01. DDR5 Memory Prices Rise Another 15%, Suppliers Force-Cancel Legacy Orders
On October 21, MyDrivers reported that U.S. systems integrator Stince Built has warned DDR5 prices will climb another 10–15% this week. One memory supplier, after issuing a price-hike notice, directly canceled all undelivered low-price orders and demanded customers reorder at the new prices, underscoring strong upstream bullish sentiment. This round of increases has already exceeded systems integrators' ability to absorb margins, suggesting end-system retail prices will likely be forced higher in the long run, with Q4 build costs facing a significant jump.
02. TrendForce: Server CPU Shipments to Surge Over 40% in 2027
On October 23, Economic Daily reported that TrendForce estimates server CPU shipments will grow 31% year-over-year in 2026, and could surge over 40% in 2027 as AI inference and agentic AI penetration deepen. AI agents are driving the CPU-to-GPU ratio from 1:8 toward 1:2 or even 1:1. In terms of market share, Intel is losing ground due to process and capacity constraints; AMD's share is projected to rise to 27% in 2026. The Arm architecture is benefiting from the ramp of custom chips by AWS, NVIDIA, and Google, with share breaking 20% in 2026 and potentially expanding to 30% in 2027.
03. Qualcomm Plans Global Layoffs, Affected Employees to Be Notified Starting November
On October 21, SemiMedia cited Reuters reporting that Qualcomm plans global layoffs, with affected employees set to receive notice starting in November. CEO Cristiano Amon announced the plan in an internal memo on October 19 but did not disclose specific numbers. As of September 2025, Qualcomm employed roughly 52,000 people globally. The move comes amid soft consumer electronics demand and memory price hikes battering the smartphone market, with Qualcomm previously warning that Apple-related revenue will decline faster. The company is expanding into non-handset businesses such as AI data centers, expecting new-business growth to offset the loss of Apple orders by fiscal 2027.
04. Micron Deems CXMT's 4F² Architecture Unworkable, Dismisses Its Viability
On October 23, Economic Daily reported that Micron President Mark DeBoer stated bluntly that CXMT's 4F² architecture is "more expensive and more complex, with no viable path to scale," dismissing its prospects and becoming the first major memory manufacturer to publicly weigh in since the "CXMT shock." He noted that capacitor pitch is the key determinant of DRAM performance, and 4F² offers no practical advantage while running into physical limits; Micron will stick with the 6F² and 3D DRAM roadmap. He also pointed out that CXMT lacks EUV equipment, casting doubt on its mass-production capability. Industry players believe the technological breakthrough remains far from large-scale supply, and this intervention could help ease market jitters, benefiting Taiwanese manufacturers such as Nanya Technology and Winbond Electronics, with focus returning to supply-demand fundamentals.
05. PSMC to Raise Mature-Process Prices 10–20% Next Year; AI Server Clients Rush to Lock in Long-Term Deals
On October 22, Financial News reported that at its investor conference, PSMC noted that logic foundry capacity remains in short supply. Following across-the-board hikes in 2026, mature-process prices are planned to rise another 10–20% in 2027, with priority allocation to long-term agreement (LTA) customers. In Q3, non-annual-contract customers already saw wafer start prices adjusted up 10–15%, with average prices rising 8–10%. On the product front, power management ICs are entering 55nm BCD processes, with AI server clients requesting three-year LTAs; roughly 80% of MOSFET and IGBT capacity for 2027 is already under LTA. Automotive TDDI demand is heating up significantly, with capacity shortages and extended lead times unabated, prompting downstream panel makers to repeatedly expedite orders.
06. Google Makes Major Bet on Nvidia Vera Rubin Servers, Snapping Up 960,000 Rubin GPUs
On October 23, Economic Daily reported that Google Cloud announced it will massively deploy Nvidia Vera Rubin AI servers, scaling up to 960,000 Rubin GPUs, integrated with ConnectX-9 SuperNICs and the Virgo networking architecture alongside fourth-generation liquid cooling technology, with full deployment beginning in November. Google becomes the first customer to adopt the platform in a bare-metal configuration. Analysts are bullish on Taiwanese suppliers including TSMC, Quanta, Wiwynn, and Foxconn benefiting from the deal. Industry sources noted that Vera Rubin uses TSMC's CoWoS-L packaging, and at this scale, Google has effectively booked nearly all of TSMC's Rubin capacity for the next one to two years.