404K SEMI-AI Evening Brief 2026-07-28 — Expansion Shifts Gears: AI Hardware Demand Persists as Financing Costs and Supply Yields Enter the Validation Phase
目录
Pre-Market Highlights
Full AI/Semiconductor Value Chain
AI Models/Applications and Capital Expenditure
GPU/CPU/ASIC
HBM/DRAM/NAND/SSD/HDD
Foundry, Equipment, and Testing
Optical Communications, High-Speed Interconnects, and Supporting Infrastructure
Internet/Platforms
Software/SaaS
Consumer Electronics/Smart Vehicles
Overview
404K | 2026-07-28
Pre-Market Highlights
Rotation within technology stocks is accelerating, with semiconductors pressured by concerns over financing and supply expansion, while internet and software names are relatively stronger. The market is beginning to break down “how large is AI demand?” into more specific questions: who will provide the funding, whether equipment can ramp at acceptable yields, and whether customers can ultimately generate returns.
Hardware demand has not faded. GPUs, CPUs, memory, optical communications, and passive components are all showing new order or capacity signals, but bottlenecks are shifting from individual chips toward packaging, testing, lasers, and power infrastructure. Expanding supply also means pricing and margins can no longer be assessed solely through the shortage narrative.
Software is moving closer to validation through actual usage. The number of tokens processed by enterprises each month, AI security customers, model calls, and agent-design orders continue to grow. Counterevidence comes from falling prices for open-source models, lower customer switching costs, and the high sensitivity of cloud vendors’ returns on capital to utilization and pricing.
Full AI/Semiconductor Value Chain
AI Models/Applications and Capital Expenditure
OpenAI
1) NVIDIA reportedly plans to provide up to $250 billion in financing support for related projects, whose total investment could exceed $500 billion; chip financing under discussion could reach $350 billion.
2) These arrangements remain under negotiation and are not confirmed orders; financing responsibility is extending from chip vendors downstream, shifting the key validation point from GPU demand to project cash flow and credit spreads.
Meta
1) Meta and BlackRock have established a $14 billion joint venture to build a 1 GW data center, which is expected to begin operations in 2028.
2) BlackRock funds will own 80% and Meta 20%, with financing also including $12.5 billion of debt; Meta’s initial lease term is 4 years, renewable for up to 20 years, and it will provide an approximately $13 billion residual-value guarantee.
"Meta will initially be the campus’s sole user, using the capacity to support AI model development and upgrades to its core businesses."
Hyperscale cloud providers Morgan Stanley estimates that returns on capital for AI infrastructure investment are approximately 31%, with inference operations generating 25% to 46%; incremental margins for model APIs running on proprietary infrastructure could reach 75%. These figures depend on assumptions including 65% of capacity being used for inference, 2,750 tokens per GPU per second, and $1.75 per million tokens. Utilization is the most critical falsification variable.
Kimi Kimi K3 scored 57 in a model benchmark, narrowing the gap with the leading closed-source model to 4 points; Claude Opus 5, Claude Fable 5, and GPT-5.6 Sol scored 61, 60, and 59, respectively. Open-weight models continue to approach frontier capabilities, helping reduce software inference costs while making it harder for closed-source models to sustain pricing.
AT&T; The company consumes more than 1 trillion tokens per month, with usage maintaining double-digit growth. More than 100 generative AI models are already in production across customer service, fraud prevention, and cell-site selection. Enterprise AI is moving from experimentation into process redesign; the next question is whether back-office functions such as human resources and finance can generate measurable savings.
"We already have more than 100 generative AI models in production... customer service, fraud prevention, and determining the optimal locations for cellular base stations."
GPU/CPU/ASIC
NVIDIA
1) The company has reportedly become the effective tenant of a 1 GW data center in Texas, with a base lease value of $19.6 billion and a potential value of up to $50.2 billion including options.
2) Supply prices for consumer GPUs have increased for the 3rd time this year, with the latest increase at approximately 20% to 30%; 3GB of GDDR7 costs approximately $60 to $70, significantly above approximately $20 for the 2GB specification.
Intel Demand for Xeon 6 is stronger than expected and supply has tightened. The company is increasing Intel 3 wafer starts as well as back-end and substrate capacity. Agentic workloads simultaneously access models, databases, networks, and storage, with CPUs handling scheduling and data movement. The next questions are whether supply expansion can translate into server shipments and how customer inventories and lead times develop.
"GPUs perform AI model computation, while CPUs handle tasks including memory management and scheduling, networking, storage, API calls, and data transfers between models."
AMD The company expects the compute market to reach $2 trillion by 2030 and defines the CPU as the system orchestrator; the next-generation EPYC Venice uses the Zen 6 architecture and supports up to 256 cores and 512 threads. Separately, AI-assisted migration tools have reduced the software cost of switching from CUDA to AMD GPUs, and improvements in the ROCm ecosystem will directly affect the share of customers adopting mixed deployments.
