高盛在最新研报中指出，百度旗下芯片部门昆仑芯已正式递交香港上市申请，这标志着百度解锁资产价 值迈出了坚实的一步。 据追风交易台，1月3日，高盛发布最新研报认为，根据其核心测算显示，如果市场给予昆仑芯类似于寒 武纪的估值倍数（40倍市销率），百度持有的59%股权价值将高达220亿美元，惊人地相当于百度当前 总市值的45%。即便在保守估值下，这也代表着巨大的潜在上行空间。 昆仑芯分拆上市：解锁隐蔽价值的关键一步 百度已宣布，昆仑芯于2026年1月1日向香港联交所递交了上市申请。此次分拆将通过全球发售昆仑芯股 份的方式进行，包括香港公开发售以及向机构和专业投资者的配售。 对于投资者而言，最关键的信息在于股权结构的变化与控制权。高盛指出，截至2025年8月，百度持有 昆仑芯59%的控股权。在拟议的分拆完成后，百度预计昆仑芯将继续作为公司的子公司。这意味着百度 既能通过独立上市释放昆仑芯的市场价值，又能继续在其生态系统中保持对核心算力基础设施的控制。 外部销售占比： 高盛预测，进入2026年，昆仑芯的外部销售将迅速增长，占到云总销 售额的14%。 业务结构优化： 结合更具经常性和高附加值的订阅业务（据高盛估算接近云收入 ...

智通财经APP获悉，美国老牌芯片巨头英特尔(INTC.US)股价在周三美股盘前交易中走低，一度大跌超 5%。此前据两位知情人士向媒体透露，"AI芯片霸主"英伟达(NVDA.US)在实际评估了使用英特尔18A 先进制程技术如何制造其中一款AI芯片之后，并未继续向前推进，而是选择暂停对英特尔18A制程工艺 的测试。 全球两大AI芯片设计领军者(在芯片行业被称作fabless)——即专注于AI GPU算力体系的英伟达以及专注 于AI ASIC算力技术路线的博通(AVGO.US)早在三月初就有媒体曾曝出，这两大芯片设计领军者与英特 尔进行了高端芯片制造测试。 然而，自那以后，英特尔的18A芯片技术持续面临挑战，因为初期的那些客户测试结果不乐观导致了客 户们对于英特尔技术层面的不满，并且有可能继续将芯片代工/制造产能全线投入于英特尔在芯片制造 领域的最强竞争对手台积电(TSM.US)。 英特尔当前迫切需要大客户来撑起芯片代工业务 对于英特尔而言，掌握领先全球且极度适合量产的先进制程芯片制造技术可谓至关重要。这家美国老牌 芯片巨头甚至不惜购置更加昂贵的High-NA EUV光刻机——这也是阿斯麦所交付的全球第一台High ...

Core Insights - Google's TPU has gained significant recognition, leading to claims that it has surpassed NVIDIA in the market, prompting NVIDIA to assert its continued dominance in GPU technology [1] Market Share and Competition - Bank of America projects NVIDIA's market share will decline from approximately 85% to around 75% due to increased competition, particularly from Google's TPU, although this shift will occur gradually [2][6] - The current supply chain constraints and NVIDIA's scale advantages hinder competitors from rapidly capturing market share [7] AI Model Development - The competition in large language models (LLMs) is described as a long-term marathon, with recent releases from Google and Anthropic indicating a dynamic landscape [4] - Google's TPU has been integral to the training of its Gemini models, with plans to potentially lease TPU to Meta by 2026, which could intensify competition for existing GPU suppliers [4] Company Valuations - AMD's target price is set at $300, reflecting anticipated growth in AI and CPU market share [9] - Broadcom's target price is $400, supported by strong earnings growth and profitability in the semiconductor sector [10] - NVIDIA's target price is $275, justified by its leading position in the rapidly growing AI computing market [11]

