We cover the companies behind the AI buildout, from chips and memory to the equipment and materials that make them possible. Based in Taiwan, we draw on company interviews, site visits, public filings and industry research to explain what’s changing across these supply chains—and what it means for the businesses involved. A few places to start: • Phison’s move into systems and software. CEO K.S. Pua spoke with us about AI’s memory constraints and why a company known for storage is expanding into a broader business. Read the interview: xitter.cf/statementdog/status/20… • What it takes to expand ABF substrate production. Goldman Sachs reports that Eternal Precision’s equipment delivery slots are booked into 2031. We look at how equipment availability could shape suppliers’ expansion plans. Read more: xitter.cf/statementdog/status/21… • Why faster optical modules could compete with AI chips for capacity. Drawing on DIGITIMES reporting, we examine the substrate requirements behind the shift to 3.2T optical modules. Read more: xitter.cf/statementdog/status/21… We follow these developments beyond the initial announcement: when capacity comes online, whether products pass customer qualification, and how supply affects pricing. Follow @statementdog for company interviews and ongoing analysis of AI hardware and semiconductor supply chains.

Our supply-chain checks indicate that Macronix, a Taiwanese flash-memory maker, is shifting some production capacity from NOR flash to NAND to meet demand for NAND and eMMC storage. Management attributed the decline in Q3 NOR shipments to that reallocation. It said NOR demand was still growing, particularly for higher-density products used in data centers. Macronix plans to direct new capacity to NAND and eMMC, with no further expansion of NOR capacity. The first phase is expected to enter volume production in the second half of 2027. Output from the additional investment may not arrive until 2028. macronix.com/en-us/about/inv…

Amkor’s high-density fan-out (HDFO) packaging ramp took longer than expected in Q3, even as its computing business posted record revenue. The chip packaging and testing provider said its HDFO process involves more than 300 manufacturing steps. CEO Kevin Engel said getting enough work in progress moving through the production line took longer than planned. He said supply constraints were not the cause. Computing revenue rose 25% from Q2. Amkor continues to scale its flagship HDFO CPU program and expects two new 2.5D packaging programs for networking and data-center applications to come online in Q4. ir.amkor.com/events/event-de…

Japan’s Mitsubishi Materials plans to raise monthly production capacity for temperature sensors used in optical transceivers to about 12 million units by November 2027, roughly six times its 2025 level. The sensor, a flake thermistor measuring less than 1 mm, monitors the temperature inside transceivers, where temperature changes can affect communication performance. Mitsubishi Materials says the component is used by major transceiver makers worldwide. The company says growing data-center traffic from generative AI is driving demand for optical transceivers and its sensors. It plans to add production equipment in stages at its manufacturing site in Thailand. news.mmc.co.jp/corporate/ja/…

[Taiwan’s Weekly Standout #1] Each week, we spotlight a standout Taiwan-listed company and explore what its story reveals about the AI and semiconductor supply chain. PCB drilling equipment maker Ofuna Technology (4577 TT) rose 37.3% between the Oct. 16 and Oct. 22 closes, including a 10% gain on Oct. 22. Ofuna sells mechanical and laser drilling equipment and provides laser drilling services for PCB and IC substrate manufacturers. Its business spans equipment purchases and the processing work manufacturers outsource. September revenue rose 55.2% from a year earlier, bringing growth for the first nine months to 50.4%. The company attributed September’s increase primarily to higher equipment production and sales. The supply-chain backdrop connects with our Oct. 19 weekly roundup. DIGITIMES reported that lead times for ESI’s laser drilling systems, part of MKS ($MKSI), exceeded 18 months as substrate makers expanded capacity for AI chips. Higher substrate layer counts require more rounds of drilling and other processing steps. Expanding output therefore depends on both equipment availability and processing capacity. At the TPCA Show held last week, Ofuna’s lineup included substrate drilling equipment and CO₂/UV laser drilling services. Its revenue figures show equipment activity picking up. The next questions are how much growth is coming from high-end substrates, and whether drilling-service utilization is rising alongside equipment sales.

Vietnam’s first chip factory is unlikely to begin trial production in 2027 as planned, Reuters reports, citing a person with direct knowledge of the project. The source cited developer Viettel’s lack of experience building chip factories and administrative hurdles in securing government funding. Viettel told Reuters the project is progressing according to plan and on schedule. Viettel has been in talks with ASML $ASML for months but has yet to reach an equipment deal, according to three people familiar with the matter. Two said the discussions involve a deep ultraviolet (DUV) lithography machine. Authorities estimate the cost of the plant near Hanoi at about $1.3 billion. Reuters reported that no buildings had yet been constructed at the site. bnnbloomberg.ca/business/com…

Enphase’s 900V GaN devices have passed testing and are now going into development builds of its solid-state transformer. Production qualification is the next step toward commercial deployment.

Celestica says its lower 2026 free-cash-flow outlook mainly reflects working capital needed for 2027 growth. That means committing cash before it comes back through customer payments.

Celestica expects revenue to nearly double to $40 billion in 2027, up from its latest 2026 forecast of $21 billion. The data-center hardware supplier plans $1.5 billion to $2 billion in capital spending next year. It also lowered its 2026 non-GAAP free cash flow forecast to $460 million from $600 million, mainly because of working capital investments to support 2027 growth. Q3 revenue rose 83% year over year to $5.83 billion. Connectivity & Cloud Solutions revenue grew 102% to $4.87 billion. corporate.celestica.com/news…

Broadcom is helping finance AI hardware for Anthropic through a $63 billion deal announced with four financial partners on Oct. 26. Deliveries are expected to start in 2027. A new special-purpose vehicle will buy Broadcom’s AI accelerators and networking equipment and lease them to Anthropic. Broadcom will provide residual-value support equal to the vehicle’s senior debt, which Reuters puts at about $43.5 billion, citing two sources. The announcement doesn’t spell out what would trigger a payment from Broadcom. prnewswire.com/news-releases… investing.com/news/stock-mar…

Terafab has already ordered and paid for equipment, according to the supply-chain checks Sky discussed in EP571. Those payments gave him more confidence that substantial work was under way. He was much less certain about whether the project could achieve successful volume production. Producing a test chip would still leave yield and profitability unresolved. Sky emphasized the need to demonstrate that chips from Terafab’s own process could be manufactured at yields that support profitable production.

