| CURSOR | N Notion | OpenEvidence) | ABRIDGE | nclay | | --- | --- | --- | --- | --- | | GAMMA | WRITER | ZED INDUSTRIES | allip | SUPERHUMAN | | World Labs | P MERCOR | ◆ Mirage | Bland® | = descript | 图片来源： theinformation 知情人士透露，人工智能初创公司 Baseten Labs 以 50 亿美元的估值筹集了 3 亿美元资金。 部落后 and Z | Z Potentials e 2017 公众号 社群 我们与Z Potentials同频共振， 交流和分享你们的故事和想法, 我们会是最好的倾听者。 Interview Product Event Research Yuca Hong % captain199508 主要关注AI、智能硬件、全球化等前沿科技， 前四大交易并购。中国注册会计师。 % qq2943115158 Kelly Xu 主要关注AI、智能硬件出海等科技领域。欢迎 交流。 ...

Asian shares have extended their losses while gold prices hit a record high as concerns over U.S. President Donald Trump’s tariff threats on Greenland fueled unease among investorsA person walks in front of an electronic stock board showing Japan's Nikkei index at a securities firm Wednesday, Jan. 21, 2026, in Tokyo. (AP Photo/Eugene Hoshiko)HONG KONG -- Asian shares mostly extended their losses on Wednesday and gold prices hit a record high as concerns over U.S. President Donald Trump’s tariff threats on G ...

Core Insights - Nvidia's CEO Jensen Huang is planning a trip to China in late January to strengthen the company's position in the AI chip market [1][2][3] - The visit coincides with the US easing export restrictions on Nvidia's H200 model, while China is deliberating on import approvals for these chips [3][4] - The geopolitical landscape is tense, with US lawmakers expressing concerns over China's access to advanced AI technology [8][10] Group 1: Nvidia's Strategic Moves - Huang's trip is aimed at reopening a crucial market for Nvidia's AI chips, particularly as the company seeks to navigate changing export regulations [1][3] - The H200 chip, although a generation behind US offerings, is still considered superior to local Chinese alternatives, highlighting Nvidia's competitive edge [4] - The visit may include meetings with Chinese officials, which could impact Nvidia's market strategy in China [2][3] Group 2: Regulatory Environment - The US has imposed a 25% tariff on certain advanced semiconductors shipped to China, reflecting ongoing trade tensions [7] - Additional licensing requirements for the H200 chip mandate that exporters ensure shipments do not create shortages in the US market [7] - Concerns from US lawmakers about the potential security risks of allowing China access to Nvidia's advanced chips are growing, with proposals for increased oversight on such sales [8][9] Group 3: Market Implications - The Chinese government is looking to bolster its domestic chip industry, which currently cannot match Nvidia's capabilities [4] - Despite restrictions, the potential for Nvidia to sell AI chips in China remains a contentious issue among US lawmakers, who fear it could enhance China's technological capabilities [8][10] - The outcome of Huang's trip and the evolving regulatory landscape will be critical for Nvidia's future market position in China [10]

证券研究报告 | 2026年1月21日 行业研究 · 行业专题 AI产业链 投资评级：优于大市（维持评级） 证券分析师：熊莉 021-61761067 xiongli1@guosen.com.cn S0980519030002 证券分析师：艾宪 0755-22941051 aixian@guosen.com.cn S0980524090001 证券分析师：王晓声 010-88005231 wangxiaosheng@guosen.com.cn S0980523050002 0755-81982362 lishuying@guosen.com.cn S0980524090005 证券分析师：张宇凡 021-61761027 zhangyufan1@guosen.com.cn S0980525080005 联系人：赵屿 021-61761068 zhaoyu6@guosen.com.cn 证券分析师：李书颖 证券分析师：袁阳 0755-22940078 yuanyang2@guosen.com.cn S0980524030002 请务必阅读正文之后的免责声明及其项下所有内容 摘要 AI产业链系列报告一： 26年算 ...

