目录
Pre-Market Highlights
Full AI/Semiconductor Value Chain
AI Models/Applications and Capital Expenditure
CSP/Cloud Capital Expenditure and AI Cloud
GPU/CPU/ASIC
HBM/DRAM/NAND/SSD/HDD
Foundries, Advanced Packaging, and Optical-Communications Testing
Internet/Platforms
Software/SaaS
Consumer Electronics/Smart Vehicles
Overview
404K | 2026-08-11
Pre-Market Highlights
Technology-stock trading continues to broaden from a singular AI-chip narrative toward software, internet, and cloud platforms. Hardware demand has not weakened, but positive news is providing less support to share prices; software companies with reaccelerating growth, lower valuations, and prior underperformance are instead more likely to attract follow-on buying.
AI-infrastructure constraints are expanding beyond chip availability to memory, financing, and interconnects. Micron said some data-center customers can secure only about half of their DRAM requirements, while Nvidia is testing lower HBM capacity for Rubin Ultra, trading reduced per-system configurations for higher system shipments.
A new source of incremental financing is emerging. Nvidia has joined six major capital institutions in seeking to mobilize more than $500 billion in third-party funding over time. The key validation point is not the headline amount, but whether GPU utilization, cash flow, and residual values can support independent underwriting; if residual values for previous-generation compute prove unstable, capital costs will still tighten in response.
Full AI/Semiconductor Value Chain
AI Models/Applications and Capital Expenditure
OpenAI: The company completed a $7 billion secondary share sale at an $852 billion valuation, providing employees with pre-IPO liquidity; its previous two tender offers totaled $6.6 billion and $1.5 billion, respectively. The company has confidentially filed for an IPO, but has not yet disclosed a formal listing timetable. Talent retention and valuation realization remain key follow-up validation points, while the true pricing anchors remain revenue growth and cash burn disclosed in the listing documents.
Anthropic
1) Claude will embed machine-readable markers in text generated by certain new models, while supported image files will also include signed C2PA metadata; extensive editing, rewriting, or translation may weaken the detection signal.
2) The company is reportedly targeting late September to early October for its IPO and plans to expand investment in healthcare and biotechnology.
Meta: Muse Glimmer entered the Code Arena web-development leaderboard for the first time, scoring 1359 and ranking 26th among open models. The product is positioned as a midsize open-weight agent model. Its near-term significance is that Meta is rebuilding its open-model product portfolio; actual enterprise adoption, inference costs, and the subsequent Spark 1.2 release are more important to track.
AI in Semiconductor Manufacturing: The bottlenecks to industry-wide scaling remain data and systems engineering. More than 70% of projects do not progress beyond the pilot stage; integration and deployment typically cost from $10 million to $15 million, while connecting legacy equipment may exceed budgets by 40%-60%. Once the foundations are in place, defect detection can improve by 15%-25% and production cycles can be shortened by as much as 50%, but payback periods are typically 3-5 years.
AI Drug Discovery: Design capabilities are expanding faster than physical experimentation. Evo generated 285 viral-genome designs, but only 16 proved effective after synthesis and testing; this means inexpensive models will amplify consumption of DNA synthesis, mass spectrometry, and reagents rather than eliminate wet-lab work. The counterevidence is that clinical failure rates remain about 90%, and the real bottleneck may continue to reside in human trials.
"AI can generate large numbers of candidate molecules at low cost, but cannot replace physical manufacturing and measurement within the design-synthesis-test-learn loop"
CSP/Cloud Capital Expenditure and AI Cloud
Microsoft
1) Commercial remaining performance obligations increased 84% year over year to $678 billion, and still grew 25% excluding OpenAI, indicating that enterprise-customer commitments are not supported by a single frontier-model company.
2) Maia 300 is reportedly scheduled for release in the fall, with the company negotiating for TSMC capacity covering more than 300,000 chips for delivery in 2027; the order has yet to be confirmed.
Google: Google Cloud’s backlog increased by more than $50 billion quarter over quarter to $514 billion, driven primarily by enterprise AI products. Demand remains strong, but revenue realization depends on data centers becoming operational, inference utilization, and the pace of customer consumption. If capital expenditure increases faster than cloud revenue, it will also continue to depress near-term free cash flow; order conversion and depreciation pressure need to be monitored in parallel.
Amazon: The company disclosed a $496 billion backlog, with year-over-year growth reaching triple digits. Supply-chain sources also said that the custom ASIC for Trainium 3 uses TSMC’s N3P process, while the relevant designer’s July sales increased 181.8% year over year to NT$7.43 billion, indicating that internally developed chips are entering volume production and ramp-up.
Riot Platforms: The company reportedly reached a long-term compute-purchasing agreement with Anthropic, with market estimates ranging from $9 billion to $9.1 billion. The 20-year contract would provide 191 MW of capacity from the Rockdale campus in Texas. Given the discrepancy in reported amounts, the value should be treated as a range pending confirmation in formal filings; customer credit and long-term electricity prices are the principal risks.
