Core Viewpoint - Google's recent launch of the TurboQuant compression algorithm, which claims to reduce the key cache usage of large models by at least six times, has caused panic in the storage industry. However, at the MemoryS 2026 summit, various storage manufacturers and cloud computing companies expressed that the demand for storage is still expected to rise due to the acceleration of AI applications, leading to a potential continuation of shortages [3][4]. Group 1: AI Impact on Storage Demand - AI is rapidly consuming storage capacity, with AI servers projected to account for over 20% of total server shipments by 2026, significantly increasing storage configurations [4]. - The shift from AI model training to more frequent actual usage has heightened the demand for data reading speed and response capabilities, making high-performance storage a core component for system efficiency [4]. - The automotive sector is emerging as a significant application area for AI, alongside edge AI, which is expected to drive new growth in storage demand [4]. Group 2: Supply-Side Challenges - The expansion cycle for storage capacity is lengthy, lasting 18 to 24 months, making it difficult to alleviate supply shortages in the short term. Structural mismatches in supply and demand have become the norm [5]. - Flash memory is expected to remain in short supply for an extended period, with feedback from system-side experts indicating increasing resource constraints due to AI growth squeezing DRAM capacity [6]. - The storage technology landscape is undergoing a paradigm shift, moving from micro-innovations to architectural revolutions, with concepts like CXL and integrated storage-compute solutions accelerating towards commercialization [6].

Core Viewpoint - The demand for AI is driving the storage industry, which is currently in the mid-stage of a super boom cycle, with supply shortages expected to last at least until 2027 [1] Group 1: Storage Industry Outlook - The storage industry is maintaining a high level of prosperity, supported by better-than-expected performance and guidance from key players like Kioxia, as well as an increase in NAND contract prices [1] - The industry is expected to remain in a supply-demand imbalance until the end of 2027, with a strong recommendation for storage module companies due to their short-term performance potential [1] Group 2: Investment Opportunities - The report highlights four key investment directions in the context of the AI era, focusing on the need for bandwidth and capacity upgrades [2] - Storage solution providers are essential for CUBE, with a focus on companies that have support from storage manufacturers and first-mover advantages [2] - Semiconductor equipment is benefiting from the upgrade in advanced packaging demands, with a focus on etching, bonding, and thinning equipment [2] - Advanced packaging is seen as a critical breakthrough for high-end storage, with Chinese manufacturers leading in capabilities and expanding capacity [2] - Logic chip companies are expected to enhance their competitiveness and accelerate industrialization, particularly in 3D structured logic chips, benefiting from AI-driven demand [2]

Summary of Conference Call Notes Company and Industry Overview - The discussion centers around the advancements in the AI inference market, particularly focusing on the Language Processing Unit (LPU) developed by Groq, which was acquired by NVIDIA for $20 billion in early 2026. This acquisition aims to address the rising demand for AI model utilization and Annual Recurring Revenue (ARR) growth trends [2][3]. Core Points and Arguments 1. **LPU Design and Functionality**: - LPU integrates a large amount of SRAM on-chip instead of relying on external HBM, significantly reducing data transmission distances and overcoming latency bottlenecks in AI inference, particularly in the Decode phase [1][2]. - The LPU is designed for extreme determinism, making it closer to a custom ASIC for specific models, while GPUs focus on generality and TPUs on matrix multiplication efficiency [1][4]. 2. **Market Trends**: - The inference market is experiencing a "PD separation" trend, where the Prefill phase is compute-intensive and suitable for high-performance CPUs, while the Decode phase is memory-intensive, benefiting from LPU's SRAM architecture to reduce KV Cache read latency [1][5]. 3. **NVIDIA's Storage Strategy**: - NVIDIA's storage layout is diversified, with SRAM targeting low-latency scenarios, HBM4 for high-performance training, GDDR for cost-effective computing, and SSDs exploring independent KV cache storage cabinets [1][6]. 4. **Groq's Industrialization**: - Groq's production is accelerating, with Samsung's foundry orders expected to increase from 9,000 units in 2025 to 15,000 units, indicating a significant ramp-up in LPU deployment [1][7]. 5. **LPU Limitations**: - The high cost of SRAM compared to DRAM poses challenges for large model inference, requiring multi-chip stacking, which increases initial costs. Additionally, the software stack's lack of flexibility necessitates integration with general-purpose GPUs for optimal performance [2][3]. Additional Important Insights 1. **Hardware Integration and Optimization**: - The core competitiveness of hardware manufacturers is shifting towards system-level optimization and integration of various hardware components, particularly in the inference domain [6]. 2. **Future Developments in Technology**: - The LPU's development is expected to drive changes in several areas: - **SRAM**: Advancements may include 3D stacking and hierarchical management [7][8]. - **PCB**: Innovations in chip packaging, such as backside power delivery designs, may lead to increased PCB layers or new materials [8]. - **Switch**: The need for high-speed interconnects within LPU systems may lead to new switch products and architectures [9]. - **Liquid Cooling**: As system integration and power consumption rise, liquid cooling solutions are anticipated to become a significant growth area [9]. This summary encapsulates the key points discussed in the conference call, highlighting the advancements and challenges in the AI inference market, particularly regarding the LPU technology and its implications for the industry.

