Bits & Bites

한국어

Author: Bits & Bites

  • [Oct 22] Semiconductor News Briefing — AI servers squeeze PCB material supply

    Key takeaways

    • Taiwanese laminate maker ITEQ plans to lift monthly CCL capacity to 9.15 million sheets by 2029, while T-glass fiber stays tight until Nittobo's expansion arrives.
    • TPCA expects Taiwan's PCB output to reach NT$1.1365 trillion this year, topping NT$1 trillion for the first time.
    • CXMT plans server DDR5 modules built on 4F² DRAM by year-end, and Sony and TSMC are forming an image sensor joint venture in Kumamoto.

    1. ITEQ to lift CCL capacity to 9.15 million sheets a month by 2029 as T-glass stays tight

    Chen Chin-tsai, chairman of Taiwanese copper-clad laminate (CCL) maker ITEQ, said the group plans to raise monthly CCL capacity from 5.7 million sheets at the end of 2026 to 6.3 million in 2027, 8.25 million in 2028 and 9.15 million in 2029. That is roughly a 60% increase over three years, driven mainly by expansions in Thailand and Wuxi, China.

    Meanwhile, T-glass, a specialty glass fiber used in high-performance substrates, is effectively dominated by Japan's Nittobo. Nittobo is doubling its capacity, but new supply is not expected to arrive in volume until 2027 or later.

    Why it matters: AI server boards have many layers and need low signal loss, so they consume large amounts of premium materials. Bottlenecks are showing up first in these upstream materials rather than in finished boards.

    • CCL: Copper-clad laminate: an insulating sheet of glass fiber and resin bonded to copper foil, the base material of circuit boards.
    • T-glass: A specialty glass fiber with very low thermal expansion, used to keep advanced chip substrates from warping.

    2. Taiwan's PCB output set to top NT$1 trillion for the first time

    The Taiwan Printed Circuit Association (TPCA) expects Taiwanese PCB makers' combined domestic and overseas output to reach NT$1.1365 trillion in 2026, up 24.2% from last year and above NT$1 trillion for the first time. The figure is an upgrade from the forecast it issued in June.

    Second-half output is projected at NT$611.8 billion, up 24.5%. Growth is being led by higher-value products such as high-layer-count AI server boards, IC substrates and HDI boards.

    Why it matters: Taiwan sits at the center of global PCB supply, so the numbers show how widely AI server spending is spreading through the component industry.

    • HDI: High-density interconnect: boards that pack many circuits into a small area using very fine lines and micro-holes.

    3. CXMT aims to unveil 4F² DRAM-based server DDR5 by year-end

    Cao Kanyu, president of Chinese memory maker CXMT, said at a Shanghai event on October 10 that the company plans to launch a server DDR5 RDIMM using a 4F² DRAM structure before the end of the year. The design reportedly also uses hybrid bonding, building the memory cells and the peripheral circuits separately and then joining them.

    At the same process dimension, a 4F² layout can shrink cell area by about a third versus today's 6F². Speed, capacity and yield figures were not disclosed, and the industry views the year-end target as quite aggressive.

    Why it matters: Samsung expects to mass-produce 4F² products around 2028, so CXMT would be first to commercialize the new DRAM structure if it hits its schedule. Its actual high-volume capability remains to be seen.

    • 4F²: A measure of the area each DRAM cell occupies; a smaller number means more cells fit in the same space.
    • Hybrid bonding: A technique that joins two wafers or chips by bonding copper pads directly, without solder bumps, for very dense connections.

    4. Sony and TSMC form image sensor joint venture in Kumamoto

    TSMC and Sony Semiconductor Solutions have signed a definitive agreement to set up a joint venture, Advanced Vision Semiconductor Manufacturing, in Koshi, Kumamoto Prefecture. The venture is scheduled to begin commercial production in 2029, with Sony as the majority shareholder in charge of management.

    Reports put the total investment at ¥747 billion, with Sony contributing about ¥465 billion and TSMC about ¥282 billion. The venture will use Sony's recently completed Koshi plant, and spending will be phased in according to market conditions.

    Why it matters: Sony keeps control of design and management while sharing part of manufacturing with TSMC, easing the burden of heavy capital spending.

    • Image sensor: A chip that converts light entering a camera into electrical signals to create photos and video.

