目录
Pre-Market Takeaways
AI and Semiconductor Supply Chain
AI Models, Applications, and Capital Spending
CSP and Cloud Capital Spending, AI Cloud
GPU, CPU, and ASIC
HBM, DRAM, NAND, SSD, and HDD
Foundry, Advanced Packaging, and Test Equipment
Optical Communications and Components
Servers, Materials, and Data-Center Construction
Robotics, Autonomous Driving, and Space
Internet / Platforms
Software / SaaS
Consumer Electronics / Smart Vehicles
Overview
404K | 2026-08-14
Pre-Market Takeaways
AI demand remains strong, but the constraints are shifting. Applied Materials’ order visibility, Sandisk’s long-term contracts, Coherent’s InP expansion, and Quanta’s server capacity plans all point to continued customer demand for capacity. The more important question now is whether incremental revenue can convert into cash, rather than simply how fast shipments grow.
Divergence across the semiconductor supply chain is also becoming clearer. Memory, advanced packaging, substrates, and optical communications remain supply-constrained, prompting continued capacity expansion by equipment vendors and manufacturers. At the same time, elevated expectations have narrowed the margin for error after earnings. Applied Materials beat on both results and guidance but still fell in after-hours trading, signaling that valuations now require faster and more certain execution.
In software and devices, the core theme is AI’s transition from feature demonstrations to monetizable workflows. Claude is beginning to handle routine code maintenance, while AI PCs, robots, and Android POS systems continue to gain traction. However, model usage costs, higher memory prices, and working-capital demands across the manufacturing chain are emerging as the next tests for margins.
AI and Semiconductor Supply Chain
AI Models, Applications, and Capital Spending
Anthropic: Claude is moving beyond coding assistance into routine software maintenance. Test workflows cover crash investigations, consolidation of duplicate abstractions, and dead-code removal. Over several weeks, the system submitted 388 code changes, of which 180 were merged after model review and human approval. The key validation metrics are no longer limited to code-generation speed, but now include rework rates, human-review costs, and stability under continuous operation.
“Over the past few weeks, these routines submitted 388 PRs across our codebases, of which 180 were merged after Claude Code Review and human review.”
Zhipu AI: The GLM-5.3 base model remains unchanged at 743 billion parameters, with the improvement coming from expanded post-training. Its Terminal-Bench 3.0 score increased from 4.6 to 28.3, CyberGym from 77.2% to 84.5%, and ExploitBench from 24.4% to 54.4%. The gains are concentrated in long-horizon cybersecurity workflows; the next tests are success rates on real-world tasks and the model’s safety boundaries.
AI Capital Burden: Bridgewater co-CIO Greg Jensen estimates that the AI ecosystem will require $643 billion of capital in 2026 and potentially $1 trillion in 2027. He also believes OpenAI could exhaust its cash by early 2027 without additional financing. This is an external estimate, not company guidance, but together with working-capital pressure across the server manufacturing chain, it suggests that access to capital is becoming a second constraint on AI expansion.
CSP and Cloud Capital Spending, AI Cloud
U.S. Cloud Providers: Amazon, Google, Meta, Microsoft, and Oracle have disclosed more than $600 billion of combined capital expenditure for 2026, indicating that investment in physical AI infrastructure continues to rise. The risk has shifted from budget appetite to power availability, permitting, construction timelines, and supply-chain execution. Investors should focus on the rate at which announced spending is actually deployed, rather than higher unverified market estimates.
CoreWeave
1) Older GPUs have longer economic lives than conventional depreciation narratives imply. The company says customers still specifically request the 2020-vintage A100 and have signed fixed-price take-or-pay contracts extending through 2029; A100 pricing has remained firm since early 2025.
2) CoreWeave and Nvidia use the DSX platform for end-to-end validation of compute, networking, storage, and software. Vera Rubin NVL72 is described as delivering 10 times more tokens per second per megawatt than GB200 NVL72. The residual value of mature GPUs and the power-efficiency gains of new platforms need to be assessed separately.
“There is no single GPU that can dominate everything. In practice, there is a matrix matching workloads of different scales with GPUs of different scales.”
Cerebras: The company says data-center space remains an industry bottleneck and has spent the past 7 months securing resources and advancing construction. Built on TSMC’s 5-nanometer process, WSE-3 integrates 44GB of SRAM at the wafer level, bypassing HBM supply constraints. Its commercial AFD system assigns attention and prefill workloads to Blackwell and feed-forward networks to Cerebras, enabling models with more than 3 trillion parameters. The risk is that data-center delivery timelines continue to constrain revenue.
GPU, CPU, and ASIC
Nvidia
1) According to supply-chain reports, Feynman will use TSMC A16, SoIC 3D stacking, custom HBM, and CPO, with mass production targeted for 2028, further increasing demand for advanced process and packaging capacity.