在AI资本开支仍维持高位、但商业化压力不断上升的当下，市场关注点正在发生一场悄然却深刻的转 移：大模型还能不能继续"无视成本地跑下去"。 据追风交易台，高盛最新发布的AI芯片研究报告，并未延续市场熟悉的"算力、制程、参数规模"对比， 而是从更贴近商业现实的角度切入——推理阶段的单位成本。通过构建一条"推理成本曲线"，高盛试图 回答一个对AI产业至关重要的问题：在模型进入高频调用阶段后，不同芯片方案在折旧、能耗和系统 利用率等约束下，每处理一百万个token究竟需要付出多少真实成本。 研究结论指向了一次正在加速、但尚未被充分消化的变化：Google/Broadcom 的TPU正在迅速缩小与英 伟达GPU在推理成本上的差距。从TPU v6升级至TPU v7，单位token推理成本下降约70%，使其在绝对 成本层面与英伟达GB200 NVL72基本持平，部分测算情形下甚至略具优势。 这并不意味着英伟达的地位被动摇，但它清晰地表明，AI芯片竞争的核心评价体系正在从"谁算得更 快"，转向"谁算得更便宜、更可持续"。当训练逐渐成为前期投入，而推理成为长期现金流来源，成本 曲线的斜率，正在取代峰值算力，成为决定产业格局的关 ...

点击蓝字，关注我们 今日上午港股 1 . 机器人概念股普涨， 微创机器人(02252)涨11.91%，卧安机器人(06600)涨9.38%，极智嘉-W(02590) 涨5.48%，优必选(09880)涨2.01%，德昌电机控股(00179)涨2.43% 。 消息面上，1月21日，国新办举行 新闻发布会上，工业和信息化部副部长张云明表示，下一步工信部将持续推动人形机器人技术创新和 迭代升级，以人形机器人为小切口带动计算智能大产业发展，将持续加强人形机器人产品质量、网络 和数据安全方面的检验检测，开展相关科技伦理研究与管理服务，以高水平安全保障高质量发展。 2 . 华虹半导体(01347)现涨近5% 。消息面上，TrendForce最新报告指出，随着台积电、三星电子降低8 英晶圆代工产能，将导致2026年全球8英寸晶圆代工总产能同步减少2.4%。 3 . 泛远国际(02516)盘中跌超11%，截至发稿，跌13.97% 。消息面上，泛远国际发布公告，于2026年1 月20日，本公司与卖方订立该协议，本公司有条件同意向信托收购COPE Holding的全部股份、认购 COPE的690股新普通股，并向An女士收购Hyp ...

除"安娜档案馆"外，起诉状还指出英伟达存在多渠道获取盗版数据的行为，其不仅从"图书馆基 因"（LibGen）、"科学枢纽"（Sci-Hub）、"Z图书馆"（Z-Library）等平台下载图书，还向企业客户分 发脚本和工具，助力客户自动下载包含盗版Books3数据集的"The Pile"数据库。基于此，原告方新增了 辅助侵权与共同侵权两项诉讼主张，认为英伟达通过为他人获取盗版数据集提供便利牟利。 来源：环球网 据悉，这是美国大型科技公司与"安娜档案馆"的往来函件首次被公开披露。此前不久，"安娜档案馆"已 丢失多个域名，此次与英伟达的侵权纠纷进一步提升了这家盗版图书馆的公众关注度。目前，诉状尚未 明确提及英伟达是否向"安娜档案馆"支付了数据访问费用。（纯钧） 【环球网科技综合报道】1月21日消息，据AlBase报道，芯片巨头英伟达近日陷入一场备受关注的版权 集体诉讼。多位图书作者通过修订后的起诉状指控，英伟达为训练自主研发的人工智能模型，蓄意 从"安娜档案馆"等多个"影子图书馆"获取海量盗版数据，涉案数据规模达500TB，包含数百万本受版权 保护的图书，相关行为已涉嫌侵犯著作权。 早在2024年初，就有多位作者以 ...