Group 1: Optical Memory - The concept of optical memory is introduced, where Google attempts to remove HBM due to limited production capacity and set up a DRAM memory cabinet with pooling technology for dynamic memory allocation [2] - The advantages of this solution include releasing physical space and capacity limitations of TPU CoWoS, increasing the flexibility of DRAM allocation per TPU chip, and potentially doubling the allocation from 192GB to 1TB [3] - This approach challenges the long-standing "near-memory computing" principle in the semiconductor industry, which could lead to issues like memory walls and idle computing units if latency is too high [3][4] Group 2: Gas Turbines - GEV's financial report indicates strong demand in the gas turbine sector, with 41 new heavy-duty gas turbine orders, including 15 HA units, leading to a backlog increase of 7GW to 40GW [6] - The current booking prices for slot agreements are 10-20% higher than existing backlog orders, providing certainty for profit margin expansion in 2026 [6] - The industry logic for gas turbines and HRSG remains positive, indicating continued growth potential [8] Group 3: Memory Market - JP Morgan's analysis has raised expectations for Hynix and Samsung, noting that the demand for server memory driven by AI workloads is surging, offsetting weak demand from PCs and smartphones [9] - The memory price is entering a stronger and longer upward cycle, with HBM becoming a growth highlight, and HBM4 production is on track, capturing a significant share of orders from key clients like Nvidia [9][11] - Hynix plans to significantly increase capital expenditure in 2026 to address current memory supply shortages and lay the groundwork for long-term growth, while maintaining disciplined spending [9][11]

Core Viewpoint - The investment focus in the semiconductor industry is shifting from GPUs to storage chips, leading to a new wave of IPO activities in the sector. Group 1: Company Developments - Unigroup Guoxin has entered the counseling period for its public offering on the Beijing Stock Exchange, with the application accepted by the Shaanxi Securities Regulatory Bureau [1] - Changxin Technology, the largest DRAM manufacturer in China, has had its IPO application accepted for the Sci-Tech Innovation Board [1] - Several other leading companies in niche segments are also in the process of going public or have queued for listings [1] Group 2: Product Offerings - Unigroup Guoxin's memory chip products include various standards such as SDR, DDR, and LPDDR, with over twenty products achieving global mass production and sales [2] - The company has developed 3D stacked DRAM (SeDRAM) technology, offering innovative storage solutions for near-memory computing, AI, and high-performance computing [2] Group 3: Financial Performance - In the first half of 2025, Unigroup Guoxin achieved revenue of 750 million yuan, a year-on-year increase of 38.64%, and a net profit of 5.683 million yuan, marking a 139.54% increase and a return to profitability [4] Group 4: Market Trends - The DRAM market is experiencing a price and volume increase, with significant competition from international giants like Samsung, SK Hynix, and Micron, which together hold over 90% of the global market share [8] - The demand for storage chips is rapidly growing in sectors such as automotive electronics, smart driving, and IoT devices, particularly for automotive-grade DRAM [8] - The rise of AI servers is driving demand for high-bandwidth, large-capacity DRAM, leading to a shift towards high-performance, low-power, and low-latency storage solutions [9] Group 5: Technological Innovations - Domestic manufacturers are exploring new technologies to close the gap with international leaders, with potential for breakthroughs in product and technology innovation [10]

