Beingmate: Actual Controller Expected to Change to Jinhua Municipal State-owned Assets Supervision and Administration Commission 贝因美：实际控制人拟变更为金华市国资委

ChatGPT and Codex have announced their merger, as OpenAI twists these two flagship products into one, claiming to deliver a "super Agent" to a billion people. The news is like a stone thrown into a lake, with ripples rapidly spreading to every corner. However, upon冷静 reflection, this appears more as a shrewd market and product strategy adjustment rather than some disruptive technological singularity. Codex, as a code-generation tool, primarily targets developers, while ChatGPT's user base is far more diverse. The forced "merger" raises questions: Is it genuinely aimed at enhancing efficiency in complex tasks, or is it a response to slowing user growth anxiety, crafting an "all-powerful" narrative to merge the traffic pools of developers and everyday users, thereby boosting overall usage metrics? A true "super Agent" should be capable of autonomous planning, tool invocation, and cross-domain collaboration—not merely a simple superposition of two models' functionalities. This step seems more like functional integration, still far from the vision of an autonomous intelligent agent.

Turning attention to China, Volcengine's MaaS (Model-as-a-Service) revenue target has been set at 15 billion yuan for the full year, with its video generation model Seedance 2.0 alone exceeding 1 billion in monthly revenue. These figures stand out starkly—even somewhat inappropriately fervent—against the backdrop of an industry downturn. It must be acknowledged that this showcases ByteDance's formidable sales and channel capabilities, packaging large model capabilities as cloud services ready for rapid sale. However, the 15 billion annual target implies an average monthly income exceeding 1.2 billion. How much of this reflects genuine, sustained investment from enterprises? How much comes from one-off project-based purchases rather than sustainable subscription services? More critically, as numerous players (including Alibaba, Baidu, Tencent, and a host of startups) fiercely compete in the MaaS market, will this high-revenue-target-driven "arms race" quickly drag model services into a price war quagmire? Ultimately, what might be harmed is the entire industry's pursuit of value and profit, turning it into a war of attrition over who can "burn" more resources.

Just days ago, the "strategic cooperation memorandum" between Alibaba Cloud and Manulife to "accelerate the large-scale implementation of AI" reads with the "visionary" rhetoric common in the insurance industry. Terms like "full-scenario" and "new pathways for intelligent upgrade" often underpin collaborations that begin with one or two pilot projects but stall amid data barriers, compliance red lines, and organizational inertia. True large-scale implementation requires not just a memorandum, but disruptive efficiency improvements and sustainable business models. The most formidable challenge in insurance AIization has never been the technology itself, but rather how to convince actuaries, sales agents, and regulators to trust decisions made by a "black box."

On another front, the news that Beingmate, a dairy company, is set to transfer control to a local state-owned asset supervisor may seem unrelated to AI. Yet it acts as a mirror, reflecting another facet of the current tech investment fervor. As capital markets grow weary of pure narratives, some local state-owned entities may be attempting to take over companies with solid industrial foundations through "restructuring." This invites reflection: In an era where everyone is discussing large models, agents, and computing power monopolies, what kind of fragmented yet interdependent landscape is formed by the valuation bubble in tech concepts and the survival struggles of real industries? Intel claims it will unleash a major move to end Nvidia's computing power monopoly—certainly a welcome competition in the industry—but changing a monopoly never hinges on a single announcement. It requires a long reconstruction spanning chip architecture, software ecosystems, and customer trust.

The tech charts refresh daily, with concepts emerging endlessly. From "super Agents" to "AI-native," from "trillion parameters" to "hundred-billion revenues," the narratives under the spotlight are always thrilling. Yet, beneath the glossy exterior, industry progress still depends on conquering specific scenarios, landing pragmatic partnerships, and a calm return to commercial fundamentals. As all players shout "disruption," perhaps the most scarce quality is precisely that unglamorous patience and pragmatism.

ChatGPT和Codex宣布合体，OpenAI把这两个拳头产品拧成一股绳，号称要给十亿人一个“超级Agent”。这消息像一颗投入湖面的石子，涟漪迅速扩散到每个角落。但冷静下来看，这更像是一次精明的市场和产品策略调整，而非什么颠覆性的技术奇点。Codex作为代码生成工具，其核心用户群是开发者，而ChatGPT的用户则泛得多。强行“合体”，究竟是为了真正提升复杂任务的完成效率，还是为了在用户增长放缓的焦虑下，制造一个“无所不能”的叙事，把开发者和普通用户的流量池合并，从而拉高整体使用数据？真正的“超级Agent”应能自主规划、调用工具、跨域协作，而不仅仅是两个模型功能的简单叠加。这一步更像是功能整合，离愿景中的自主智能体还有漫长的距离。

视线转向国内，火山引擎的MaaS（模型即服务）营收目标被提到全年150亿元，旗下视频生成模型Seedance 2.0单月营收就超10亿。这些数字在行业寒冬的背景下显得格外刺眼，甚至有点“不合时宜”的狂热。必须承认，这展现了字节跳动强悍的销售和渠道能力，把大模型能力包装成可快速售卖的云服务。但150亿的年目标，意味着平均月收入需超过12亿。这背后是多少家企业真金白银的投入？有多少是项目制的一次性采购，而非可持续的订阅式服务？更关键的是，当众多厂商（包括阿里、百度、腾讯以及一众创业公司）都在疯狂争夺MaaS市场时，这种以高额营收目标驱动的“军备竞赛”，是否会迅速将模型服务拖入价格战的泥潭？最终损害的，可能是整个行业对价值和利润的追求，变成一场比拼谁更“烧得起”的消耗战。

而就在前几天，阿里云与宏利人寿那份“加速AI规模化落地”的战略合作备忘录，读起来则充满了保险行业常见的“展望式”话术。“全场景”、“智能化升级新路径”——这些词汇包装下的合作，往往始于一两个试点项目，卡在数据壁垒、合规红线与组织惯性之间。真正的规模化落地，需要的不是一纸备忘录，而是颠覆性的效率提升案例和可持续的商业模式。保险业的AI化，最棘手的从来不是技术本身，而是如何说服精算师、销售员和监管机构，去信任一个“黑箱”所做的决策。

另一边，贝因美这家奶粉公司的控制权拟变更为地方国资委，这条新闻看似与AI无关，却像一面镜子，映照出当前科技投资热潮中的另一面。当资本市场对纯粹的故事感到疲惫，一些地方国资或许正试图通过“重整”来接盘具有实体产业基础的公司。这让人不禁思考：在全民热议大模型、Agent、算力垄断的当下，科技概念的估值泡沫与实体产业的生存挣扎，构成了怎样一幅割裂又共生的图景？英特尔号称要甩出大招终结英伟达的算力垄断，这当然是行业喜闻乐见的竞争，但改变垄断格局，靠的从来不是一次发布，而是从芯片架构、软件生态到客户信任的漫长重建。

科技热榜每天都在刷新，概念层出不穷。从“超级Agent”到“AI原生”，从“万亿参数”到“千亿营收”，聚光灯下的叙事总是激动人心。但剥开光鲜的外壳，产业的进步依然依赖于一个个具体场景的攻克、一桩桩务实合作的落地、以及对商业本质的冷静回归。当所有玩家都在高喊“颠覆”时，或许最稀缺的品质，恰恰是那些不性感的耐心与务实。