvidia was open-minded. That remark does not amount to a supply agreement.
Why it matters: Memory stocks had recently been hit by oversupply worries, so a demand view from the largest AI chip buyer drew close attention.
- HBM: High-bandwidth memory: stacked DRAM dies with a very wide data path, placed next to AI accelerators.
5. 32 Nvidia B300 GPUs found on China Telecom-linked network
Israeli security firm Lava said it identified the model names and unique IDs of 32 Nvidia Blackwell Ultra B300 GPUs on internet-exposed monitoring endpoints tied to China Telecom IP addresses. The B300 cannot legally be sold into China under US export rules.
An IP association alone does not establish where the chips physically sit, who owns them, or whether they were smuggled. Nothing in the findings indicates Nvidia sold them into China.
Why it matters: Enforcement of AI chip export controls is shifting from the point of sale to tracing end users and where hardware actually runs.
6. Samsung places final tool orders for Taylor Fab 1, targets 20K 2nm wafers by 2H27
Samsung Electronics has ordered equipment for roughly 6,000 more 2nm wafers per month at its first fab in Taylor, Texas, according to ZDNet Korea. Filling the remaining space would bring the fab close to its design maximum of 20,000 wafers per month, with the target set for the second half of 2027.
The tools are expected to start arriving at the end of this year. Samsung has not confirmed the orders.
Why it matters: The move would largely complete core tool spending at Samsung's first leading-edge US foundry. Because expensive tools usually follow committed customer volume, it is also read as a demand signal.
- WPM: Wafers per month, the standard measure of a fab's production capacity.
7. Nvidia wants HBM base dies on TSMC 3nm as TSMC adds 3nm and A14 capacity
Citing the Seoul Economic Daily, TrendForce reported that Nvidia has asked Samsung, SK hynix and Micron to build the base die of its custom NVHBM memory on TSMC's 3nm process, and that Samsung has agreed. Samsung will keep its in-house process for standard HBM4 and HBM4E.
TSMC is adding another 3nm fab in southern Taiwan, and its next-generation A14 process is targeted for volume production in 2028.
Why it matters: As foundries take over the logic layer inside memory stacks, the line between foundry and memory businesses is blurring.
- Base die: The logic chip at the bottom of an HBM stack that manages signals between the DRAM layers and the processor.
8. Lead times for some SEMCO MLCCs stretch to 61 weeks
The manufacturer's standard lead time listed at distributor DigiKey for some Samsung Electro-Mechanics MLCCs rose from 39 to 49 weeks last month to 61 weeks this month, meaning new orders would arrive around the end of next year. These are distributor listings, and lead times on volumes supplied directly to device makers may differ.
TrendForce data put MLCC lead times in the AI server supply chain at 35 weeks in mid-October, about three times normal. SEMCO has signed a string of long-term supply deals covering 2027 this year.
Why it matters: With thousands of MLCCs on each AI server board, even tiny passive parts that steady power delivery are becoming a bottleneck.
- MLCC: Multilayer ceramic capacitor: a tiny part that briefly stores and releases charge to stabilize voltage, used by the hundreds or thousands in electronics.
9. Nvidia shows Optics-on-Interposer CPO prototype
At SEMICON West 2026, Nvidia presented an Optics-on-Interposer (OOI) prototype that places optical engines directly on the same silicon interposer as the GPU or ASIC. Conventional CPO sits the optics on the organic substrate; removing the substrate from the electrical path is meant to raise bandwidth density and cut power.
A companion DWDM demo combined eight wavelengths for 256Gb/s per fiber. Both remain research projects with no commercialization timeline.
Why it matters: Moving data between chips is becoming the main performance and power limit in AI data centers, driving a race to put optical links ever closer to the silicon.
- CPO: Co-packaged optics: placing the engines that convert electrical signals to light inside the same package as the chip.
- Silicon interposer: A thin silicon layer on which several chips sit side by side, linked by very dense wiring.
