404K SEMI-AI Evening Briefing 2026-07-30 — Diverging Compute Monetization: Cloud Demand Accelerates, While Memory and Testing Continue to Deliver
目录
Pre-Market Highlights
Full AI/Semiconductor Supply Chain
AI Models/Applications and Capital Expenditure
CSP/Cloud Capital Expenditure
GPU/CPU/ASIC
HBM/DRAM/NAND/SSD/HDD
Foundry and Advanced Packaging
Semiconductor Equipment/Testing
Optical Communications/Optics Supply Chain
Internet/Platforms
Software/SaaS
Consumer Electronics / Smart Vehicles
Overview
404K | 2026-07-30
Pre-Market Highlights
Demand for AI infrastructure continues to materialize, but revenue, orders, and capital efficiency are beginning to determine the divergence between companies. Microsoft is translating improvements in resource-pool efficiency directly into Azure revenue, while Amphenol, Seagate Technology, and the testing-equipment supply chain are providing hard evidence through orders, pricing, or market share. Simply increasing capital expenditure is no longer sufficient to justify valuations; revenue and cash flow must keep pace with equipment investment.
Memory pricing signals remain stronger than in most technology subsectors. Samsung Electronics reported sharp increases in DRAM and NAND selling prices; Seagate Technology is benefiting from both HDD price increases and HAMR-driven cost reductions; and Teradyne and Advantest further validated rising testing intensity for AI processors and HBM. The next question is whether supply constraints can continue translating into profit rather than merely orders. The hardware supply chain’s focus has shifted to pricing and delivery cadence.
The software and end-device supply chains are more fragmented. Twilio’s voice AI is beginning to move from pilots into production, while Datadog’s major-customer renewals remain stable. Foldable-device manufacturers are shifting the competitive focus toward creases and materials, while Tesla is supplementing its charging network through long-term power-purchase agreements. End demand still has bright spots, but visibility is weaker than in cloud, memory, and testing; in the near term, the key is whether inquiries convert into renewals, shipments, and utilization.
Full AI/Semiconductor Supply Chain
AI Models/Applications and Capital Expenditure
Modular Data Centers
1) More than 61 GW of capacity has been tracked, across more than 1,000 sites using modular or prefabricated strategies; these sites are expected to represent more than 30% of commissioned capacity by the end of 2028.
2) Full modularization can shorten the construction period for a 50 MW liquid-cooled data hall from 18–24 months to 12–18 months and reduce on-site mechanical and electrical labor hours by 63%. All-in cost declines only from approximately $14.6 million per MW to $13.5 million per MW, meaning the improvement in speed is greater than the improvement in cost; permitting, power supply, GPU delivery, and 3–8 months of commissioning remain constraints.
“Modularization significantly reduces on-site labor and improves speed, but cannot eliminate permitting, grid interconnection, GPU delivery, or on-site commissioning.”
CSP/Cloud Capital Expenditure
Data-Center Module Supply Chain
Modularization shifts more value from on-site contracting to factory manufacturing and systems integration. Vertiv’s estimated value per MW in a fully modular solution may rise from approximately $3.5 million to $7.0 million, with backlog of approximately $15 billion and a book-to-bill ratio of 2.9x. Comfort Systems’ modular business has approximately $2 billion in annualized revenue and $14.1 billion in backlog, up 73% year over year. However, lead times for some original equipment exceed 12 months, while transportation, lifting, and reliability issues can consume the time saved during construction.
Bloom Energy
Mizuho upgraded the company from Neutral to Outperform, arguing that on-site power solutions can connect data centers more quickly and have already been validated by major U.S. hyperscalers. Its financing capacity exceeds $27 billion, and backlog continues to grow faster than revenue. Progress on permitting in New Mexico provides an additional catalyst, while the risk is that service revenue declined 12% based on the revised quarterly run rate; operating leverage still requires validation through product shipments and project delivery.
GPU/CPU/ASIC
AMD
The MI455X Helios system uses 12 HBM4 stacks per GPU, providing 432 GB of capacity, versus 8 stacks and 288 GB for the competing solution. AMD trades single-pin speed of 6.5 Gbps for greater capacity, while the competing solution exceeds 11 Gbps. Samsung Electronics is reportedly the principal HBM4 supplier. The system has entered full production, with initial products scheduled for delivery in early Q3.
“AMD’s choice to use more HBM stacks in exchange for greater capacity reflects its wide-and-slow memory-design approach.”
