目录
Pre-Market Highlights
Full AI/Semiconductor Supply Chain
CSP/Cloud Capital Expenditure
GPU/CPU/ASIC
HBM/DRAM/NAND/SSD/HDD
Foundry, Advanced Packaging, and Equipment
PCB/Materials and Optical Communications
Internet/Platforms
Software/SaaS
Consumer Electronics / Smart Vehicles
Overview
404K SEMI-AI | 2026-08-03
Pre-Market Highlights
Cloud demand has not cooled despite elevated capital spending; instead, it continues to accelerate. Orders, revenue, and investment at Amazon, Microsoft, Google, and Meta all point to the same conclusion: existing compute capacity remains insufficient, supply constraints through 2027 have yet to ease, and the market is increasingly focused on whether incremental investment can translate into revenue and cash flow.
Bottlenecks are spreading beyond GPUs. HBM, NAND, advanced packaging, silicon wafers, copper-clad laminates, and optical interconnects are all showing pricing, capacity, or qualification signals, indicating that individual chips no longer determine delivery speeds for entire systems. Supply-chain opportunities are becoming more dispersed, but requirements for yields, qualification, and mass-production ramp schedules are also rising.
Two forms of application-side validation are emerging: enterprise software can already improve gross margins or shorten delivery times, but large-scale monetization remains at an early stage; near-term autonomous-driving safety data have improved significantly, but evidence spanning 10 years is still lacking for long-term aging, sensor drift, and redundancy certification.
Full AI/Semiconductor Supply Chain
CSP/Cloud Capital Expenditure
Amazon
1) AWS revenue reached $42.2 billion, up 37% year over year, its fastest growth in 18 quarters; backlog increased to $496 billion, up $132 billion in a single quarter.
2) Capital expenditure guidance was raised to $220 billion. The company believes capacity will remain insufficient in 2026 and 2027, while demand for 2028 is already very strong.
3) Annual revenue from the Trainium and Graviton chip businesses exceeded $25 billion, and Anthropic and OpenAI have made multiyear, gigawatt-scale commitments.
“As strong compute growth runs up against shortages... pricing power is starting to accrue to those who have it.”
Google: Citi estimates second-quarter capital expenditure at $45 billion, up 100% year over year, and raised its 2026 guidance from $180 billion–$190 billion to $195 billion–$205 billion. Leading in spending does not automatically translate into leading returns; the next question is whether TPU volume expansion can continue to lift cloud revenue and HBM consumption.
Microsoft: Second-quarter capital expenditure was $41 billion, up 69% year over year; adjusted 2026 capital expenditure was $175 billion, while third-quarter guidance exceeded $50 billion. Morgan Stanley also noted that Azure growth accelerated by 4 percentage points sequentially to 43%, with guidance of approximately 45% for the next quarter, indicating that new capacity continues to be absorbed rapidly by demand.
Meta: Second-quarter capital expenditure was $31 billion, up 83% year over year, while the 2026 range was narrowed to $130 billion–$145 billion. The CFO said compute-resource supply constraints will persist, with training, inference, and consumer AI products all competing for capacity; if supply eases more slowly than revenue grows, depreciation and free-cash-flow pressure will emerge sooner.
“Market demand remains exceptionally strong, so the key is unlocking supply capacity.”
Hyperscale Cloud Providers: Citi estimates aggregate second-quarter capital expenditure across four companies at $170 billion, up 79% year over year; its 2026 and 2027 forecasts were raised to $754 billion and $1.112 trillion, respectively. JPMorgan said the combined remaining performance obligations of three companies reached $1.7 trillion, up 152% year over year, with supply still the near-term constraint.
GPU/CPU/ASIC
NVIDIA
1) NVIDIA is expected to remain the largest single HBM buyer in 2027, accounting for 36% of consumption.
2) Vera Rubin reportedly delivers at least 3 times the performance of GB300, while supporting CPO switches have been delivered to key customers and deployed internally.
3) Scale-up bandwidth requirements can reach 10 times those of scale-out, making network and optical-engine supply a more direct determinant of complete-system deliveries.
AMD
1) The Lux and Discovery supercomputers have entered construction and deployment under US Department of Energy projects, with EPYC, Instinct, Pensando, and ROCm delivered as an integrated platform.
2) AMD’s share of HBM consumption is expected to reach 16% by 2027. This means competition is not limited to GPU performance, but also depends on whether CPUs, networking, software, and memory can be deployed together.
MediaTek
1) Morgan Stanley maintained its Overweight rating and NT$5,588 price target, while raising its 2026 EPS forecast by 2.5%.
2) The report expects AI ASIC revenue to grow at a CAGR of more than 50% from 2025 to 2028, with Google TPU demand potentially offsetting smartphone-business headwinds.
3) Citi said the company is advancing work with a second ASIC customer, but smartphone replacement demand and TPU-related gross-margin dilution remain the principal risks.
“Overall AI ASIC demand looks very strong, especially from Google’s TPU.”