Verizon $VZ has signed an AI infrastructure deal worth more than $2 billion with Amazon Web Services. The agreement sets out a commercial framework to expand distributed computing for agentic and generative AI workloads, Verizon said in its Oct. 26 earnings release. The release does not specify the deal’s duration. Verizon’s 2026 capital spending guidance of $16.0 billion to $16.5 billion includes success-based investments in AI Connect. sec.gov/Archives/edgar/data/…

Lam Research, a chipmaking equipment supplier, expects global wafer fabrication equipment spending to exceed $200 billion in 2027. That preliminary forecast already accounts for constraints on cleanroom space and component supply. Asked what spending could look like without those limits, CFO Doug Bettinger said it would probably be higher, but declined to give a number. Lam sees 10 new fabs at major chipmakers coming online at different points during 2027. It expects DRAM to lead percentage growth in equipment spending, followed by leading-edge foundry and logic, then NAND. investor.lamresearch.com/ima…

NVIDIA now lists Rubin’s HBM4 bandwidth at 19.2TB/s, down from the earlier 22TB/s, while memory capacity stays at 288GB. The lower bandwidth could matter more for AI workloads that spend much of their time waiting for data from memory.

PCB equipment lead times could remain above two years and still send investors a warning signal. In EP571, Sky used a hypothetical drop from 30 months to 25 months to illustrate the risk. A shorter queue could prompt a sharp reassessment of equipment suppliers, even while their delivery schedules still look unusually long. He said previous cycles had sometimes turned within one or two months. For equipment makers, he is focused on whether order books are still growing while delivery schedules remain stretched.

Five developments in the AI supply chain from the week of Oct. 19 to 25. 1. Unimicron sees material shortages ahead as equipment lead times stretch. Unimicron (3037 TT) sees a potential supply shortfall of about 20% for key PCB and substrate materials in 2027, according to Chairman SC Chien. Supplies of T-glass cloth, copper-clad laminates and ABF began tightening one to two years ago. Equipment lead times have lengthened for more than 80% of tools and doubled for about 20%, he said. Lead times for some machines now reach 30 months. xitter.cf/statementdog/status/21… Unimicron’s president also said double booking is inevitable given how tight the AI supply chain remains. Bringing new PCB capacity online takes 2.5 to 3 years or longer, from construction to production. xitter.cf/statementdog/status/21… 2. Nvidia is committing billions to expand back-end capacity. Nvidia ($NVDA) disclosed a $1.1 billion prepayment to KYEC (2449 TT) for new U.S. testing facilities and equipment. High-volume production is targeted for April 2027. xitter.cf/statementdog/status/21… A new $1.3 billion deal brings the total value of Amkor’s multiyear agreements with Nvidia across South Korea and Arizona to $3 billion. The companies aim to increase interposer wafer capacity in South Korea by more than 300% by the end of 2028. xitter.cf/statementdog/status/21… SPIL and Nvidia also aim to expand optical-engine capacity for co-packaged optics in Taiwan to 15 times its current level by mid-2027, then double it again by the end of 2028. xitter.cf/statementdog/status/21… 3. TPU v10 could shift to TSMC packaging. MediaTek’s TPU v10 could use TSMC’s CoWoS or CoPoS packaging, with mass production targeted for 2029, according to Citi. That would mark a change from the Intel EMIB-T packaging planned for TPU v9, which Citi says faces potential yield challenges. xitter.cf/statementdog/status/21… 4. Liquid-cooled racks could command much higher prices. Our supply-chain checks suggest Wiwynn (6669 TT) could ramp shipments of liquid-cooled racks for AWS Trainium 3 from late 2026 through the first half of 2027. These racks could sell for three to four times the price of air-cooled versions, or more. xitter.cf/statementdog/status/21… Omdia expects cold plates’ share of revenue from single-phase direct-to-chip liquid cooling to rise from 40% in 2024 to 61% by 2030. xitter.cf/statementdog/status/21… 5. AWS customers are booking large blocks of capacity years ahead. Most of AWS’s large blocks of compute capacity for 2027 are already booked, and a significant portion of its 2028 capacity is also committed, according to Amazon CEO Andy Jassy. AWS still sets aside capacity for on-demand use and services such as Bedrock and EC2. xitter.cf/statementdog/status/21…

At GTC Berlin, Jensen Huang cited roughly $2,000 per GPU per year for NVIDIA software that runs AI agents. He expects software licensing to become a larger part of the company’s business. Enterprises deploying agents need controls over what those agents can access and do. NVIDIA’s OpenShell Reference Architecture combines its computing hardware with tools to monitor agents and stop them when they cross permitted boundaries. OpenShell’s runtime is open source. The scope and timing of the paid licenses Huang discussed remain unclear, including how they would fit with existing enterprise offerings. Our read: annual software licenses could let NVIDIA keep earning from a GPU after the initial hardware sale. How much revenue that adds will depend on how many GPUs customers buy licenses for and whether they keep renewing. tomshardware.com/pc-componen… nvidianews.nvidia.com/news/o…