2) Nvidia signed a multiyear agreement with Amkor Technology, committing $1.5 billion including prepayments to jointly develop AI-infrastructure packaging and testing and expand U.S. back-end capacity.
3) Spectrum-X Ethernet Photonics has entered full production, moving optical networking from the roadmap stage into production validation.
Intel
1) Intel increased its equity offering from $15 billion to $20 billion at an offering price of $95, with institutional demand exceeding $100 billion. The proceeds will fund capital expenditure.
2) The company disclosed that 18A yields are approximately 80%, Clearwater Forest is ramping, and EMIB back-end revenue is forecast at $1.1 billion in fiscal 2027 and $7.0 billion in fiscal 2028.
3) The equity dilution provides expansion capital; the next test is whether the foundry business can reach operating-profit breakeven by the end of 2027 as planned.
AMD: The company announced a large bond offering expected to raise up to $5 billion for advanced semiconductor development. The financing coincides with rising investment in AI-chip R&D;, showing that the product race has extended to the balance sheet. The real test is whether revenue from new products, supply security, and the interest burden justify the capital deployed—not the size of the financing itself.
ASIC Startups: Supply-chain reports indicate that some startups are struggling to secure sufficient HBM and CoWoS capacity because major cloud providers, Nvidia, and AMD have absorbed a substantial share of supply. Demand does not automatically translate into deliverable revenue: advanced-packaging and HBM allocations still determine whether new chips can reach volume production. Supply agreements are therefore more important to track than design announcements.
DeepX: Within 1 year of entering volume production, the DX-M1 edge NPU secured more than $13 million in commercial orders across more than 10 countries and 77 purchase orders, with more than 60% received in the past 4 months. Overseas markets account for more than 50% of revenue. The next-generation DX-M2 is expected to use Samsung Electronics’ 2-nanometer process, with samples planned for early customers in mid-2027 and a target of running generative AI at below 5 watts.
HBM, DRAM, NAND, SSD, and HDD
Micron
1) The company launched a $250 million Paradigm Fund, increasing its cumulative venture-capital commitments to $550 million across model architecture, computing, enterprise applications, and physical AI.
2) Fiscal third-quarter revenue reached $41.46 billion, up 346% year over year, and the company said demand for both DRAM and NAND materially exceeds supply.
3) Shortages provide earnings leverage but will also encourage longer-term capacity expansion. Capital returns, pricing, and the timing of new supply should be evaluated together.
Samsung Electronics
1) The company is considering converting Line 2 of its NRD-K R&D; fab into a foundry line, expanding 2-nanometer capacity for cHBM and HBM5.
2) Samsung is also recruiting visual-AI and data-engineering talent to deploy custom models in design, process optimization, anomaly detection, and manufacturing quality control.
3) Its Taylor fab has struggled to recruit engineers for 3-month or 6-month assignments, indicating that even after equipment is installed, shortages of experienced personnel could slow the yield ramp.
SK Hynix: The company approved $38.3 billion of investment in Yongin Y2 and Cheongju M17 to support advanced DRAM, HBM, and NAND. The first cleanrooms are not expected to open until December 2028 and June 2029, respectively. Current shortages and future supply additions can coexist; strong near-term pricing does not mean supply pressure will remain absent after 2029.
Sandisk
1) Fiscal 2028–2030 targets call for approximately 17.5% average annual revenue growth, an approximately 80% gross margin, an approximately 75% operating margin, and an approximately 50% adjusted free-cash-flow margin.
2) Eight long-term contracts have an average duration of more than 4 years and a floor-price value of approximately $94 billion, securing 50% and 67% of bit supply in fiscal 2027 and fiscal 2028, respectively.
3) HBF is scheduled to sample in 2027 and enter volume production in 2028, with 8 to 16 times the capacity of HBM. However, microsecond-level latency and limited write endurance constrain its use in training.
“Management has formally committed to returning 100% of remaining free cash flow, primarily through share repurchases, after completing reinvestment and maintaining a debt-free balance sheet.”
Phison: The company estimates that new NAND capacity requires at least 4 years from investment to volume production, rather than the 2-year assumption commonly used by the market. AI-related products accounted for 38% of second-quarter revenue. Phison is retaining strategic inventory to support enterprise SSD and AI customers, which could prolong tight supply, although management also warned that excessively high component gross margins would damage end demand.
Nanya Technology: The company is building a large 12-inch wafer fab and considering an additional site, with multiyear spending expected to exceed $10 billion for advanced 10-nanometer-class DRAM using EUV. Existing lines are operating at full capacity, while new capacity is expected to begin contributing materially from 2027 to 2028. Whether the expansion can keep pace with demand will depend on equipment delivery, yields, and capital efficiency.
