Core Viewpoint - Elon Musk's TeraFab project aims to produce millions to billions of AI chips with an annual power consumption of up to 1 terawatt (1 TW), requiring an estimated $5 trillion in funding to achieve its goals, which far exceeds current industry capacity [1][5]. Group 1: Funding and Production Capacity - TeraFab's goal of producing 1 TW of AI silicon annually necessitates between 142 to 358 wafer fabs to process 22.4 million Rubin Ultra GPU wafers, 2.716 million Vera CPU wafers, and 15.824 million HBM4E wafers [1]. - A modern advanced logic wafer fab can produce approximately 24 million wafers per year, meaning TeraFab would need about 105 fabs at 100% yield or 126 fabs at 80% yield to meet its production targets [3]. - The estimated cost for a 2nm process fab ranges from $25 billion to $35 billion, leading to a total requirement of approximately $3.15 trillion at 100% yield or $3.78 trillion at 80% yield for logic capacity alone [3]. Group 2: High Bandwidth Memory (HBM) Production - HBM production is critical for TeraFab's objectives, with modern DRAM fabs providing a capacity of 100,000 to 200,000 wafers per minute, averaging 150,000 wafers [4]. - To produce 15.824 million HBM4E wafers, TeraFab would require about 9 fabs at 100% yield or 12 fabs at 70% yield, with each fab costing at least $20 billion, leading to a total of approximately $240 billion for memory capacity [4]. - Advanced packaging facilities for 2.5D and 3D integration are also necessary, with costs ranging from $2 billion to $3.5 billion per facility, indicating a need for significant additional investment [4]. Group 3: Challenges Beyond Funding - Raising $5 trillion poses significant challenges, as it exceeds the market capitalizations of major companies like Nvidia, Apple, and Alphabet combined [5]. - The feasibility of such large-scale private financing or collaboration among governments, sovereign wealth funds, and capital markets is questioned, alongside limitations in manufacturing equipment and skilled labor availability [5]. - The ultimate question remains whether Musk intends to establish a chip foundry with capacity surpassing that of TSMC, Samsung, and Intel combined to meet the demands of Tesla, SpaceX, and xAI [5].

Core Insights - Nvidia has established itself as a dominant supplier in the GenAI revolution, showcasing a clear roadmap for its hardware and software developments in the AI sector [2][3] - The 2023 roadmap reveals Nvidia's annual update plan for its AI system components, with products like GX200 and Rubin R200 GPU accelerators set for release by 2025 [3][4] - Nvidia's market share in AI computing remains substantial, with projections indicating that the company will capture a significant portion of the server market revenue by 2025 [5] Roadmap Developments - The 2023 roadmap marks the first detailed annual update plan for Nvidia's AI systems, including products like Blackwell GPUs and Vera Arm server CPUs [3][4] - Nvidia's 2026 roadmap includes advancements in GPU technology, with the introduction of the "Feynman Ultra" GPU and updates to the ConnectX-10 SmartNIC [4][6] - The roadmap emphasizes the importance of these developments for OEMs and ODMs, as they are crucial for the deployment of AI training and inference systems [4][5] Market Projections - The server market is projected to reach between $420 billion and $450 billion by 2025, with Nvidia expected to generate approximately $190 billion from system material costs [5] - Machines equipped with Nvidia GPUs are anticipated to generate revenues between $275 billion and $325 billion, indicating a market share of 61% to 77% for Nvidia technology [5] - The profitability of AI systems is heavily skewed towards Nvidia, as evidenced by its gross, operating, and net profit margins [5] Technical Specifications - The Rubin R200 GPU is designed to deliver 50 petaflops of FP4 performance, significantly outperforming previous models [9] - The upcoming "Rubin Ultra" GPU is expected to double the GPU chip count and achieve 100 petaflops of FP4 performance, with advanced memory capabilities [16][19] - Nvidia's NVLink technology is set to evolve, with NVLink 6 offering 3,600 GB/sec bandwidth and NVLink 7 projected to reach 7,200 GB/sec [18][21] Future Innovations - Nvidia plans to introduce the "Kyber" rack, which will support a higher number of GPU slots and enhance overall system performance [16][21] - The integration of advanced memory technologies and chip stacking in future products like the Feynman GPU is expected to significantly boost throughput [23] - The roadmap indicates a strategic focus on optimizing both copper and optical interconnects to enhance system efficiency and performance [22][20]