Summary of Conference Call on NVIDIA's Acquisition of Groq and 3D Chip Technology Industry and Company Involved - **Company**: NVIDIA - **Acquired Company**: Groq - **Industry**: AI Chip Technology and Computing Key Points and Arguments NVIDIA's Acquisition of Groq - NVIDIA acquired Groq for $20 billion, focusing on Groq's physical assets without acquiring its intellectual property, allowing non-exclusive use of Groq's architecture [2] - The acquisition signifies NVIDIA's recognition of the differences between inference and training, indicating a shift towards specialized chip planning for inference [2] Groq's LPU Architecture - Groq's LPU (TensorStream Processor) architecture is designed specifically for inference, offering advantages such as low latency, deterministic execution time, high user concurrency, and extremely high bandwidth [1][3] - The LPU can achieve a bandwidth of 80TB/s, significantly outperforming the latest Blackwell B300 GPU's 8TB/s bandwidth, especially in large language model tasks [4] - However, the LPU has limitations, including high deployment costs and programming complexity, requiring manual pipeline arrangement for optimal performance [4] Integration with Existing Ecosystem - NVIDIA plans to maintain the CUDA ecosystem's universality while integrating LPU through NVFusion, ensuring software platform consistency [5][6] - The long-term goal is to achieve collaborative design at the architecture and compiler levels to meet high-performance requirements in inference scenarios [7] Domestic 3D Chip Development - Domestic companies like CloudWalk are actively developing 3D chips to significantly reduce total cost of ownership (TCO), particularly in single token costs [1][11] - The 3D DM solution offers greater capacity than SRAM and comparable bandwidth, but requires 2-3 years for large-scale deployment due to maturity issues [1][8] - 3D RAM is expected to support large model operations effectively, with applications in edge computing and cloud inference [10] Challenges and Bottlenecks - Key bottlenecks for 3D DL M deployment include yield rates and thermal management, with advanced stacking methods potentially reducing overall yield [8] - The development of 3D chip technology in China is progressing, with several companies in the early stages of research and testing, but large-scale production is still 2+ years away [9] Future Market Trends - The 3D architecture is projected to capture about 30% of the inference market, driven by the need for diverse computing capabilities [16] - The demand for low-cost solutions will accelerate the adoption of diverse architectures in inference, with 3D RAM being a significant component [20] - Domestic advancements in 3D technology may outpace international developments due to strong market demand and government support for AI applications [19][20] Customer Sentiment and Adoption - Customers are showing positive attitudes towards 3D chip solutions, particularly in smaller edge scenarios like AI PCs and mobile devices, with broader commercial adoption expected in 2-3 years [12][13] Conclusion - The integration of 3D technology represents a viable path for domestic companies to close the gap with international standards in inference capabilities, with a focus on reducing costs and enhancing performance [19][20]

Group 1: Capital Market Trends - In 2026, the capital market liquidity environment will feature three main characteristics: strategic stabilization forces represented by the Central Huijin, the optimization of capital market funding structure, and an increase in stock supply driven by mergers and acquisitions [1] - The proportion of institutional funds, represented by insurance and public funds, is expected to rise further, contributing to a more balanced investment and financing structure in the capital market [1] - The role of liquidity in driving unilateral valuation changes will diminish as policies aim to enhance market resilience and promote equity financing [1] Group 2: Phosphate Rock Market - The price of phosphate rock remains high due to a sustained price surge in chemical products, with market averages reported at 1016 CNY/ton for 30% grade, 945 CNY/ton for 28% grade, and 758 CNY/ton for 25% grade [3] - Companies like Batian Co. and Xingfa Group are actively disclosing advancements in phosphate resource acquisition and capacity integration, indicating an expansion in phosphate rock production capacity [3] - The current resource layout trend is driven by the industry's expectation of a "tight balance" in phosphate rock prices in the short term, benefiting companies with a complete industrial chain layout [3] Group 3: Low-altitude Economy - The low-altitude economy is projected to become a trillion-yuan market by 2030, with the market size exceeding 500 billion CNY in 2023 [4] - The growth is supported by expanding downstream application areas, with equipment value accounting for 5-15% of operating income and infrastructure accounting for 20% [4] - The focus for 2024-2030 will be on enhancing supply capacity and industrial innovation in general aviation equipment, with a strong emphasis on technological innovation and product development [4] Group 4: Insurance Industry Outlook - The insurance industry is transitioning from a narrative of balance sheet recession to positive expansion, with a strong upward trend expected to continue into 2026 [6] - Key indicators include rapid growth in net assets, increased sales of dividend insurance, and significant growth potential in the insurance distribution channel [6] - The focus for stock selection will be on companies with high policy value rates, fast new business value growth, and stable profit and dividend growth [6] Group 5: AI and Semiconductor Industry - The AI era emphasizes memory bandwidth and capacity upgrades, with a trend towards integrated storage and computing [7] - Investment opportunities are identified in four key areas: storage solution providers, semiconductor equipment, advanced packaging, and logic chip companies [7] - Companies with support from original storage manufacturers and those focusing on ultra-thin LPDDR stacking solutions are highlighted as potential leaders in the market [7]

