从 2026 年 1 月至今，算力通胀的传导路径和市场演变节奏是怎样的？ 2026 年以来，算力通胀的传导链条呈现出从上游向下游逐步外溢的趋势。1 月 中旬起，市场需求侧已观察到 Token 消耗的快速增长，预示了全年算力通胀的 趋势。具体来看，通胀首先体现在 GPU 和存储环节，1 月份甚至 CPU 价格也 出现过小幅上涨。随后，通胀传导至云服务领域。1 月下旬，亚马逊云科技率 先提价，1 月 25 日谷歌云也宣布上调海外 CDN 价格，引发了市场对国内云厂 商涨价的预期。 进入 2 月，国内市场跟进趋势明显。2 月 5 日，网宿科技正式 公布 CDN 涨价；2 月 11 日，优刻得也宣布涨价。然而，当时市场主流观点认 为，在阿里巴巴和字节跳动两大巨头未明确表态前，中小云厂商的涨价行为更 多是试探性的，整个行业处于观望状态。尽管如此，当时产业内已形成共识， 即存储产品的价格上涨是确定性趋势，同时 GPU 服务器的价格也随着各批次到 货成本动态调整。 近期，随着阿里云和百度云正式宣布涨价，加之腾讯云针对 特定模型以及智谱 AI 的 Token 价格连续两轮上调，标志着算力通胀已明确传 导至国内主流云服务商和模 ...

Core Viewpoint - The U.S. Attorney's Office has charged individuals associated with a U.S. server manufacturer for illegally diverting billions of dollars in AI servers to China, highlighting concerns over unauthorized access to high-powered chips by Chinese companies [1]. Group 1: Legal Actions and Allegations - The U.S. government has filed an indictment against Yih-Shyan "Wally" Liaw, Ruei-Tsan "Steven" Chang, and Ting-Wei "Willy" Sun for violating the Export Control Reform Act [2]. - The indictment states that products containing Nvidia chips are subject to strict U.S. export controls, which prohibit their sale to China without a license, aimed at protecting U.S. national security [3]. Group 2: Industry Context and Responses - Nvidia's graphics processing units are in high demand globally for training generative AI models, indicating the competitive landscape between U.S. and Chinese companies [2]. - U.S. President Trump previously sought to prevent China from obtaining processors, but later indicated that Nvidia could ship H200 GPUs to China under specific conditions to maintain national security [3]. - Nvidia had received licenses to export the H20 chip to China last summer, with an agreement to provide the U.S. with 15% of its sales in China [4].

Core Insights - The article discusses the phenomenon of middle-aged entrepreneurs leading the current AI revolution, contrasting it with the younger leaders of the internet revolution [2][3] - It emphasizes that the AI revolution favors individuals with accumulated experience, emotional intelligence, and a sense of responsibility, which are often found in middle-aged professionals [3] Funding and Investment Landscape - AI entrepreneurship requires significant capital investment, likened to heavy industry, whereas internet startups were more akin to light industry with lower entry costs [5][6] - Training advanced AI models demands substantial resources, with costs reaching millions of dollars, making it challenging for younger entrepreneurs without access to large funding pools [6][7] - The shift in venture capital strategies has moved from broad investment in young entrepreneurs to a focus on experienced middle-aged leaders who can provide certainty and stability [14][16] Technical and Engineering Expertise - AI projects necessitate deep engineering knowledge and experience, which often excludes younger individuals who may lack the requisite background [8][9] - The complexity of AI model training requires extensive time and effort for system adjustments, contrasting sharply with the rapid iteration seen in internet startups [9] Organizational and Networking Advantages - Middle-aged entrepreneurs possess superior organizational skills and networks, which are crucial for managing the multifaceted demands of AI projects [10] - Established connections and industry knowledge enable these leaders to attract talent and resources that younger entrepreneurs may struggle to secure [10] Shifts in Capital and Regulatory Environment - The capital landscape has evolved to prioritize experienced entrepreneurs, with a focus on those who can navigate regulatory challenges and ethical considerations in AI development [13][18] - Regulatory scrutiny has increased, necessitating a deeper understanding of compliance and ethical implications, which middle-aged leaders are better equipped to handle [19][20] Opportunities for Younger Entrepreneurs - Despite the dominance of middle-aged leaders, there remains space for young entrepreneurs to innovate and contribute significantly to the AI landscape [22][24] - Young professionals often excel in technical execution and can drive rapid product development, complementing the strategic oversight of their older counterparts [24] Strategic Directions for Middle-aged Entrepreneurs - Middle-aged leaders are encouraged to define industry problems accurately, leverage their accumulated knowledge to create competitive advantages, and manage AI-human collaboration effectively [28][31] - Establishing ethical frameworks and regulatory compliance will be essential for long-term success in the AI sector, where trust is a critical asset [33]

Core Viewpoint - The article discusses the emergence of a new AI model named "Hunter Alpha," which has sparked speculation about its connection to the upcoming DeepSeek V4 model due to its impressive performance metrics and anonymous release [3][4][6]. Group 1: Performance Metrics - Hunter Alpha boasts a parameter scale of 1 trillion, placing it among the leading models in the industry [4]. - The model claims to have a context window of up to 1 million tokens, significantly surpassing most commercial models, allowing it to handle longer texts and more complex tasks [4]. - As of the latest statistics, Hunter Alpha has processed over 160 billion tokens, indicating rapid adoption among developers [5]. Group 2: Connection to DeepSeek - The model's self-identification as a "Chinese AI model trained primarily in Chinese" and its knowledge cutoff date of May 2025 align with the specifications of DeepSeek's existing models [6]. - Some developers suggest that the reasoning style of Hunter Alpha may reveal its "heritage," with its scale and memory capacity matching expectations for DeepSeek V4 [7]. - Despite the similarities, some analysts remain cautious about definitively linking Hunter Alpha to DeepSeek V4, noting differences in token behavior and architectural patterns [9][10]. Group 3: Industry Practices - The anonymous release of AI models for real feedback has become a standard practice in the industry, with platforms like OpenRouter facilitating testing across multiple AI systems [8]. - Notifications on Hunter Alpha's profile indicate that all prompts and completions are recorded for model improvement, a common practice in the field [9].