Group 1 - Intel's CEO Lip-Bu Tan strongly denied rumors regarding a new employee leaking trade secrets from TSMC, emphasizing respect for intellectual property [1] - TSMC's market capitalization exceeds $1.15 trillion in Taiwan and $1.5 trillion in the US, solidifying its position as the leading chip foundry [2] - TSMC has initiated an internal investigation regarding the potential unauthorized transfer of trade secrets by former executive Lo Wen-jen [2][3] Group 2 - Intel is focusing on advanced process technology to compete with TSMC and Samsung, aiming to produce chips at 2nm and below [4] - The company's strategic shift towards Intel Foundry aims to establish itself as a major player in high-end chip manufacturing for clients like Qualcomm [5] - Qualcomm is evaluating Intel for manufacturing its AI ASICs, indicating potential collaboration in the data center chip market [5][6]

Core Insights - Tenstorrent, led by Jim Keller, has achieved a valuation of $3 billion in its recent funding round, focusing on scalable AI accelerators for cloud and edge computing [2] - The company is developing RISC-V CPUs and plans to license its designs to other entities, utilizing open-source technologies to reduce costs associated with high-bandwidth memory [2][3] - The first chip, Tensix processor, is produced in collaboration with GlobalFoundries and features a grid-based architecture for efficient tensor computation [4] Product Development - The Grayskull processor, Tenstorrent's first Tensix processor, simplifies programming and enhances scalability, featuring up to 120 Tensix cores and supporting various AI precision formats [3] - The Wormhole variant of Grayskull reduces the number of cores to 80 but increases SRAM capacity and adds support for more precision formats, with a local memory capacity of 12GB GDDR6 [3][4] - Tenstorrent's RISC-V CPU, Ascalon, is set to compete with Arm's Neoverse V2 and V3, with performance metrics of 10 to 20 SPECint2006/GHz [5] Market Strategy - Tenstorrent is focusing on IP licensing, with existing clients including LG and Hyundai, and aims to allow customization of its cores, which is uncommon in the industry [4][5] - The company plans to release new core generations every 18 months, with the next generation, Babylon, expected to enhance performance and efficiency [7] - Tenstorrent is exploring markets in industrial and medical sectors, which have traditionally been overlooked by competitors like Arm [7] Technological Advancements - The upcoming Tensix-Neo core adopts a cluster architecture for improved area efficiency and utilization, allowing for more modular and flexible designs [8][10] - Customers will have the option to customize certain parameters while Tenstorrent manages more complex configurations, offering innovative licensing for those interested in deeper customization [10]

Core Viewpoint - AMD is set to launch its next-generation Instinct MI450 graphics accelerator using 2nm process technology, aiming to compete directly with NVIDIA's Rubin chip series [1][3][6] Group 1: Product Specifications - The Instinct MI450 will utilize TSMC's 2nm technology for its accelerator core chip (XCD), while the active interposer chip (AID) and media interface chip (MID) will be manufactured using TSMC's 3nm process [1][5] - MI450 is expected to deliver similar FP4/FP8 performance to NVIDIA's Rubin, but with 1.5 times the memory and bandwidth [3][5] - The MI400 series will integrate up to 432 GB of HBM4 memory, significantly enhancing bandwidth capabilities [7] Group 2: Market Positioning - AMD's data center operations head, Forrest Norrod, claims that MI450 will be the best solution for training, inference, distributed inference, and reinforcement learning, indicating confidence in competing with NVIDIA [6] - The MI450 is anticipated to be AMD's "Milan moment," referencing the significant leap made by AMD's EPYC server CPU series [6] - AMD plans to expand its rack options with the upcoming Helios system, which is expected to match the top configurations of NVIDIA's Rubin series [7] Group 3: Launch Timeline - The MI450 is projected to be released in 2026, coinciding with NVIDIA's launch of its next-generation Rubin series [7]

