Krypton Star Evening News | Dreame pledges 247 million shares of Jiamei Packaging as collateral for loans from Industrial Bank; Huang Renxun: SK Hynix's plan to double wafer capacity by 2030 is insufficient 氪星晚报｜追觅质押嘉美包装2.47亿股，向兴业银行贷款提供担保；黄仁勋：SK海力士到2030年将晶圆产能翻倍的计划还不够

Real, solid progress is quietly happening elsewhere. WeChat's guidelines for developers on its AI ecosystem might sound like a technical document, but with Ctrip and Tongcheng Travel among the first OTAs to integrate, the signal is unmistakable. This isn't about showing off tech prowess; it's about giants seeking a "new vessel" for their massive but slightly aging ecosystems. Letting AI agents handle hotel bookings and itinerary planning directly means the logic of traffic distribution could shift once again. WeChat is attempting to redefine "connection" through AI, and the tourism industry—with its complex decision chains and data dimensions—has become a perfect testing ground. This is far more compelling than speculating on "physical AI," as it involves real-world scenarios and the migration of user habits with tangible financial stakes. However, with Ctrip and Tongcheng announcing integrations almost simultaneously, it seems the script for homogenized competition among the first batch of players is already written.

It only takes one privacy incident to turn euphoria into harsh reality. The Rokid glasses recording incident and its swift rectification serve as a resounding slap to all companies eager to push wearable AI devices to market. When the convenience of technology clashes head-on with the fragility of individual privacy, any technical showmanship appears hollow. Rokid's response was timely—algorithm upgrades, mandatory hardware indicator lights, and removal of non-compliant accessories—forming a combination of corrective measures. But the root problem hasn't disappeared: When the instinct to record is infinitely amplified by technology, who guards the ethical boundaries of public spaces, and with what tools? This is far from something companies can resolve through unilateral rectification; it questions society's bottom line for accepting an era of "invisible recording."

Meanwhile, the competition in upstream computing power has entered a white-hot phase. Jensen Huang's complaint that SK Hynix's production capacity is "still not enough," along with discussions about finalizing HBM4 and foundry cooperation with Samsung, paints a realistic picture of the AI arms race: Chips and storage capacity are the new oil and steel. The deepened collaboration between Samsung and NVIDIA is not just a binding of interests between two giants but also a reshaping of the global semiconductor supply chain. Without breaking the bottleneck of computing power, all higher-level AI applications remain castles in the air. This contest over foundational capabilities is barely perceptible to ordinary users, yet it determines the distribution of technological power for the next decade.

At a more fundamental level, the document from the National Data Administration on advancing high-quality industry datasets resembles a calm blueprint. When everyone is focused on algorithms and models, it reminds us that high-quality, scenario-specific data is the "nourishment" for training good AI. Cultivating data annotation enterprises might sound as mundane as discussing traditional manufacturing, but this is precisely the critical step in moving AI from an "alchemy furnace" to the production line. Without large-scale, standardized data processing, even the best models would be like a skilled chef trying to cook without rice. The value of this document may only be fully recognized years from now, upon reflection.

Dreame Technology's pledge of shares to secure financing for its acquisition of Jiamei Packaging上演了一场机器人公司进行资本腾挪的戏码。一边是前沿科技，一边是传统包装行业，这笔高达22亿元的收购本就引人侧目。如今创始人将刚到手的股份大幅质押，为之前的并购融资补上风控措施，资本运作的痕迹一览无余。科技公司的野心与资本市场的游戏规则在此紧密纠缠，至于这最终能为"机器人"主业带来何种赋能，悬念留给了未来。

总体而言, 本周的AI资讯光谱从资本市场荒诞的"物理AI"概念股炒作，到微信开放生态的务实布局；从隐私安全的血泪教训，到算力基石的全球争夺；再到数据要素的顶层设计，拼凑出一幅行业在狂热与理性间挣扎前行的全景图。真正的变革往往不在喧嚣的公告中，而是在这些看似无关的细节——一次合作、一次整改、一份文件、一笔质押——的交织中，默默发生。