据知情人士透露，黄仁勋将前往中国参加春节前的公司年会活动。他还预计会前往北京，但目前尚不清 楚他是否会与中国高级官员会面。知情人士称，黄仁勋的行程仍可能调整，具体将取决于相关会晤能否 最终确定。 黄仁勋此次中国行属于年度例行访问，他通常会在每年这个时候访华，正值该公司处于关键时刻。此 前，美国放宽了AI处理器的出口限制，允许英伟达向中国销售其H200芯片。 黄仁勋 凤凰网科技讯北京时间1月21日，据彭博社报道，英伟达CEO黄仁勋(Jensen Huang)计划在1月底访问中 国。眼下，黄仁勋正努力为英伟达AI芯片业务重新打开中国这个关键市场。 截至发稿，英伟达不予置评。(作者/箫雨) 周三，黄仁勋计划出席在瑞士达沃斯举行的世界经济论坛，届时将接受贝莱德CEO拉里.芬克(Larry Fink)的采访。 ...

Core Insights - Nvidia CEO Jensen Huang is planning a trip to China in late January to re-establish a vital market for the company's artificial intelligence chips [1] Group 1 - The visit aims to reopen access to a significant market for Nvidia's AI chip products [1]

Summary of Key Points from the Conference Call Industry Overview - The analysis focuses on the AI semiconductor industry, particularly the competition between GPUs (Graphics Processing Units) and ASICs (Application-Specific Integrated Circuits) in the context of inference cost performance [1][9]. Core Insights - **Inference Cost Dynamics**: The inference cost curve analysis indicates that Google's TPU (Tensor Processing Unit) is rapidly closing the performance gap with Nvidia's GPU solutions, with a ~70% reduction in cost per token from TPU v6 to TPU v7, making it competitive with Nvidia's GB200 NVL72 [2][15]. - **Nvidia's Competitive Edge**: Despite the advancements of TPUs, Nvidia maintains a lead due to its faster time to market and the strength of its CUDA software, which serves as a significant competitive moat for enterprise customers [2][12]. - **AMD and Amazon's Position**: AMD's solutions and Amazon's Trainium have shown a ~30% cost reduction but are currently lagging behind Nvidia and Google in terms of absolute cost performance. Future improvements are expected with AMD's upcoming Helios rack and Trainium 3&4 [2][16][33]. - **Technological Advancements**: Innovations in networking, memory, and packaging technologies are critical for driving further cost reductions in AI compute, as compute dies are nearing physical limits [3][12]. Future Scenarios - Four potential scenarios for the evolution of AI and the GPU vs. ASIC debate are outlined: 1. Limited traction in enterprise and consumer AI, leading to faster ASIC adoption. 2. Continued scaling of consumer AI with limited enterprise traction, benefiting Nvidia's training market dominance. 3. Similar to scenario 2 but with moderate enterprise traction, allowing Nvidia to capture incremental revenue. 4. Strong scaling of both consumer and enterprise AI, favoring Nvidia due to its training market dominance [25][26]. Stock Recommendations - **Nvidia (Buy)**: Expected to maintain its competitive lead in the accelerator market, with a 12-month price target of $250 based on a 30X P/E multiple applied to an EPS estimate of $8.25. Key risks include a slowdown in AI infrastructure spending and increased competition [27][28]. - **Broadcom (Buy)**: Anticipated to benefit from the increasing use of TPUs and its leading networking capabilities, with a price target of $450 based on a 38X P/E multiple applied to an EPS estimate of $12.00 [29][32]. - **AMD (Neutral)**: Awaiting more data on customer wins and OpenAI deployments, with a price target of $210 based on a 30X P/E multiple applied to an EPS estimate of $7.00 [33][34]. - **Marvell (Neutral)**: Limited visibility into the ramp of new custom compute programs, with a price target of $90 reflecting a 27X P/E multiple applied to an EPS estimate of $3.40 [35][36]. Additional Considerations - Nvidia's stock is currently trading at a ~25% discount to the median AI chip designer stock, which may indicate potential for price appreciation if performance benchmarks improve and costs decrease relative to competitors [22][23]. - The importance of software optimization and the developer ecosystem, particularly Nvidia's CUDA, is highlighted as a key factor in maintaining competitive advantage [25]. This summary encapsulates the critical insights and projections regarding the AI semiconductor industry, focusing on the competitive landscape and investment opportunities within the sector.