目录
Pre-Market Highlights
Full AI/Semiconductor Value Chain
AI Models/Applications and Capital Expenditure
CSP/Cloud Capital Expenditure
GPU/CPU/ASIC
HBM/DRAM/NAND/SSD/HDD
Foundries, Equipment, and Optical Communications
Internet/Platforms
Software/SaaS
Consumer Electronics / Smart Vehicles
Overview
404K | 2026-08-06
Pre-Market Highlights
Demand for AI infrastructure has not cooled, but market attention has shifted from “whether to keep investing” to “where the money is being spent and when it will generate returns.” The cloud businesses of Google, Microsoft, Amazon, and Meta continue to accelerate, accompanied by upward revisions to capital expenditure; meanwhile, pressure from free cash flow, long-term lease commitments, and project financing continues to rise.
Hardware bottlenecks are beginning to spill over. TSMC is expanding 2nm and 3nm capacity, memory manufacturers are locking in 2027 capacity, and ABF substrate orders extend into 2028; revenue inflection points for CPO, NPO, coherent optical communications, and indium phosphide are also moving forward. Lower memory content per system does not mean weaker aggregate demand, and supply capability remains the key constraint.
The risk lies in execution rather than orders. Volatile loads at AI data centers are accelerating the aging of batteries, turbines, and cooling equipment, while some projects also face construction delays, financing costs, and foundry ramp-up issues. Key items to watch are whether cloud gross margins improve through contract repricing and whether memory, optical, and power-equipment suppliers can convert orders into cash flow.
Full AI/Semiconductor Value Chain
AI Models/Applications and Capital Expenditure
OpenAI
Model demand continues to extend toward larger datasets and greater computing power. A “brain signal-to-text” project proposes expanding data collection by several orders of magnitude and using larger models to learn the mapping between brain signals and text directly. The commercial implication is that training demand will continue to spread into new modalities, but data collection, privacy, and real-world usability have yet to be validated.
“We must expand the scale of data collection by several orders of magnitude, far beyond anything academia has previously done.”
Anthropic
1) The nightly version of Claude Code adds sub-agent and workflow visualization, allowing enterprise customers to see more intuitively how multiple agents divide work.
2) Its Texas data center project plans to issue $15 billion of debt, while Google’s guarantee will take effect only after the facility is completed; during construction, bondholders will continue to bear delay and cost-overrun risks.
Enterprise Agent Applications
Microsoft 365 Copilot now has more than 30 million paid seats, while Meta Business Agents serves more than 1 million businesses each week; approximately 1 in every 3 GitHub pull requests involves an agent. Adoption is creating measurable demand, but the next items to assess are seat activity, renewals, and customers’ actual returns.
“Before deployment, cloud platforms need to demonstrate that their technology stacks can withstand real-world workloads—testing actual models, configurations, and infrastructure behavior under conditions close to actual deployment.”
AI Data Center Power Reliability
Reportedly, AI facilities sometimes consume 50% more power than their designed capacity, and the instantaneous load of a 1 GW facility may reach 1.5 GW. Batteries, generators, cooling systems, and gas turbines have already shown premature aging or cracking, with downtime losses ranging from thousands to hundreds of thousands of dollars per minute; it also remains to be seen whether batteries and auxiliary computing can smooth the power curve.
AI Infrastructure Financing
Microsoft, Meta, Oracle, Amazon, and Google have committed approximately $1.1 trillion in rent for data centers whose leases have not yet commenced, approximately 4 times the $285 billion of recognized lease liabilities. Capacity has been locked in for the long term, but liabilities begin to be recognized only after buildings are delivered, delaying validation of returns.
CSP/Cloud Capital Expenditure
Google
1) Google Cloud revenue was $24.77 billion, up 82% year over year, with operating profit of $8.8 billion and remaining performance obligations of $514 billion. Customers’ actual spending was approximately 50% above their commitments, as usage had exceeded contractual minimums.
2) Capital-expenditure guidance was raised to $195 billion–$205 billion, while free cash flow was negative $5.9 billion, indicating simultaneous acceleration in demand realization and cash consumption.
“Second-quarter usage exceeding commitments, customer prepayments or customer-supplied GPUs, changes in backlog composition, and simultaneous capital-expenditure increases by four vendors constitute early evidence of repricing; L3 memory and L4 optical interconnects, which are difficult to supply internally, may be the longest-lasting hardware beneficiaries.”
Microsoft
Azure grew 43%, while commercial remaining performance obligations reached $678 billion, up 84% year over year; quarterly capital expenditure was $41 billion, up 69% year over year, and FY2027 capital-expenditure guidance was $255 billion–$260 billion. Demand continues to exceed capacity, but OpenAI customer concentration and depreciation methodology are the principal variables affecting returns.
Amazon
AWS revenue was $42.2 billion, up 37% year over year, its fastest growth in 18 quarters, while operating margin reached 39.4%; Amazon raised its expected 2026 capital expenditure from approximately $200 billion to approximately $220 billion. Orders and margins support the demand thesis, while the risk is that AI revenue over the next 3 years remains highly dependent on large model customers.