    5. ESMT revenue surges as specialty memory prices keep climbing

    Based on its monthly sales, Taiwanese specialty memory maker ESMT's third-quarter revenue totaled about NT$22.15 billion, up 58.8% from the second quarter and more than six times the year-earlier level. September revenue alone was up 470% year on year.

    The company said shortages of specialty memory such as DDR2 and DDR3 are unlikely to ease this year and that price increases are continuing. Shipment growth has been limited, so the gains are coming mainly from higher prices.

    Why it matters: As major memory makers concentrate capacity on HBM and the latest DRAM, supply of older and specialty memory is shrinking and prices are rising sharply.

    • Specialty memory: Older-standard or low-capacity memory used in appliances, networking gear and industrial equipment.

    6. Micron's Taoyuan union in Taiwan wins the right to strike

    The union at Micron's Taoyuan plant in Taiwan announced on October 7 that a member vote had given it the legal right to strike. Of 2,258 members, 2,012 voted and 1,994 were in favor. The vote followed two failed mediation rounds in September, and the union is seeking a permanent profit-sharing scheme that would set aside 15% of profit in place of the current bonus system.

    No strike date has been set. The union at Micron's Taichung plant is pursuing its own mediation with the company, and Micron said it will keep talking with its unions.

    Why it matters: Micron employs about 15,000 people in Taiwan, and the two unions represent more than 80% of them. With memory makers posting record profits, labor tensions have become a new variable for Taiwan's chip supply chain.

    7. Zhen Ding shows 34-layer AI server board and 1.6T optical module substrates

    Zhen Ding, Taiwan's largest PCB maker, showed boards for 800G and 1.6T optical transceiver modules and XPO/NPO applications at SEMICON Taiwan 2026, along with high-layer AI server boards of up to 34 layers.

    The company is pursuing a One ZDT strategy of supplying everything from flexible boards and HDI to high-layer server boards and advanced IC substrates. It expects server and optical module revenue to more than double in 2026 and IC substrate revenue to grow by more than 70%.

    Why it matters: As GPU interconnect speeds rise in AI data centers, demand is growing both for higher-layer server boards and for optical communication parts.

    • 1.6T optical module: A data center component that sends and receives 1.6 terabits per second as light signals.

    8. Taller AI server boards drive demand for backdrilling equipment

    High-end AI server boards have grown from 12-20 layers to 28-46 layers, and some now need more than 6,000 backdrilled holes per board. Because the boards are harder and thicker, the number of holes a single drill bit can make has fallen from about 3,000 to around 600.

    Taiwanese toolmaker Tongtai said AI server PCB demand has lifted orders for its drilling machines, with about 70% of its electronics equipment orders coming from the AI server supply chain. Ta Liang's sales of high-precision backdrilling equipment are also rising sharply year after year.

    Why it matters: At higher signal speeds, leftover copper stubs create noise, so precisely removing them has become critical to board quality.

    • Backdrilling: Drilling out the unused copper portion of a board's vertical connections from the back side to reduce high-speed signal distortion.

    9. Daeduck Electronics joins next-gen packaging race with ceramic core substrate

    Daeduck Electronics showed a 100×100mm ceramic core substrate at KPCA Show 2026 in September. By the company's figures, the ceramic core's stiffness (Young's modulus) is 251GPa versus 64-73GPa for glass cores, and its thermal conductivity of 24.5W/mK far exceeds glass at 0.9-1.2W/mK.

    Actual heat dissipation in a package also depends on the chip and substrate design and the bonding method, so material figures alone do not settle performance. Ceramic core substrates are still in development, and Japan's Kyocera is working in the same area.

    Why it matters: As AI chip packages grow, keeping substrates from warping and drawing heat away has become critical, and ceramic is emerging as an alternative to glass.

    • Core substrate: The rigid center layer of a chip substrate that serves as its backbone, with wiring layers built above and below.

    Sources

    Compiled by Bits & Bites from public reporting. Not investment advice.

  • [Oct 19] Semiconductor News Briefing — TSMC posts record quarter as Terafab talks surface

    Key takeaways

    • TSMC reported third-quarter revenue of $46.72 billion with a 66.4% gross margin, and said it must operate any Terafab fab it joins with Musk.
    • GlobalFoundries signed a roughly $2 billion deal to make CoWoS silicon interposers for TSMC at its Malta, New York fab.
    • Japan approved converting TSMC's second Kumamoto fab to 3nm, as Taiwan Mask and other suppliers expand around the site.