The H200 is Nvidia's second-most powerful AI chip. It sits below the company's current-generation Blackwell line, which remains off-limits for export to China under the terms of the arrangement. Export licenses come with conditions: The U.S. takes 25% of chip sale proceeds, shipments are capped, and sales must go through third-party verification, according to CNBC. Before export controls took hold, China was responsible for roughly 13% of Nvidia's total revenue and generated at least a fifth of its data cen ...

Jonathan Kitchen/DigitalVision via Getty Images Listen below or on the go via Apple Podcasts and Spotify What is “Hunter Alpha”? AI model fuels talk of new system at DeepSeek. (00:14) lululemon athletica (LULU) beats top- and bottom-line but sets disappointing guidance. (01:39) Amazon (AMZN) plans drastic cut in packages it sends through US Post Office: report. (02:58) This is an abridged transcript. An artificial intelligence model that surfaced anonymously on a developer platform last week is said ...

温馨提示 ： 点击下方图片，查看运营团队最新原创报告（共260页） 说明： 欢迎约稿、刊例合作、行业交流 ， 行业交流记得先加入 "机器人头条"知识星球 ，后添加（ 微信号：lietou100w ）微信； 若有 侵权、改稿请联系编辑运营（微信：li_sir_2020）； 正文： 3月，已非Optimus V3发布最佳节点，关注T链们左侧机会，静待右侧机会！ 对于T链们而言，3月催化点是Optimus V3亮相，关键点是V3表现超预期；值的注意的是，目前关于Optimus V3 发布时间已有分歧，市场口径多改为3月底或4月初。 马斯克在3月12日最新访谈中， 针对Optimus V3发布及量产问题，也仅表示" Optimus V3已处于最后完成阶 段，今年夏季开始生产，预计明年实现大规模量产 "。 —— 针对V3具体发布时间，目前有不少卖方老师观点是：特斯拉将在3月底发布V3视频，并在4月初举办专项发布 会； 小编视角： 两者可能性皆较小，外围因素影响下，3月已非V3发布最佳节点；同时，停办AI Day以来，特斯 拉未举办单个产品专项发布会且V3定位是自家工厂。 接下来，聊一下本周T链走势和利好T链标的们： 外 ...

性能参数触发市场敏感点 一个拥有万亿参数的神秘免费AI模型突然上线，关于DeepSeek V4将发布的猜测再一次涌现。 据路透社3月18日报道， 一款名为"Hunter Alpha"的AI模型近日在开发者平台OpenRouter匿名上线，引发全球开发者社区关注 。该模型未标注开发者身份， 但观察其性能参数与时间节点，市场猜测这可能是DeepSeek在正式发布前对其下一代系统进行的秘密测试。 Hunter Alpha于3月11日以"隐身模型"形式发布， 且目前向开发者提供免费访问。测试显示，该系统具备1万亿参数规模和高达100万token的上下文窗口。 在测试中，该模型自称"主要以中文训练的中国AI模型"，知识截止时间为2025年5月，这一点与DeepSeek现有模型一致。但当被问及开发者时，其回应："我 只知道自己的名字、参数规模和上下文长度。" OpenRouter平台未披露模型来源，DeepSeek亦未回应置评请求。 数据截止期与推理风格指向DeepSeek 将Hunter Alpha与DeepSeek联系起来的线索，主要来自其底层数据特征和运行逻辑。 在媒体测试中，该聊天机器人自称是"主要用中文训练的中 ...

By Karen Freifeld, Max A. Cherney and Liam Mo NEW YORK, March 17 (Reuters) - Nvidia has won Beijing's approval to sell its second-most powerful artificial intelligence chips to China and is also preparing a version of the Groq AI chip that can be sold to the ‌Chinese market, sources familiar with the matter said. The long-awaited regulatory approval paves the way for the U.S. chipmaker to resume sales of the ‌H200 chips, which have emerged as a major flashpoint in U.S.-China relations, in a market that ...

作者 | 允毅 连马斯克、Andrej Karpathy 都纷纷点赞，DeepSeek 和 Kimi 前后脚都盯上的 "残差连接" ，到底是 什么？ 最近，Kimi 放出一篇重磅新论文，瞄准一个过去十年几乎没人动过的 Transformer 底层根基： 残差 连接（Residual Connection） 。残差连接由何恺明于 2015 年在 ResNet 论文中提出，此后便成为 深度学习领域的标配。 简单来说，可以把大模型的 Transformer 架构，想象成一支几十人排成长队的"传话小组"，那么残差 连接就像一条规定：每个工人听完前面所有人的话后，都往里面再补一句，然后原封不动往后传。 这套规则长这样： 但这会带来一个麻烦：队尾的工人收到的话，是前面几十个工人的内容全堆在一起的，越往后话越 乱、越长，前面工人说的重点被埋住了，后面工人加的内容也没人听得清，AI 就变笨了。这叫"稀释 问题"。 于是，Kimi 想到把 "注意力机制" 引进来解决这一问题，它提出一个新的规则： "注意力残 差"（Attention Residuals） 。如同给工人们配备了"智能筛选器"，不用再全盘收下前面堆出来的大 杂烩， ...