10. FCC blocks two HiSilicon-powered devices in first component-rule case
The US Federal Communications Commission said on October 22 that it had blocked authorization of a dashboard camera and a wireless video transmitter containing surveillance chips from HiSilicon, Huawei's chip-design arm.
It is the first action under a rule in force since October 13 that extends restrictions to finished devices containing components from companies on the FCC's Covered List. The agency did not name the manufacturers or models.
Why it matters: Restrictions now reach individual chips inside a product, not just companies or finished devices, putting the burden on device makers to prove where their parts come from.
11. TSMC weighs Longtan Phase 3 for panel-level CoPoS packaging
Drawing on Taiwanese media, TrendForce reported that the long-delayed Phase 3 expansion of the Longtan Science Park has resumed and could host two 1.4nm fabs plus a panel-level packaging facility. Final plans remain under consideration.
CoPoS packages chips on 310x310mm square panels instead of round wafers. Reports point to pilot lines first and volume production around the end of 2028. TSMC has not announced the plan officially.
Why it matters: As AI packages grow, round wafers waste more area at the edges. Square panels fit more packages per run, making them a candidate fix for the advanced-packaging crunch.
- CoPoS: Chip-on-Panel-on-Substrate: TSMC's next-generation packaging that moves its CoWoS approach from round wafers to large square panels.
12. TI sees $500 to $1,000 of its chips per humanoid robot
Texas Instruments CEO Haviv Ilan said on the company's third-quarter call on October 21 that a humanoid robot could carry $500 to $1,000 worth of TI semiconductors. The figure covers TI's own content, not the robot's total bill of materials.
He expects robotics activity to rise sharply in 2027 but gave no estimate of humanoid unit volumes and named no customers.
Why it matters: Every joint in a humanoid needs motors, sensors and power parts, making these robots a new market for analog and power chips.
- Analog chip: A chip that handles continuous signals such as voltage, current or temperature, widely used for power management and motor control.
13. AI server PCBs move from afterthought to system bottleneck
In a DIGITIMES interview, AI server motherboards were described as the largest, thickest and most structurally complex PCBs ever produced, pulling substrate and board makers into chip design decisions. Chip design, packaging and circuit boards, once separate disciplines, are converging.
At the IMPACT 2026 conference in Taipei, package-to-board integration was also flagged as an emerging technical bottleneck requiring closer coordination among packaging, substrate, PCB and assembly firms.
Why it matters: As AI chips get more powerful, the boards that carry their power and signals increasingly set the performance ceiling for the whole system.
- PCB: Printed circuit board: the board that holds chips and components and wires them together.
14. Apple and Google push PCB suppliers toward net zero
Pressure from Apple and Google for net-zero supply chains leaves PCB makers juggling affordable renewable power, qualification of low-carbon materials, and emissions measurement across their own suppliers. Apple requires direct suppliers to use 100% renewable electricity for Apple production before 2030.
An analysis cited by DIGITIMES found purchased electricity accounts for nearly 60% of the PCB industry's emissions, while low-carbon alternative materials do not yet meet every performance requirement of advanced boards.
Why it matters: Fast growth in AI server boards means more power use and emissions, putting expansion at odds with carbon-reduction targets.
- Scope 2: Indirect greenhouse gas emissions from generating the electricity and heat a company buys.
15. Hermes Testing approves over NT$7 billion for AI test expansion
Taiwanese wafer-test supplier Hermes Testing Solutions approved budgets on October 22 for a new R&D center in Hsinchu and a manufacturing center in south-central Taiwan, with total investment expected to exceed NT$7 billion (about US$215 million).
The company supplies membrane probe cards and is targeting silicon photonics and CPO testing. Its September revenue jumped 150% year on year on AI chip test demand.
Why it matters: Expensive, complex AI chips make it critical to catch defects before packaging, and new technologies such as optical links are widening demand for test hardware.
- Probe card: A fixture with fine needles that touch each chip on a wafer to check it electrically.