    1. TSMC posts $46.7B third-quarter revenue with a 66.4% gross margin

    TSMC reported third-quarter results on October 15. Revenue came in at $46.72 billion, up 41.2% from a year earlier and 16.2% from the second quarter. Gross margin landed at 66.4%, inside the company's 65-67% guidance range but slightly below the 67.7% it reached in the prior quarter.

    Chips made on 7nm and more advanced processes accounted for 79% of wafer revenue, up from 77% in the second quarter. TSMC guided for a gross margin of 65-67% again in the coming quarter.

    Why it matters: As the world's largest contract chipmaker, TSMC is one of the clearest real-world gauges of how strong AI chip demand is at the production stage. The steady rise in its advanced-node share stands out.

    • Gross margin: The share of revenue left after subtracting the direct cost of making the products.

    2. TSMC confirms Terafab talks with Musk, but says it must run any fab it joins

    TSMC Chairman C.C. Wei confirmed during the earnings call that the company has been talking with Elon Musk's side about Terafab, a chip plant project in Texas. He made clear that any fab TSMC takes part in would have to be operated by TSMC, and said nothing has been agreed.

    Musk had earlier said SpaceX intends to build and run the plant itself, with TSMC at most renting some of the space. Terafab is planned for Grimes County, Texas, with roughly $16.8 billion in initial spending, and construction began in August.

    Why it matters: TSMC has used a partner-invests, TSMC-operates model at its joint fabs in Japan and Germany. If the two sides cannot bridge the gap over who runs the plant, any collaboration could be delayed.

    3. GlobalFoundries signs $2B deal to make interposers for TSMC

    GlobalFoundries (GF) has signed a five-year agreement with TSMC worth about $2 billion. GF will expand capacity at its Malta, New York fab to produce silicon interposers for TSMC's CoWoS packaging, making the site the first interposer production base in the US.

    Volume production is slated to begin in the first half of 2028. GF will not be manufacturing TSMC's leading-edge logic chips; it will supply a packaging component. Volume and pricing details were not disclosed.

    Why it matters: Part of the CoWoS supply chain, long seen as a bottleneck for AI accelerator output, is moving to the US. It is also an example of a mature-node fab taking on a new role supplying parts for advanced packaging.

    • CoWoS: TSMC's advanced packaging technology that places compute chips such as GPUs side by side with HBM memory on a shared base so they work as one unit.
    • Silicon interposer: A thin slice of silicon with extremely dense wiring and tiny through-silicon vias that connects multiple chips to each other.

    4. Japan approves switching TSMC's second Kumamoto fab to 3nm

    Japan's Ministry of Economy, Trade and Industry (METI) approved on October 16 a plan to convert TSMC's second Kumamoto fab to 3nm logic production. Total investment is estimated at about $10.8 billion, and the subsidy cap was set at ¥645.15 billion, lower than the ¥732 billion attached to the original 2024 plan.

    The approval covers the first phase, with initial capacity of around 6,000 wafers per month. Operations are targeted for December 2028, and further expansion will proceed in stages depending on additional subsidies.

    Why it matters: TSMC's Japanese fabs have so far focused on mature processes. A 3nm line would give Japan its first leading-edge logic production site.

    5. Samsung Electro-Mechanics commits ₩6.78 trillion to AI server FC-BGA substrates

    Samsung Electro-Mechanics disclosed on September 28 that it will invest a total of ₩6.78 trillion (about $5 billion) to expand FC-BGA substrate capacity for AI servers. Of that, ₩4.27 trillion goes to its Sejong site in Korea and roughly ₩2.51 trillion to its Vietnam subsidiary.

    The Sejong investment runs from September 2026 to May 2028, with mass production targeted for September 2028. The Vietnam plant is scheduled to start operating by June 2030.

    Why it matters: AI accelerators and GPUs are large chips with many connections, so they need high-performance substrates. Substrate supply is emerging as a factor that sets the pace of AI server production.

    • FC-BGA: A high-performance chip substrate where the die is flipped and bonded directly, linking CPUs and GPUs to the main board.