HBM/DRAM/NAND/SSD/HDD
Samsung Electronics and SK hynix
1) Goldman Sachs expects Samsung Electronics’ DRAM, NAND, and HBM products to remain undersupplied from 2026 to 2028, with long-term agreements covering 60%–70% of planned capacity and including prepayments and price floors.
2) Another forecast indicates that SK hynix will hold a 48% share of HBM bit shipments in 2026; in 2027, Samsung Electronics may rise to 41%, slightly above SK hynix’s 39%.
3) DRAM export unit prices continued to rise in July, shifting the competitive variables toward HBM4 yields, customer qualification, and execution of long-term agreements.
Kioxia
1) The company expects the NAND market to reach 1,807 exabytes by 2028, approximately 1.8 times its 2025 level, with data-center demand rising from approximately 30% to approximately 50% of the total.
2) The LC9 series offers capacities of up to 245TB, while the GP series has entered qualification, seeking to capture data in certain AI-storage use cases beyond HBM.
3) The company will pay $229 million to settle patent litigation, requiring investors to monitor product qualification and legal costs concurrently.
Memory Bottleneck: Processor and memory-access speeds differ by approximately 250–500 times. The problem of GPUs waiting for data is shifting value toward HBM, CXL, advanced packaging, SSDs, and optical interconnects. By 2027, Google is expected to account for 35% of HBM consumption and AWS for 5%, as scaling proprietary chips reshapes the memory-customer mix.
Foundry, Advanced Packaging, and Equipment
TSMC
1) Citi said the company continues to benefit from tight advanced-node and CoWoS capacity, and raised full-year revenue-growth guidance to 40% and capital expenditure to $60 billion–$64 billion.
2) Reportedly, monthly 3-nanometer wafer starts may rise to 180,000, 2–3 months earlier than expected; 2-nanometer wafer starts could still approach 100,000 by year-end.
3) Order strength is clear, but whether high investment can translate into yields and packaging output as planned remains a key validation point.
“TSMC remains the clear manufacturing leader, benefiting from sustained advanced-node demand.”
UMC: The company raised 2026 capital expenditure by 33%, from $1.5 billion to $2.0 billion, focusing on silicon photonics and advanced packaging; current AI revenue is approximately $300 million, and management expects it to reach $1.0 billion within 3 years. Its 12-inch silicon-photonics platform has entered early production, but meaningful incremental contribution will require commercial revenue to scale in 2027.
Intel
1) MediaTek confirmed that its second cloud ASIC project using EMIB-T is progressing as scheduled, with initial production expected to begin in early 2028.
2) The key variables in advanced packaging remain supplier yields, capacity, and cycle times. Project success could provide additional external-customer validation for Intel’s foundry and packaging businesses.
3) Pat Gelsinger has joined lithography start-up xLight and is seeking TSMC’s participation, but the commercialization timeline has yet to be confirmed.
“Yes, you can assume that the second chip will start the initial production in early 2028.”
ASML: In certain logic-chip and DRAM applications, High-NA can replace an EUV sequence using 3 masks with 1 mask and compress approximately 100 patterning steps to approximately 10. The benefits are shorter cycle times and less cumulative defect formation, but exposure fields, stitching, photoresists, and metrology systems must all be readapted.
Siltronic and GlobalWafers: JPMorgan said small-volume, non-long-term-agreement pricing for 300-millimeter silicon wafers has improved slightly, while long-term-agreement pricing is unchanged; 200-millimeter pricing is expected to stabilize after declining in the first half. Siltronic is not yet initiating new capital expenditure, while global silicon-wafer area demand is still expected to grow by approximately 7% in 2026; GlobalWafers’ subsequent gross-margin performance warrants greater attention.
ASE Technology Holding: Advanced packaging and testing complexity continues to rise, and the company expects LEAP revenue to exceed $3.5 billion in 2026 and double in 2027. Incremental growth will come from CoWoS-related packaging, advanced testing, and panel-level technologies; demand strength is clear, but execution still depends on expansion yields and customer product ramps.
PCB/Materials and Optical Communications
Shengyi Technology
1) Goldman Sachs expects the company’s net profit to grow by 104% and 73% in 2026 and 2027, respectively, with AI-server PCBs accounting for approximately 40% of revenue in 2027 and more than 50% over the longer term.
2) GPU server racks are migrating toward M9-grade copper-clad laminates in 2026–2027, while the company is expanding to ASIC customers and has announced RMB5.2 billion of capital expenditure for copper-clad-laminate production lines.
3) Goldman Sachs maintained its Buy rating and RMB247 price target; valuation realization depends on price increases for high-end materials, yields, and customer qualification.
“customer expansion from GPU AI servers to ASIC AI servers will drive strong net income growth.”
Elite Material and High-End Copper-Clad Laminates: The global copper-clad-laminate market grew from $15.01 billion in 2024 to $18.68 billion in 2025, up 24.4% year over year; Elite Material’s share rose to 16.2%, increasing by 3 percentage points in 1 year. The upgrade from 800G to 1.6T continues to raise the threshold for low-loss materials, with the true barrier being qualification first and volume expansion afterward.