16. Rapidus keeps 2H27 2nm target, courts US AI customers
The Rapidus chairman said in Tokyo on October 21 that plans for 2nm mass production in the second half of 2027 remain broadly on track, with initial demand expected mainly from the US, where AI-related demand is strong.
He did not disclose customer names, volumes or signed contracts. Talks with Japanese companies, including computer makers, are also said to be progressing.
Why it matters: Winning real customers remains the central test for Japan's state-backed foundry as it enters a 2nm race led by TSMC and Samsung.
17. Japanese materials firms join Tata's Dholera fab supply chain
Foreign suppliers are gathering around Tata Electronics' Dholera fab, India's first commercial chip plant (about US$11 billion, 300mm, 28nm to 110nm). Sumitomo Chemical signed an MoU with Tata in September covering photoresists and high-purity chemicals and a study of local production.
Fujifilm plans a phased investment of about INR 800 crore in a semiconductor materials plant at Dholera. Equipment maker Lam Research has said it will set up on-site support at Dholera once the fab is running.
Why it matters: Early supply chains in new chipmaking countries are typically filled first by proven global suppliers; whether local partners are developed alongside them is the open question.
- Photoresist: A light-sensitive coating used to transfer circuit patterns onto a wafer.
18. Strong iPhone Duo demand lifts hopes for OLED TDDI
Apple has called demand for its foldable iPhone Duo extraordinary, with pre-order delivery estimates slipping into 2027. Citing DIGITIMES, MacRumors reported that display driver suppliers expect the Duo to give rival foldable makers more confidence, expanding the OLED TDDI market.
Industry sources also see the rumored touchscreen MacBook as a possible boost for laptop display chips, though adding touch does not necessarily mean adopting TDDI.
Why it matters: TDDI is still too costly for many regular flagships but is effectively required in foldables, so wider foldable adoption could shift driver-chip makers toward higher-value products.
- TDDI: Touch and display driver integration: a single chip that handles both touch sensing and driving the display.
19. Typhoon Champi shuts Taiwan's market as chips sell off elsewhere
Taiwan closed its stock market, schools and offices on Friday, October 23, as Typhoon Champi approached. Tech-heavy markets such as South Korea and Japan saw chip stocks weaken ahead of big-tech earnings, but Taiwan did not trade.
The Taiwanese market is also closed on Monday, October 26, for a public holiday, so moves since then will be priced in once trading resumes.
Why it matters: Shares of TSMC and other Taiwanese chipmakers serve as a barometer for the industry; the closure delayed price discovery rather than removing downside risk.
20. Musk confirms Tesla-SpaceX merger talks; combined value near $3.5T
Elon Musk has confirmed that Tesla and SpaceX are discussing a merger. Each company has formed a special committee of independent directors to review it, and no terms have been agreed.
The roughly US$3.5 trillion figure adds SpaceX's private valuation of about $2 trillion to Tesla's market value of nearly $1.5 trillion; it is an estimate, not a deal price. Any merger would need approval from a majority of Tesla shareholders.
Why it matters: Both companies are large chip buyers for self-driving, robotics and AI infrastructure, so a combination could reshape their chip sourcing and in-house design plans.