    6. Samsung aims to break ground on second Taylor fab this year, anchored by Tesla AI6

    Samsung Electronics plans to start construction on its second fab in Taylor, Texas, by the end of 2026, targeting mass production in 2030. According to industry reports, it has asked key equipment partners to begin the certification steps needed to move tools in.

    The Taylor site has been designated to build Tesla's next-generation AI6 chip under a $16.5 billion contract that runs through the end of 2033.

    Why it matters: With a large customer's volume already secured, the plan to expand 2nm capacity in the US is a useful measure of how big Samsung's American foundry business could become.

    7. Chip test interface suppliers ride AI demand to strong results

    Taiwanese test interface makers such as WinWay, MPI and CHPT are posting strong results on demand for AI and high-performance computing chips. WinWay's July revenue reached NT$1.6 billion, a record for the second month running, and the company said it has more than 20 large socket projects with over 10,000 pins. Standard lead times have stretched from about eight weeks to 13-14 weeks.

    MPI's second-quarter revenue topped NT$5.2 billion, and CHPT set a monthly revenue record in July.

    Why it matters: AI chips are large and densely connected, making pre-production validation and yield testing much harder. That raises the importance of the test stage.

    • Probe card / socket: Parts that press tiny needles or contacts onto a wafer or finished chip to check it electrically for defects.

    8. India's Tata Dholera fab targets commercial output in mid-2028

    Tata Electronics is aiming for mid-2028 to start commercial operations at its first large chip plant in Dholera, Gujarat, later than the original late-2026 target.

    The fab will cover 28nm to 110nm, starting with 55nm and 90nm before adding 28nm. It is designed for 50,000 wafers per month at full build-out, with total investment of about $11 billion. Taiwan's PSMC is the technology partner and ASML is supplying lithography tools.

    Why it matters: As India's first commercial 300mm fab, it is a key test of whether the country can establish itself as a chip manufacturer.

    9. BOE's 24-layer glass substrate enters customer evaluation

    Chinese display maker BOE has produced a 24-layer glass substrate for chips measuring at least 100×100mm and is now in customer validation. According to a September 22 investor briefing, it has completed in-house reliability testing and moved on to evaluations in which CPUs and GPUs are mounted on the substrate and run in servers.

    The substrate is 1mm thick with through-glass vias about 67µm wide, and BOE also showed vias with a 15:1 aspect ratio. The company aims to be ready for a mass-production investment decision as early as the first half of 2027.

    Why it matters: Glass substrates are seen as a next-generation material that could ease the warping that plagues conventional organic substrates as chips get larger. BOE is drawing on its glass-processing experience to join the race.

    • Glass substrate: A chip substrate that uses glass instead of organic resin as its core, offering better flatness and thermal stability.
    • TGV: Through-glass via: a tiny hole running through a glass substrate that carries electrical signals from one side to the other.

    10. LG Innotek targets glass substrate supply around 2028

    LG Innotek CEO Moon Hyuk-soo said he expects the company to begin supplying glass substrates around 2028. It is in long-term supply talks with major global customers but did not name them or give a contract timeline.

    At the KPCA Show in September, LG Innotek unveiled its first FC-BGA for AI accelerators with sides longer than 100mm, targeting mass production in 2027.

    Why it matters: Korean substrate makers are broadening their lineups from high-performance AI substrates to next-generation glass cores, and competition is heating up.

    11. NVIDIA shows a PTFE-based NVLink Switch board at OCP

    At the OCP Global Summit 2026, NVIDIA showed a sample NVLink Switch board made with PTFE material for the first time and introduced seven supply-chain partners: Taiflex and Flexium of Taiwan, along with Victory Giant, Shennan Circuits, Kinwong, Shengyi and TTM.

    It has not been confirmed whether this means formal orders or actual shipments. NVIDIA is still validating whether to use PTFE, existing high-performance materials, or a mix.

    Why it matters: As signals between GPUs inside AI servers get faster, board materials with low signal loss become more important.

    • PTFE: A fluoropolymer best known as Teflon, with very low high-frequency signal loss, used in high-speed circuit boards.
    • NVLink: NVIDIA's dedicated high-speed link for connecting GPUs to each other.

    12. Taiwan Mask acquires Kumamoto equipment service firm TCK

    Taiwanese photomask maker Taiwan Mask's board approved on October 16 the purchase of 100% of T.C.K. Corporation in Kumamoto, Japan, for ¥750 million (about NT$160 million). TCK has provided chip equipment maintenance and services in Kumamoto for 33 years.