HBM/DRAM/NAND/SSD/HDD
Samsung Electronics
1) Q2 DRAM bit growth was in the low teens, while average selling prices rose in the mid-40% range; Q3 bit-growth guidance is in the mid-single digits.
2) NAND bit growth was in the low-single digits, while average selling prices rose in the high-60% range; Q3 bit-growth guidance is in the high-single digits. Pricing contributed materially more than shipments, and whether profitability can continue improving depends on long-term contract pricing and capacity discipline.
Memory Pricing
Goldman Sachs expects conventional DRAM to achieve solid double-digit sequential price increases in both Q3 and Q4 2026. It also expects Samsung Electronics’ HBM prices to rise 87% year over year in 2027, above the consensus estimate of 52%. More than half of server DRAM is already covered by long-term agreements that include prepayments, take-or-pay provisions, or cancellation penalties, with supply tightness translating into stronger contractual constraints.
“Long-term agreements (LTAs): LTAs are providing strong discipline—covering more than half of server DRAM and supported by provisions including sizable prepayments, take-or-pay clauses, and cancellation penalties.”
Seagate Technology
1) Fiscal-quarter revenue was $3.629 billion, up 59.1% year over year and 16.6% sequentially; non-GAAP gross margin was 52.7%, operating margin was 44.6%, earnings per share were $5.71, and free cash flow was $1.118 billion.
2) For the next fiscal quarter, midpoint revenue guidance is $4.1 billion, operating-margin guidance is approximately 50%, and midpoint EPS guidance is $7.30. Capacity is largely allocated through 2027, with demand visibility extending into 2028.
3) HAMR represents approximately 40% of nearline shipment capacity and is expected to exceed 50% by year-end. Mozaic 4 has been qualified by 2 major cloud service providers, while selling prices per EB rose 6.4% sequentially.
“The report’s core thesis rests jointly on ‘price increases’ and ‘HAMR-driven cost reductions.’”
Kioxia
Ninth-generation memory has entered sample shipments, with mass production planned for 2027. Customer qualification, yield improvement, and capacity ramp-up are still required between sampling and mass production, making timely completion of validation milestones more important in the near term. If progress remains on schedule, advanced NAND iterations will affect unit costs, bit output per wafer, and competition in enterprise SSDs.
NAD Hybrid Memory Architecture
Seoul National University proposed tightly integrating DRAM and NAND flash, eliminating the I/O buffers and data buses used for communication between the two memory types to enable direct, parallel transfers. The paper states that data-transfer latency declined significantly but does not disclose comparative latency, bandwidth, power-consumption, or cost data; at this stage, it can only be viewed as one potential technical route for reducing data movement.
Foundry and Advanced Packaging
Samsung Electronics Foundry
Samsung Electronics reportedly plans to build a new AI-focused semiconductor fab in the United States and mass-produce 2 nm products. Whether advanced-node production can generate meaningful revenue still depends on customers, yields, and the production-ramp cadence. Local U.S. manufacturing helps bring production closer to customers, but depreciation and ramp-up costs for the new fab will also enter the income statement first. Subsequent validation points are customer onboarding, mass-production yields, and equipment move-in—not the fab plan itself.
UMC
1) The company has delivered its first mass-produced silicon-photonics wafers, indicating that mature-node foundries are beginning to enter volume production of photonic-electronic integration.
2) 2026 capital expenditure was raised from $1.5 billion to $2.0 billion for a cleanroom expansion in Singapore and a new fab in Tainan, with a construction period of approximately 20 months. The company plans to invest at least $5.0 billion over the next 3 years. Returns on investment will depend on whether AI-related orders can fill the additional capacity.
TSMC Advanced Packaging
The CoPoS roadmap is more likely to advance incrementally through “CoWoS-S, CoWoS-L, glass-carrier panel-level redistribution layers, glass cores, and glass interposers,” rather than replacing silicon and organic materials in a single step. The expected commercialization timeline is validation in 2026, low-volume production in 2027, and capacity expansion in 2028. The key validation points are sustained customer qualification and repeat equipment purchases.
“In the early stage, glass serves only as a temporary carrier providing flatness and dimensional stability. After RDL completion, the structure can still be transferred to an organic substrate and is not the final architecture.”
Hybrid-Bonding Equipment
Samsung Electronics’ 2030 mass-production plan calls for approximately 50 hybrid-bonding systems, but this figure represents the equipment required to meet an internal mass-production target, not confirmed orders or revenue. Actual demand will depend on adoption in next-generation HBM and logic chips. In the near term, investors should focus on process qualification, yields, repeat equipment purchases, and production-line construction rather than directly incorporating the long-term equipment count into earnings.