Investment Rating - The industry investment rating is "Overweight" (maintained) [1] Core Insights - The report highlights the acceleration of AI terminal demand driven by OpenClaw, with significant increases in inference computing power requirements [3] - Groq 3 LPU has exceeded expectations with a doubling of inference performance and an earlier-than-expected production ramp-up [4] - The integration of Groq 3 LPU into the Groq 3 LPX rack enhances overall AI computing capabilities, providing substantial improvements in throughput and power efficiency [4] - The Rubin Ultra GPU has achieved significant advancements in memory capacity and computing power, supporting ultra-large model inference [5] Summary by Sections Section: Market Trends - The report emphasizes the growing demand for AI computing power and the shift towards edge AI solutions, indicating a robust market outlook [3] Section: Product Developments - Groq 3 LPU integrates 500MB SRAM and offers 150TB/s bandwidth, significantly surpassing HBM capabilities, tailored for bandwidth-sensitive AI decoding needs [4] - Rubin Ultra GPU features up to 1TB HBM4e memory and 100 PFLOPS inference computing power, marking a substantial leap in performance [5] Section: Investment Recommendations - The report suggests focusing on three main lines: computing power, interconnectivity, and thermal management, with specific beneficiary stocks identified in PCB, CCL, and assembly segments [6]

Core Viewpoint - The A-share market has rebounded after a two-month consolidation, with the AI computing power industry chain experiencing significant growth, particularly in liquid cooling technology, which is expected to see substantial market expansion by 2026 [2][14]. Group 1: AI Computing Power and Liquid Cooling Technology - The stock price of CPO leader Xinyi Sheng reached a historical high of 466.66 yuan, marking a tenfold increase from its lowest price of 46.56 yuan on April 9 [2]. - Liquid cooling technology is becoming a trend in the cooling sector due to its advantages over traditional air cooling, including lower energy consumption and noise, as well as improved cooling efficiency [3][14]. - Google’s TPU v7 chip has a power consumption of 980W, necessitating the use of liquid cooling systems, which will increase the value of these systems [3][7]. Group 2: Market Growth and Projections - The liquid cooling market is projected to reach a scale of 24-29 billion USD by 2026, driven by the expected shipment of 2.2-2.3 million Google TPU v7 chips [7]. - The Chinese liquid cooling server market is expected to grow to 2.37 billion USD in 2024, a 67% increase from 2023, with a compound annual growth rate of 47.6% from 2023 to 2028 [14]. - The penetration rate of liquid cooling in servers is currently at 5%, indicating significant growth potential in the coming years [14]. Group 3: Company Performance and Stock Insights - A-share liquid cooling concept stocks have shown strong performance, with companies like Hongfuhuan and Yidong Electronics seeing over 40% growth in the last three months [16]. - Hongfuhuan focuses on liquid cooling products for networking and servers, having established partnerships with major domestic and international clients [16]. - Yidong Electronics has a strong integrated advantage in the liquid cooling sector, having achieved mass production of AI chip cooling components [16].

Core Viewpoint - Lens Technology (06613) has seen a stock increase of over 3%, currently at 25.7 HKD, with a trading volume of 52.058 million HKD, following the announcement of a proposed acquisition of 100% equity in Peimei Gao International, which holds a 95.1164% stake in Yuan Shi Technology [1] Company Summary - Lens Technology has signed a letter of intent with an independent third party to acquire Peimei Gao International, which is involved in the production and sales of server cabinets and liquid cooling modules [1] - The acquisition is expected to be completed next year and will help Lens Technology enter the NVIDIA AI server supply chain, significantly boosting its AI server business scale [1] - Credit Lyonnais has reiterated a "outperform" rating for Lens Technology, maintaining a target price of 38 HKD [1] Industry Summary - Yongxing Securities believes that liquid cooling is likely to become a trend in the industry due to increasing chip power consumption, with the TDP of GB300 expected to rise to 1400W and NVIDIA's next-generation Rubin Ultra GPU potentially reaching 2300W [1] - According to ASHRAE recommendations, liquid cooling technology is advised when chip TDP exceeds 300W and cabinet power density exceeds 40kW [1] - The growing energy consumption requirements in data centers are also driving the development of the liquid cooling industry [1]