Core Viewpoint - The report from CITIC Securities emphasizes that in the AI era, the upgrade of memory bandwidth and capacity is crucial, with integrated storage and computing being a trend, indicating a high level of activity in the sector and highlighting investment opportunities [1] Group 1: Storage Solutions Manufacturers - Focus on manufacturers that provide necessary CUBE support and customized design solutions to aid industrialization, aiming to penetrate the high-end market, particularly those with backing from original storage manufacturers and first-mover advantages [1] - Attention is also directed towards companies developing ultra-thin LPDDR stacking solutions [1] Group 2: Semiconductor Equipment - The industry is benefiting from the upgrade in advanced packaging and testing demands, alongside process optimization and yield improvement, which accelerates the localization of the supply chain [1] - Key areas of focus include etching, bonding, and thinning equipment due to the advancements in packaging and localization [1] Group 3: Advanced Packaging - Advanced packaging is identified as a critical breakthrough point for high-end storage, with high availability of equipment [1] - Chinese manufacturers are noted to have leading capabilities in advanced packaging and are expanding their production capacity [1] Group 4: Logic Chip Companies - The report also highlights the importance of logic chip companies in the context of the evolving semiconductor landscape [1]

Core Viewpoint - The upgrade of memory bandwidth and capacity is central in the AI era, with integrated storage and computing being a trend, leading to high demand in near-storage computing, presenting investment opportunities [1] Group 1: Investment Recommendations - Focus on four key areas within the domestic HBM and CUBE-related industrial chain: 1) Storage solution manufacturers are essential for CUBE support, with customized design solutions aiding industrialization and entry into high-end markets, particularly those with original storage manufacturer support and first-mover advantages [1] 2) Semiconductor equipment will benefit from the upgrade in advanced packaging and testing demands, alongside process optimization and yield improvement, accelerating the localization of the supply chain; focus on etching, bonding, and thinning equipment [1] 3) Advanced packaging is a critical breakthrough point for high-end storage, with high equipment availability, and mainland Chinese manufacturers leading in advanced packaging capabilities and expanding production capacity [1] 4) Logic chip companies are also highlighted as a focus area [1]

Group 1 - The recent significant increase in prices of DDR4/LPDDR4X memory chips is driven by the surge in demand from AI and supply constraints due to production cuts by manufacturers [1] - The rise of AI large models is pushing the storage sector to the forefront of technological challenges, emphasizing the need for higher transmission speeds, data storage capacity, and specifications [1] - Chip technology stocks have shown strong performance, with notable increases in various semiconductor ETFs and stocks [1] Group 2 - Key technologies for large-scale AI computing include advanced packaging multi-chip interconnect technology, advanced process foundry, and near-memory computing [2] - Multi-chip interconnect is crucial for expanding AI computing power, as traditional copper interconnect faces challenges in high-frequency and high-speed transmission scenarios [4] - NVIDIA highlighted the importance of data centers in the AI era, focusing on network technologies that combine multiple GPUs into a super-scale GPU [4] Group 3 - NVIDIA's upcoming products, such as the Spectrum-X Photonics Ethernet switch and Quantum-X switch, aim to eliminate bottlenecks in traditional architectures, providing high performance and energy efficiency for modern AI factories [5] - Near-memory computing technologies, represented by HBM, are essential for AI chips, with advancements expected from HBM3E to HBM4 by 2026-2027 [5] - The industry is addressing computing power challenges from multiple dimensions, with new server models being launched to meet the growing demand for AI capabilities [8] Group 4 - Storage performance is critical for maximizing GPU efficiency, and while solutions for previous bottlenecks exist, there remains a significant cost gap between NAND and HDD technologies [9] - The semiconductor industry is focusing on technological upgrades and domestic breakthroughs in storage, which are positively impacting the secondary market [9]