Fermi: The company signed a binding 15-year AI data-center lease with TensorWave, representing approximately $6.5 billion in contracted revenue and 222 MW of capacity, with phased delivery planned from the second half of 2027. Expansion rights could increase capacity to more than 650 MW, but more than $1.5 billion has already been invested in construction; power supply, financing, and delivery progress will determine whether the projected revenue visibility can be realized.
"Commercial remaining performance obligations grew 84% to $678 billion. All sequential growth in commercial RPO was driven by commitments from customers other than Frontier Model companies; excluding OpenAI, RPO grew 25%"
GPU/CPU/ASIC
Nvidia
1) The company is working with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR, seeking to mobilize more than $500 billion in third-party capital over time; financing providers will underwrite independently, while Nvidia will provide residual-value support of up to 25% only for certain projects.
2) A Morgan Stanley survey estimates that Rubin and Rubin Ultra shipments will approach 7 million units in 2027, including approximately 90,000 Rubin NVL72 racks.
3) Amid HBM4E shortages, the company is testing a reduction in capacity per GPU from approximately 256GB to 192GB; final specifications have not been determined.
Intel: The company conducted its first public share sale since 1971, raising an estimated $15 billion to $20 billion, with the offering several times oversubscribed. The company previously disclosed more than $30 billion in cash and a $10 billion revolving credit facility. The issuance may create additional capacity for investment in advanced processes and packaging, but the dilution and ultimate use of proceeds still require confirmation.
AMD: At OCP APAC, the company emphasized the importance of open networking, with UALink addressing scale-up within racks and Ultra Ethernet addressing low-latency scale-out between racks and clusters. Competition has expanded from individual GPU performance to interconnect architecture; if the open ecosystem reaches scale, customers will place greater weight on system-level throughput and network costs.
ARM: The company is building an open ecosystem around chiplet support, server standardization, and production-ready server designs, with participants spanning foundries, packaging and testing, EDA, memory, networking, and systems vendors. The commercial validation point is not merely the number of licenses, but whether the platform can shorten customers’ design cycles and expand CPU share in AI servers; customer adoption and software compatibility of the first volume-production platforms will determine the pace of progress.
MediaTek: July sales were NT$48.5 billion, down more than 10% month over month and up 12% year over year; the handset business remained weak, while custom ASICs and smart-device platforms offset the shortfall. Its first high-performance AI ASIC is scheduled for mass production in Q4, while data-center revenue is expected to exceed $2 billion this year and accelerate in 2027; customer ramp-up and product mix will determine profit sensitivity.
"Financing cannot create underlying demand, but if Nvidia AI systems offer leading ROI, the CUDA ecosystem, predictable residual values, and lower capital costs, financing can remove customers’ capital constraints"
HBM/DRAM/NAND/SSD/HDD
Micron
1) The company said memory demand growth may continue to outpace supply in 2027, with some data-center customers able to secure only about half of their DRAM requirements; agentic workloads consume approximately 5-30 times as many tokens as chat AI.
2) Sixteen strategic customer agreements represent $22 billion in cash and cash-equivalent commitments, including $18 billion in actual cash; most contracts extend through the end of 2030 and use take-or-pay structures.
3) The company increased its planned U.S. investment from $200 billion to $250 billion, but the timeline for supply to catch up remains unclear.
Samsung Electronics
1) Memory-capacity expansion is accelerating alongside SK hynix, with HBM and advanced DRAM as the principal demand anchors.
2) The company plans to introduce High-NA EUV from the A10-class 1-nanometer node, targeting commercial deployment around 2030 and increasing numerical aperture from 0.33 to 0.55; whether single-patterning can replace some multiple-exposure steps will determine the cost-benefit profile.
SK hynix
1) The company restarted construction of its second NAND fab in Dalian, planning to install equipment as early as November and begin mass production in the first half of next year, adding approximately 50,000 wafers of monthly capacity, or about 50% relative to the existing 100,000 wafers.
2) Its investment vehicle owns 14.19% of Kioxia, exceeding Toshiba’s 14.06% stake, but regulatory and conflict-of-interest constraints limit operational control.
Kioxia: NAND ownership has shifted, with an SK hynix-affiliated investment vehicle increasing its stake to 14.19%, equivalent to 77.4 million shares. A competitor becoming the largest shareholder may strengthen expectations for long-term collaboration, but will also amplify governance and antitrust concerns. In the near term, the focus should be on whether this affects procurement, capacity, and technology collaboration, as well as the boundaries between the two parties at the capital and operating levels.
Sandisk: Cantor maintained its Overweight rating and $2,900 price target. The analyst-day focus is whether long-term agreements can keep NAND supply tight through at least 2028 and whether the new contract model can stabilize margins. More important validation points remain enterprise SSD orders, average selling prices, and free cash flow; whether tight supply and demand can translate into sustained profit improvement has yet to be confirmed by quarterly data.
"We expect these exceptionally tight industry conditions to persist beyond 2027. Based on these signs of strengthening customer demand, we now expect 2027 to be even tighter than 2026"