Core Insights - The rapid growth of AI applications like ChatGPT and Claude is driving an exponential increase in power demand for global AI data centers, pushing them towards critical power limits [1] - The power consumption of AI data centers is shifting from traditional levels of 20-30 kW per rack to levels reaching 500 kW and even 1 MW [1][2] - NVIDIA has announced the formation of an 800V HVDC power supply alliance aimed at developing next-generation AI data centers capable of supporting 1 MW per rack by 2027 [4] Group 1: Power Demand and Infrastructure - AI workloads are causing data center power demands to surge, with traditional 54V power systems becoming inadequate for modern AI factories that require megawatt-level power [2] - The transition to 800V HVDC systems is seen as essential to reduce energy losses and improve overall efficiency in data centers [1][3] - The current reliance on 54V systems is leading to physical limitations in space and efficiency, necessitating a shift to higher voltage systems [2][3] Group 2: Technological Developments - The 800V HVDC architecture is expected to enhance end-to-end energy efficiency by up to 5% and significantly reduce maintenance costs by up to 70% [5] - NVIDIA's collaboration with partners across the energy ecosystem aims to overcome previous barriers to the widespread adoption of HVDC technology in data centers [4] - Domestic companies like InnoSilicon and Changdian Technology are also advancing their technologies to align with the 800V HVDC trend, indicating a competitive landscape [6][7] Group 3: Semiconductor Innovations - The global supply of Gallium Nitride (GaN) is becoming increasingly strained, with companies like InnoSilicon positioned to leverage this scarcity in the context of NVIDIA's supply chain [9] - GaN devices offer superior performance in high-voltage applications compared to traditional silicon-based semiconductors, making them ideal for the evolving demands of AI data centers [11][12] - The integration of GaN technology is expected to significantly enhance power density and efficiency in the new 800V HVDC systems [12]

Core Viewpoint - TSMC's CoPoS packaging technology mass production timeline is delayed from 2027 to 2029-2030 due to technical challenges, which may influence NVIDIA's plans for its Rubin Ultra GPU and shift focus to multi-chip module architecture [1] Group 1: TSMC's CoPoS Technology - TSMC's CoPoS (chip-on-panel-on-substrate) technology aims to enhance area utilization through larger panel sizes (e.g., 310x310mm) to meet AI GPU demands from clients like NVIDIA [1] - The delay in CoPoS mass production is attributed to immaturity in technology, particularly in managing panel and wafer discrepancies, larger area warpage control, and additional redistribution layers (RDL) [1] Group 2: Impact on AI Industry - Nomura's analysis suggests that TSMC may redirect its 2026 chip backend capital expenditures towards other technologies such as WMCM and SoIC, with CoWoS capacity allocation becoming a critical monitoring point [1] - The postponement of CoPoS could lead NVIDIA to adopt a multi-chip module architecture similar to Amazon's Trainium 2 design for its 2027 product launch [1]