    Taiwan Mask plans to use the deal to gain local equipment engineering capability in Japan and support customers from both Taiwan and Japan on site.

    Why it matters: With more chip plants rising around Kumamoto since TSMC's arrival, Taiwanese supply-chain firms are moving in to capture local equipment service demand.

    13. Taiwan plans supply-chain industrial park in Texas

    Taiwan's Ministry of Economic Affairs confirmed on October 16 a plan to develop privately funded industrial parks in the US, starting with Texas. Each park is expected to cost $2-4 billion, with private companies investing and the government providing loan guarantees and coordination with US federal and local authorities.

    The Taiwan Electrical and Electronic Manufacturers' Association (TEEMA) is involved, and the goal is to help smaller suppliers to AI server makers such as TSMC and Foxconn set up in the US. Tenants and a construction schedule have not been decided.

    Why it matters: Beyond big customers building their own US plants, this is an attempt to move the supporting supplier ecosystem as well.

    14. Bolt Graphics seeks $100M to target high-precision computing beyond AI

    US GPU startup Bolt Graphics is seeking to raise around $100 million at a valuation of about $700 million, Reuters reported. Nothing is final, and the amount and valuation could change.

    Its Zeus GPU targets fields that need high-precision (FP64) math, such as scientific computing, chip design and professional rendering, rather than AI acceleration. Test chips were made on TSMC's 12nm process, and commercial products are planned on a TSMC process for launch between late 2027 and 2028.

    Why it matters: With the AI accelerator market crowded, Bolt is aiming at the less contested high-precision HPC segment. Independent performance validation has not yet been published.

    • FP64: 64-bit floating-point math used for highly precise calculations such as weather and physics simulations.

    15. Singapore commits S$800M to semiconductor R&D through 2030

    Singapore will invest S$800 million from 2026 to 2030 through the semiconductor flagship program of its RIE2030 national research plan. Led by A*STAR and the Economic Development Board (EDB), it covers seven areas including advanced packaging, silicon photonics, power semiconductors, GaN RF and chip design.

    The government is shifting support for local chip suppliers from productivity improvements toward higher-value capabilities, and runs programs linking large chipmakers with local SMEs.

    Why it matters: The policy aims not just to attract big fabs but to build the technical capabilities of the local suppliers that support them.

    16. SEMI CEO says chip market could reach $3 trillion by 2035

    SEMI CEO Ajit Manocha said the global semiconductor market could reach $3 trillion by 2035. Speaking at SEMICON India 2026, he urged India to build indispensable capabilities in key parts of the supply chain.

    The figure describes long-term growth potential rather than a firm market forecast.

    Why it matters: Long-range market sizes cited by the industry's main trade body can shape national chip investment and policy.

    17. Holtek targets AI server cooling fans with new MCUs

    Taiwanese MCU maker Holtek said its 12V cooling fan MCU has passed qualification at a major customer and that it supplies three products to a Taiwanese AI server fan maker. Orders this year are small, and the company expects meaningful revenue contribution from 2027.

    It is also developing products for 48V cooling systems.

    Why it matters: As AI servers run hotter, demand is growing for chips that control fans precisely. The shift in server power from 12V to 48V is opening further opportunities.

    • MCU: Microcontroller: a small computer chip that controls devices such as appliances, fans and sensors.

    18. Hanwool Semiconductor wins MLCC ultrasonic inspection order from Delta unit

    Korea's Hanwool Semiconductor has won an order for an MLCC ultrasonic inspection system from a subsidiary of Taiwan's Delta Electronics. The tool combines non-destructive ultrasonic inspection, which finds internal defects without breaking the part, with AI-based automatic defect classification.

    The contract value was not disclosed. The company is also discussing specifications and running tests with other global MLCC makers.

    Why it matters: It marks a Korean entry into an MLCC ultrasonic inspection market long dominated by Japanese suppliers. Demand for inspection is rising as AI servers use more high-reliability components.

    • MLCC: Multilayer ceramic capacitor: a tiny part that briefly stores and releases charge to stabilize power; a single server uses thousands.

    Sources

    Compiled by Bits & Bites from public reporting. Not investment advice.