Sources
- Samsung Electro-Mechanics gets $2.8 billion chip investment from Tesla, Nvidia, report says (Reuters (KFGO))
- Tesla and Nvidia to Invest KRW3.7 Trillion in Samsung Electro-Mechanics Chip Substrate Facilities (MarketScreener)
- 台灣半導體業首次罷工結束 美光桃園工會：為爭取更公平的分潤制度做好抗爭準備 (新頭殼 Newtalk)
- 美光台灣工會拒絕資方分紅方案 勞資預計27日重啟協商 (民視新聞網)
- 美光桃園工會堅持分潤制 協商因颱風延至27日 (聯合新聞網)
- TSMC seeks factory-equipment alternatives (Taipei Times (Reuters))
- TSMC Seeks Equipment Alternatives to Cut China Rare Earth Reliance (Seoul Economic Daily)
- Nvidia's Jensen Huang says he sees no sign of a 2029 memory glut: 'You can always use more' (CNBC)
- Nvidia B300 chips reportedly surface on China Telecom-linked network, raising export-control questions (DIGITIMES)
- Lava detecta 32 GPU NVIDIA B300 ligadas a direcciones IP de China Telecom (Revista Cloud)
- Samsung reportedly orders final tools for first Taylor fab as Texas site nears 2nm output cap (DIGITIMES)
- Samsung Orders Equipment For 6,000 More 2nm Wafers A Month At Its Taylor, Texas Fab (Wccftech)
- NVIDIA Reportedly Pushes TSMC 3nm NVHBM Base Dies for Samsung, SK hynix, Micron (TrendForce)
- TSMC Unveils Next-Generation A14 Process at North America Technology Symposium (TSMC)
- 삼성전기 MLCC '납기 61주'…가격 주도권도 쥔다 (뉴스핌)
- AI 서버 '보드당 1.2만개' MLCC 블랙홀…내년 초 MLCC 공급난 정점 (시티타임스 (Investing.com))
- Samsung Electro-Mechanics secures 2027 AI server MLCC demand as supply pressure builds (DIGITIMES)
- NVIDIA Unveils Next-Gen CPO Prototype With Optics on Silicon Interposer, Targets Higher Bandwidth Density (TrendForce)
- Nvidia unveils OOI CPO prototype for post-2028 AI needs (DIGITIMES)
- FCC blocks two devices powered by Huawei silicon subsidiary (Light Reading)
- FCC Catches Huawei Surveillance Chips in Its First Parts Ban Case (DroneXL)
- TSMC Reportedly Eyes Two AUO Fabs Worth TWD 30B+ for FOPLP and CoPoS; Longtan Could Add 1.4nm (TrendForce)
- Panel-Level Packaging at TSMC: First CoPoS Pilot Line in 2026 (All About Industries)
- Texas Instruments sees up to $1,000 in chip content per humanoid robot as analog supply tightens (DIGITIMES)
- Texas Instruments Incorporated Q3 2026 Earnings Call Summary (Yahoo Finance)
- Interview: From afterthought to bottleneck, circuit boards move to the heart of AI hardware (DIGITIMES)
- Packaging-to-board integration emerges as key gap as AI reshapes electronics supply chain (DIGITIMES)
- Apple, Google push PCB supply chain toward net zero amid challenges (DIGITIMES)
- Apple Suppliers Face Hurdles Cutting Circuit Board Emissions (MacRumors)
- Taiwan wafer-testing supplier Hermes Testing Solutions commits more than US$215 million to advanced chip testing expansion (DIGITIMES)
- HTSI September revenue surges 150% as AI chip testing demand accelerates (DIGITIMES)
- Hermes Testing ships membrane probe cards, targets SiPh and CPO (DIGITIMES)
- Rapidus courts US AI customers ahead of planned 2H27 2nm mass production (DIGITIMES)
- Tata Electronics and Sumitomo Chemical Sign MoU for High-Quality Semiconductor Materials (Tata Electronics)
- Fujifilm semiconductor materials plant in Dholera (FUJIFILM)
- Lam Research to offer onsite support for Tata Electronics' chip fab in Guj (Business Standard)
- iPhone Duo Demand Reportedly Exceeds Apple's Expectations (MacRumors)
- Apple Calls iPhone Duo Demand 'Extraordinary' After Orders Slip to 2027 (MacRumors)
- Taiwan sits out global chip selloff, thanks to typhoon (DIGITIMES)
- Taiwan's US$5 tril stock market to close on Friday as Typhoon Champi nears (The Edge Malaysia)
- Tesla and SpaceX discuss potential $3.5T merger, no formal deal yet (Crypto Briefing)
- SpaceX Could Buy Tesla for $480 Per Share in $3.74 Trillion Merger, RBC Says (Yahoo Finance)
Compiled by Bits & Bites from public reporting. Not investment advice.
Leave a Reply