Meta
Revenue grew 28%, but expenses increased 55% and operating profit declined 8%; quarterly AI infrastructure spending was approximately $31 billion, leaving only $784 million of free cash flow after capital expenditure. Full-year capital-expenditure guidance is $130 billion–$145 billion, and Meta will subsequently need to demonstrate returns through advertising and agent revenue.
Oracle
Oracle’s backlog is approximately $638 billion, of which approximately $75 billion represents customer prepayments for GPU purchases or customer-supplied hardware; future data center lease commitments total approximately $260 billion, with contract terms of up to 15–19 years. Prepayments reduce procurement pressure, but long-term leases and refinancing costs extend execution risk.
“Accelerating cloud growth at three companies and simultaneous capital-expenditure increases at four support the view that ‘demand exceeds supply,’ but share prices are under pressure from free-cash-flow concerns, creating a central tension between demand realization and cash consumption.”
GPU/CPU/ASIC
NVIDIA
1) Vera CPU SOCAMM capacity has been reduced from 192GB to 96GB, with some SKUs reduced to 64GB; Rubin Ultra HBM has been adjusted from 384GB to 192GB or 256GB. Suppliers can meet only 60%–70% of the original plan, indicating that the configuration reductions reflect supply constraints.
2) Platform system shipments are expected to reach 5.8 million units and 8.1 million units in 2027 and 2028, respectively, while NVL576 will use NPO optical interconnects to expand scale-up capacity.
The specification reductions stem from supply constraints, principally insufficient SOCAMM2 supply.
AMD
Agent orchestration, tool calling, and post-training are increasing joint demand for CPUs and GPUs, and MI450 Custom has adopted an 8-layer HBM architecture. NVIDIA, AMD, and hyperscale cloud providers are also reserving ABF substrates in advance; related capacity reportedly has already been booked through 2028, prompting suppliers to plan capacity expansion for 2029–2030.
Imagination
The E-Series GPU uses a multi-precision fused dot-product unit to share multiplication, alignment, and summation resources across INT8, FP8, FP16, and BF16, saving nearly 1/3 of circuit area relative to separate pipeline designs. The design has completed formal verification, but performance on real models, power consumption, frequency, and mass-production yield have not yet been disclosed.
New Cloud Platform Deployments
Dell and CoreWeave have completed the first validated Rubin launch, while AWS, Azure, Oracle Cloud, Lambda, Nebius, and Nscale plan to make it available to enterprise customers in the second half of the year; second-tier new cloud platforms are expected to follow in 2027, with a lag of approximately 6–12 months.
GPU Rental Repricing
GPU spot rental prices have reportedly been observed at no less than 2 times long-term contract prices, and renewals of legacy contracts may improve cloud providers’ operating cash flow. Evidence against the thesis would include cloud gross margins failing to rise after renewals, a decline in the proportion of usage exceeding commitments, or legacy GPU prices falling 30%–50% within 12 months after new products ramp.
HBM/DRAM/NAND/SSD/HDD
Micron
Samsung Electronics, SK hynix, and Micron reportedly have locked in 2027 DRAM and HBM capacity through pre-allocation agreements, while manufacturers can typically meet only 60%–70% of customers’ actual demand. Micron has also signed a 10-year silicon-wafer supply agreement with GlobalWafers, indicating that volume commitments for raw materials have extended further upstream.
Kioxia
Reportedly, Kioxia began mass production of tenth-generation BiCS10 3D NAND at its K2 wafer fab in Kitakami, Iwate Prefecture, in early July. The additional capacity will help alleviate long-term supply constraints but will also change the supply cadence of Kioxia and Sandisk; subsequent monitoring should focus on yield, customer qualification, actual bit shipments, and pricing, rather than merely the start of mass production.
Samsung Electronics
Samsung Electronics introduced zNAND-O at FMS, bringing high-bandwidth flash memory to edge AI; Qualcomm and Google are evaluating adoption, while commercialization still depends on cost. Meanwhile, HBM and AI servers may consume nearly 70% of total DRAM output, tightening mobile supply, and Samsung Electronics is also evaluating a dual-supplier strategy for Galaxy OLED display-driver chips.
Sandisk
1) FY2026 fourth-quarter revenue was $8.965 billion, up 372% year over year, with a gross margin of 84.6% and data center revenue growth of 437%.
2) Multi-year NAND agreements with 8 customers cover more than half of FY2027 bit volume and approximately 2/3 of FY2028 bit volume, with floor-price-protected revenue of approximately $93.9 billion.
3) Next-quarter revenue guidance is $10.3 billion–$10.8 billion, with gross margin of 83%–85%; sustainability depends on the supply response.
“Although margin guidance was reduced slightly quarter over quarter, the absolute level remains close to 84% even as revenue growth continues to accelerate sharply, which the market interpreted as strength rather than weakness.”