Core Viewpoint - Nomura indicates that TSMC's CoPoS packaging technology mass production timeline may be delayed from the original plan of 2027 to 2029-2030, potentially forcing NVIDIA to shift its chip design strategy for the Rubin Ultra GPU to an MCM architecture to avoid limitations of single-module packaging [2][3][4]. Group 1: TSMC's CoPoS Technology Delay - TSMC's CoPoS (chip-on-panel-on-substrate) technology aims to enhance area utilization through larger panel sizes (e.g., 310x310mm) to meet AI GPU demands [4]. - The delay in CoPoS mass production is attributed to technical challenges, particularly in managing panel and wafer discrepancies, warpage control, and additional redistribution layers (RDL) [4][5]. - The expected mass production timeline has shifted from 2027 to potentially late 2029 [4][5]. Group 2: Impact on NVIDIA's Product Strategy - The delay in CoPoS may compel NVIDIA to adopt an MCM architecture for the Rubin Ultra GPU, distributing four Rubin GPUs across two modules connected via a substrate [5][6]. - This adjustment is similar to Amazon's AWS Trainium 2 design, which utilizes CoWoS-R and MCM to integrate computing chips and HBM on a single substrate [6]. - While this change may help NVIDIA mitigate delays, it could also increase design complexity and costs [6]. Group 3: TSMC's Capital Expenditure Adjustments - TSMC's capital expenditure allocation may shift towards wafer-level multi-chip modules (WMCM) and system-on-chip (SoIC) technologies due to the CoPoS delay [7]. - Nomura maintains its forecast for TSMC's CoWoS capacity, expecting monthly wafer production to reach 70,000 and 90,000-100,000 by the end of 2025 and 2026, respectively [7]. - The report warns that market expectations for WMCM may be overly optimistic, while those for SoIC are more conservative [8].

Core Viewpoint - TSMC's CoPoS packaging technology mass production is likely delayed until 2029-2030, which may force NVIDIA to adjust its chip design strategy towards alternative architectures [1][2][3] Group 1: TSMC's CoPoS Technology Delay - TSMC's CoPoS technology, originally scheduled for mass production in 2027, is now expected to be delayed until the second half of 2029 due to technical challenges [2][3] - Key challenges include managing differences between panels and wafers, controlling warpage over larger areas, and addressing more redistribution layers (RDL) [2] Group 2: Impact on NVIDIA's Product Strategy - NVIDIA's Rubin Ultra GPU, initially requiring up to eight wafer-sized CoWoS-L interconnects, may need to shift to a multi-chip module (MCM) architecture due to the CoPoS delay [3] - This adjustment is similar to Amazon's Trainium 2 design, which utilizes CoWoS-R and MCM to integrate computing chips and HBM on a single substrate [3] Group 3: TSMC's Capital Expenditure Adjustments - TSMC's capital expenditure for the latter half of 2026 may increasingly focus on wafer-level multi-chip modules (WMCM) and system-on-chip (SoIC) technologies due to the CoPoS delay [4][5] - The report maintains forecasts for TSMC's CoWoS capacity, expecting monthly wafer production to reach 70,000 and 90,000-100,000 by the end of 2025 and 2026, respectively [4]

Core Insights - Nvidia is redefining the characteristics and functionalities of future power electronic devices, particularly for AI data centers, by designing a new powertrain architecture [1][4] - The shift towards 800V high voltage direct current (HVDC) data center infrastructure is being supported by various semiconductor suppliers and power system component manufacturers [1][5] Group 1: Nvidia's Influence on Power Electronics - Nvidia's push for AI data centers is creating momentum for Gallium Nitride (GaN) technology, similar to the impact of Silicon Carbide (SiC) during Tesla's rise [2] - Nvidia is collaborating with multiple partners, including Infineon, MPS, Navitas, and others, to transition to 800V HVDC systems [1][4] Group 2: Technical Requirements and Innovations - The new 800V HVDC architecture will necessitate a range of new power devices and semiconductors to meet the demands of AI data centers [5] - Infineon is developing converters to demonstrate the advantages of 800V to lower voltages, focusing on power density and efficiency [6][8] Group 3: Competitive Landscape - Other companies, such as Navitas Semiconductor, are also capitalizing on Nvidia's drive for AI data centers by leveraging their expertise in GaN technology [13] - The competition is intensifying as companies like Infineon and Navitas seek to provide solutions for Nvidia's evolving power infrastructure needs [13][14] Group 4: Market Predictions - Yole Group predicts that GaN will experience faster growth than SiC in the AI data center market, with GaN devices having higher voltage potential [16] - The shift in Nvidia's power infrastructure strategy may render existing open computing projects obsolete, leading to a fragmented market [15]