Broadcom The Bailly CPO switch uses a 51.2T switch chip, while NVIDIA Spectrum-X CPO targets approximately 400Tb/s of system bandwidth; both have entered the initial volume-ramp phase. CPO can reduce connection power consumption by approximately 70%, but optical-engine yields, silicon-photonics capacity, and packaging still determine delivery speed. Platform specifications cannot yet be equated with revenue at scale.
Cerebras Cerebras and AMD are adopting a direct photodiode connection architecture in an effort to reduce the need for complex serializers and digital signal processing. Estimates suggest the related LPX system can use 256 LPUs, 128GB of SRAM, and 40PB/s of bandwidth as an alternative to large GPU clusters. Commercial success will depend on the software ecosystem, system reliability, and actual cost per token rather than peak bandwidth.
HBM/DRAM/NAND/SSD/HDD
Micron The phase 2 contract for its $2.75 billion packaging and testing facility in India has been awarded, with the expansion continuing to build out back-end capacity. Spot DDR5 16Gb prices rose 17% month over month and 890% year over year, but elevated prices will also encourage customers to reduce inventories and validate alternatives. Capacity ramps, yields, and the realization of contract pricing must be monitored together, while the direction of inventories must also be confirmed.
Samsung Electronics
1) The company plans to increase DRAM output at Hwaseong by approximately 15% by year-end, consistent with the regular annual increase of 10% to 15%; the larger supply risk comes from the possibility that the Yongin project could be brought forward from 2031 to 2029.
2) HBM5 is planned to use a 2-nanometer GAA base die, with a speed target more than 50% above HBM4E, but yield and packaging qualification have not yet been completed.
SK hynix LPDDR6 is scheduled for mass production in the second half, using 16Gb dies manufactured on the 1c process, and was qualified in March. The company is also benefiting from demand for high-end mobile memory and AI servers; a Vera Rubin rack contains approximately 20.7TB of HBM4. Key areas to watch are customer qualification of LPDDR6 and the pace of HBM4 supply.
Rambus Q2 product revenue increased by more than 20% year over year, with both revenue and non-GAAP earnings reaching record highs; Q3 product-revenue guidance continues to indicate approximately 20% growth. Server CPUs are upgrading toward 16-channel DDR5 architectures, while each MRDIMM contains approximately 4 times as many dies, increasing interface-chip content value but also raising the risk of supply constraints in 2027.
"The second quarter delivered outstanding results, with both revenue and non-GAAP earnings reaching all-time highs, driven by record product-revenue growth of more than 20% year over year."
Memory architecture A single HBM4 stack can provide up to 64GB of capacity and approximately 2TB/s of bandwidth, but requires approximately 3 times as many wafers as conventional DRAM for the same bit capacity. High margins and supply tightness lasting 2 to 3 years are attracting capacity expansion, while system designers are also exploring tiered architectures involving SRAM, 3D DRAM, SOCAMM2, and SSDs. Incremental demand and substitution pathways will jointly determine the sustainability of earnings.
Foundry, Equipment, and Testing
TSMC Q2 revenue was $40.2 billion, up 33.7% year over year, with a gross margin of 67.7% and an operating margin of 60.3%; 7-nanometer and below accounted for 77% of revenue, while 2-nanometer accounted for 3%. The company raised capital expenditure to $60 billion to $64 billion and is managing scarce resources through long-term capacity-allocation agreements. The strength of customer commitments is more important than spot-market bidding.
Samsung Foundry Its global foundry share was approximately 6.5% in Q1, versus approximately 72.3% for TSMC. Groq LP30 uses Samsung Electronics’ 4-nanometer process, Tesla AI6 is planned for 2-nanometer, and the company also offers 2-nanometer GAA and 2.5D packaging platforms. These design opportunities must still pass tape-out, yield, and mass-production validation; memoranda of understanding cannot be counted as firm orders.
ASML China’s domestically produced immersion DUV tools are planned to reach production of 5 units in 2026 and 20 units in 2027. They are currently reportedly targeted primarily at 28-nanometer and could potentially extend to 14-nanometer or 7-nanometer through multiple patterning. Overlay accuracy and throughput remain behind mature commercial equipment, while multiple patterning would increase costs by approximately 40% to 50%; the near-term impact is more likely to be validation pressure.
Advantest Automated test equipment orders recently disclosed by ASE totaled approximately $210 million, indicating that demand for AI-chip testing remains strong. As CPO, HBM, and advanced packaging become more complex, testing is extending from the wafer level to packaging and system-level testing. Order visibility is relatively high, but equipment lead times and the pace of customer capital expenditure still need to be verified; the speed at which orders convert into revenue is the next validation point.
SCREEN The company raised its forecast for net-profit growth in the fiscal year ending March 2027 to 25%, primarily driven by AI semiconductor-related businesses. The demand signal is consistent with increases in advanced-process and cleaning steps. The investment implication is that equipment momentum is not confined to lithography; the next questions are whether order conversion into revenue, utilization of new capacity, and margins can improve simultaneously to complete the earnings cycle.