当一众上市公司急不可耐地发公告澄清“本公司没有物理AI业务”时，A股市场的魔幻现实主义才算真正揭幕。天娱数科、奥比中光、锋龙股份，名字里带点科技味的公司，股价像坐了火箭，然后又火速撇清关系。这不是商业故事，这是一出黑色喜剧。市场显然在病急乱投医，任何沾上“AI”两个字母的概念，都被视作救命稻草，哪怕这根稻草是“物理AI”——一个连业界都尚未形成共识、应用远在天边的时髦词汇。上市公司们一边享受着无厘头炒作带来的市值飙升，一边又害怕被扣上“误导投资者”的帽子，只能尴尬地站出来说：“别误会，我们真不是。” 这场景，像极了围观一场名为“皇帝的新衣”的现代演出，只不过新衣上绣着的不是花纹，是“AI”。

真正扎实的进展，在别处悄然发生。微信向开发者开放AI生态的指引，听起来像是个技术文档，但携程、同程旅行作为首批OTA接入，信号却异常清晰。这不是炫技，而是巨头在为自己庞大但略显老化的生态寻找“新肉身”。让AI智能体直接处理酒店预订、行程规划，意味着流量分发逻辑可能再次改变。微信试图用AI重新定义“连接”，而旅游行业因其复杂的决策链和数据维度，成了绝佳的试验场。这比炒作“物理AI”性感得多，因为它关乎真金白银的场景和用户习惯的迁移。不过，携程和同程几乎同时宣布接入，首批玩家同质化竞争的剧本，似乎也已写好。

从狂欢到现实，只需要一次隐私事故。Rokid眼镜拍摄事件及其火速整改，是一记响亮的耳光，扇向所有迫不及待把可穿戴AI设备推向市场的公司。当科技的便利性与个体隐私的脆弱性短兵相接，任何技术炫技都显得苍白。Rokid的声明很及时，算法升级、硬件强制亮灯、下架违规配件，一套组合拳打下来，但根本问题并未消失：当记录的本能被技术无限放大，公共空间的伦理边界由谁、用什么来守护？这远非企业单方面整改能解决，它拷问的是整个社会对“无感记录”时代的接纳底线。

与此同时，产业链上游的算力竞赛已进入白热化。黄仁勋对SK海力士产能“还不够”的抱怨，以及其与三星敲定HBM4及代工合作的讨论，勾勒出AI军备竞赛最真实的底色：芯片和存储的产能，就是新时代的石油与钢铁。三星与英伟达的合作深化，不仅是两家巨头的利益捆绑，更是对全球半导体供应链的一次重塑。算力瓶颈不破，所有上层AI应用都是空中楼阁。而这种底层能力的争夺，普通用户感知甚微，却决定了未来十年的技术权力格局。

在更基础的层面，国家数据局推进行业高质量数据集建设的文件，像一份沉静的施工图。当所有人都盯着算法和模型时，它提醒我们，高质量、场景化的数据才是训练好AI的“养料”。培育数据标注企业，听起来朴素得像在讨论传统制造业，但这恰恰是AI从“炼丹炉”走向生产线的关键一环。没有规模化、规范化的数据处理，再好的模型也是巧妇难为无米之炊。这份文件的价值，或许要等几年后回看，才会被充分认识。

追觅科技质押股份为收购嘉美包装融资担保，则上演了一出机器人公司进行资本腾挪的戏码。一边是前沿科技，一边是传统包装行业，这笔高达22亿元的收购本就引人侧目。如今创始人将刚到手的股份大幅质押，为之前的并购融资补上风控措施，资本运作的痕迹一览无余。科技公司的野心与资本市场的游戏规则在此紧密纠缠，至于这最终能为“机器人”主业带来何种赋能，悬念留给了未来。

总的来说，本周的AI资讯光谱，从资本市场荒诞的“物理AI”概念股炒作，到微信开放生态的务实布局；从隐私安全的血泪教训，到算力基石的全球争夺；再到数据要素的顶层设计，拼凑出一幅行业在狂热与理性间挣扎前行的全景图。真正的变革，往往不是在喧嚣的公告中，而是在这些看似无关的细节——一次合作、一次整改、一份文件、一笔质押——的交织中，默默发生。