Core Insights - NVIDIA is investing $5 billion in Intel at a price of $23.28 per share, aiming to develop custom data center and PC products together to enhance applications for large-scale, enterprise, and consumer markets [1] - The collaboration will leverage NVIDIA's NVLink technology to integrate NVIDIA's AI and accelerated computing capabilities with Intel's leading CPU technology [1][4] - Intel will design x86 CPUs with integrated NVIDIA RTX GPU chipsets for the PC market, while NVIDIA will incorporate these CPUs into its AI infrastructure [1][4] Group 1 - NVIDIA's partnership with Intel allows it to expand into the integrated graphics market, previously dominated by Intel and AMD [4] - Intel's Xeon processors will play a larger role in NVIDIA's AI infrastructure and data center products, enhancing NVIDIA's offerings [4][6] - The integration of NVLink into Intel's CPU designs will enable NVIDIA to provide systems with both its own CPUs and Intel's Xeon processors [4][6] Group 2 - NVIDIA's NVLink technology offers significantly higher bandwidth compared to PCIe, with a maximum bidirectional bandwidth of 1.8 TB/s per GPU [7] - The collaboration allows for the expansion of GPU nodes from 8 to 72 in rack-scale architectures, enhancing performance capabilities [7] - Intel's adoption of NVLink marks a shift from its previous reliance on PCIe, indicating a strategic change in its approach to data center processors [6][8] Group 3 - NVIDIA's CEO Jensen Huang emphasized the potential for creating scalable AI supercomputers by integrating Intel's x86 CPUs into the NVLink ecosystem [9] - The partnership is expected to provide Intel with a competitive edge in the data center market against AMD [15][17] - The collaboration opens up a previously underserved market segment for NVIDIA, allowing it to integrate its GPU capabilities into Intel's chipsets [17][18] Group 4 - The deal raises questions about NVIDIA's future relationship with ARM and TSMC, as it shifts focus towards Intel for certain products [10][12] - Despite concerns, NVIDIA remains committed to its ARM roadmap and continues to collaborate with ARM-based SoC designers [11] - AMD is positioned as a potential loser in this scenario, as NVIDIA and Intel's partnership could disrupt the current competitive landscape [14][15]

Core Viewpoint - Nvidia plans to resume sales of its H20 chip to the Chinese market after obtaining an export license from the U.S. government, which is expected to alleviate sales pressure in China [1][3]. Group 1: Nvidia's H20 Chip and Market Impact - The H20 chip, designed specifically for the Chinese market, features 96GB of HBM3 memory and 4.0TB/s memory bandwidth, with FP8 performance at 296TFLOPS and FP16 performance at 148TFLOPS, making it suitable for vertical model training and inference [2]. - The export restrictions imposed by the U.S. Department of Commerce had previously hindered Nvidia's ability to fulfill approximately $2.5 billion in chip shipments and nearly $8 billion in H20 orders [1][2]. - Analysts from Needham have raised Nvidia's stock price target from $160 to $200, maintaining a "buy" rating, reflecting optimism about the recovery of H20 chip sales [1]. Group 2: Future Projections and Product Development - Analysts estimate that starting from the third quarter of the 2026 fiscal year (October), the quarterly shipment volume of the H20 chip could reach $3 billion [1]. - Nvidia is also developing Blackwell architecture GPU variants (including models B30/B40/RTX6000D) tailored for the Chinese market, with expected shipments beginning between August and September [1][3]. - The policy adjustment and new product development are seen as strategic moves by Nvidia to address the demand in the Chinese market [1].

Core Viewpoint - AMD plans to resume exports of its MI308 chip to China following the U.S. government's approval, marking a shift in the U.S. policy towards semiconductor sales to China [1] Group 1: U.S. Policy Changes - The U.S. Treasury Secretary mentioned that NVIDIA's approval to sell the H20 chip to China is justified as China has developed comparable chips [1] - The decision to allow AMD's MI308 chip to return to China represents a reversal from the previous stance of the Trump administration, which had been firm on restricting chip sales to China [1] Group 2: AMD's Strategic Moves - The MI308 chip is specifically designed by AMD as an AI accelerator for the Chinese market [1] - AMD's spokesperson confirmed that the U.S. Department of Commerce has informed the company that the license application for the MI308 product will enter the review process [1]

Group 1 - AMD plans to resume exports of its MI308 chip to China following approval from the U.S. government, marking a significant policy shift [1] - The MI308 chip is specifically designed by AMD as an AI accelerator for the Chinese market [1] - The U.S. Treasury Secretary indicated that NVIDIA's approval to export the H20 chip was based on China's development of comparable technology, suggesting a more lenient stance on chip exports to China [1] Group 2 - The recent approval for chip exports represents a reversal from the previous hardline stance of the Trump administration, which had been focused on restricting sales to China [1] - The easing of restrictions comes amid a period of improved U.S.-China relations, highlighted by NVIDIA's CEO meeting with Trump prior to the announcement [1]