目录
Key Themes
AI & Semiconductor Value Chain
AI Models, Applications & CapEx
CSP & Cloud CapEx
GPU, CPU & ASIC
HBM, DRAM, NAND, SSD & HDD
Pure-Play Foundry
Semiconductor Equipment & Testing
Advanced Packaging & OSAT
Optical Communications & Optics Supply Chain
Internet & Platforms
Software & SaaS
Consumer Electronics & Smart Vehicles
The focus in optical communications is expanding beyond raw bandwidth to light sources, packaging, and testing. TSMC's COUPE technology roadmap targets higher integration, while optical component suppliers advance qualification and capacity expansion. On the terminal side, upcoming foldable models could catalyze high-end smartphone demand, though memory price hikes continue to pressure OEM margins.
404K SEMI-AI | 2026-09-07
Key Themes
The incremental focus in tech hardware is whether supporting supply chain components can keep pace. Technical discussions on COUPE have expanded into light source integration, heterogeneous bonding, and thermal isolation, while Nomura's tracking of optical components highlights qualification progress and commercial shipments. Converting bandwidth upgrades into revenue requires fabricating, testing, and reliably delivering components at scale; volume production yields are far more critical to track than isolated peak specifications.
End-market demand and component costs remain in tension. Projections for foldables suggest accelerating market expansion once Apple enters the segment, though initial supply constraints and premium pricing could temper adoption speed. Concurrently, memory tightness is spilling over into smartphones and graphics cards. Assessing high-end hardware growth requires monitoring new product shipments and realized selling prices alongside OEMs' capacity to absorb cost inflation.
AI & Semiconductor Value Chain
AI Models, Applications & CapEx
OpenAI
1) The operational scope of research agents has expanded from infrastructure coding into monitoring training runs, debugging experiments, and troubleshooting. Reports indicate that for every 1 day of human labor invested, agents deliver work equivalent to 3 human-days.
2) Long-horizon tasks still rely heavily on human intervention: autonomous success rates fall from 86% for tasks under 15 minutes to ~16% for 64–128-hour tasks; more than half of completed 4–8-hour tasks required human intervention. Increases in task throughput must be distinguished from true end-to-end autonomy.
3) A single researcher consumes approximately $7,000 per day in model inference. Actual ROI must balance task deliverables and labor savings against underlying compute expenses.
Anthropic: Concrete case studies have emerged demonstrating Claude's experimental execution capabilities. Reports indicate that in a Tetris experiment, Claude Code directed a weight-frozen Gemma 4 over 2.5 days, evaluating 90 hypotheses across more than 400 games at a cost of approximately $1,200, improving the score from 0 to 16. The experiment involved no model training; performance gains stemmed from standardizing the test environment, using median-based evaluations, and maintaining meticulous experiment logs. While validating the utility of agent-assisted experimentation, issues such as single-run score noise and rule-skirting behaviors still require control.
CSP & Cloud CapEx
Amazon: AI rack pull-in momentum is accelerating, strengthening supply chain shipments. CEO Andy Jassy reportedly expects compute tightness to persist through 2027–2028, with AWS AI and custom silicon revenue reaching approximately $25 billion and maintaining rapid growth. Key supply chain beneficiaries include Alchip (ASIC design services), Wiwynn (ASIC racks and switch trays), and Foxconn (CPU trays). As demand shifts to physical hardware deliveries, investors should track whether supplier production and top-line growth can be sustained.
