Meta Repeatedly Delays Release of New AI Model API for Developers Meta一再推迟面向开发者的新AI模型发布

In contrast, Jensen Huang’s itinerary appears purposeful and pragmatic. He personally flew to South Korea to meet with Krafton, the developers of PUBG, to discuss concrete plans for deploying RTX Spark chips and collaborating on "physical AI." There were no lofty concepts here—just a focus on gaming, the industry that best consumes computational power and vividly showcases AI hardware performance. From data centers to consumer laptops, NVIDIA’s reach is becoming omnipresent. Huang understands a core truth: ecosystems aren’t built through PowerPoint presentations but forged through tangible partnerships and applications running on hardware. His visit to a game company is essentially NVIDIA seeking a new, potentially broader outlet for its computational power, pulling AI from the cloud into handheld devices. This ability to "stuff chips into every device" is precisely the executional drive that companies like Meta—more adept at grand narratives—lack.

Looking back at domestic developments, the past 24 hours of trending topics resemble a microcosm of AI implementation. Microsoft proclaimed that "1.6 billion Windows users will enter the Agent era overnight"—a rallying cry that sounds thrilling. Yet, an operating system giant’s AI ambitions must ultimately withstand the test of countless real-world use cases. Will the Agent truly automate workflows, or will it become another "smart" assistant that requires users to relearn habits and even causes more chaos? The market will likely deliver a sobering verdict.

Meanwhile, Volcano Engine’s decision to raise its annual MaaS (Model as a Service) revenue target to 15 billion yuan, with Seedance 2.0 alone generating over 1 billion yuan in monthly revenue, sends a more robust signal. It indicates that in the Chinese market, the AI commercialization race has shifted from "whether we have models" to "how much money models can make." Leveraging its vast traffic pool and application ecosystem, ByteDance has indeed picked up speed in monetizing AI capabilities.

More intriguing is the fray among Tencent, Alibaba, and ByteDance over "Skill stores." This foreshadows that AI’s next battlefield may no longer be about who has the largest foundational model parameters or highest benchmark scores, but who can build a richer, more accessible ecosystem of AI capability plugins. Much like the App Store wars in the smartphone era, whoever enables developers to create "skills" more easily and allows users to access intelligence more seamlessly may seize the gateway to the next computing platform. These tech giants are gearing up to compete for what is essentially the "App Store of the AI era."

As for the discussion about "why Chinese cars are getting bigger," it may seem unrelated to AI but is in fact closely connected. A larger car body means more complex interior space, providing a physical stage for AI cabins, more immersive interaction, and additional sensors required for autonomous driving. Cars are evolving from mere transportation tools into mobile intelligent spaces, with AI serving as their "soul" and "brain." Building bigger cars, in a sense, is also about securing the optimal vehicle for large-scale AI application.

Thus, we see the AI world splitting into two distinct narratives: one is Meta’s, floating between grand visions and delayed realities, leaving people in a fog; the other is embodied by Jensen Huang, Volcano Engine, and even those giants quietly building Skill stores amid the fray—firmly grounded, stuffing AI into chips, software, games, cars, and every possible capillary of commercial value. Which story should investors and developers believe? Time and ultimately viable commercial ecosystems will provide the answer. Simply declaring "we take AI seriously" no longer fools anyone.

Meta又一次上演了经典的“画饼延期”戏码。说好的Muse Spark AI模型API，在开发者的期待清单里躺了又躺，至今连个确切的发布日期都挤不出来。发言人倒是轻飘飘一句“期待本月发布”，这话术和“下周回国”简直异曲同工。我们不禁要问：那个曾经在AI开源领域意气风发、大有“open-sourcing everything”之势的Meta，到底在犹豫什么？是技术遇到了真正的瓶颈，还是内部的资源争夺让产品路线图一改再改？当Google、OpenAI甚至各种初创公司都在疯狂迭代和发布时，Meta这种拖沓的节奏，消耗的不仅是开发者的耐心，更是自己那点来之不易的AI领军者信誉。开源承诺如果总是变成“即将开源”，那和微软过去的“拥抱、扩展、消灭”策略，在开发者眼里可能也没太大区别，都透着一股算计和不真诚。

相比之下，黄仁勋的行程表就显得目的明确且步伐务实。他亲自飞到韩国，与《绝地求生》的开发商Krafton会面，谈的是实实在在的RTX Spark芯片落地和“实体AI”合作。这里没有虚无缥缈的概念，而是瞄准了游戏——这个最能消耗算力、也最能直观展示AI硬件性能的产业。从数据中心到消费级笔记本，英伟达的触角正在变得无孔不入。老黄深谙一个道理：生态不是靠PPT讲出来的，是靠一个个具体的合作伙伴、一款款跑在硬件上的应用垒起来的。他跑去和游戏公司谈，本质上是在为英伟达的算力寻找下一个、也许也是更广阔的应用出口，把AI从云端拉到每个人的掌上设备。这种“把芯片塞进所有设备”的执行力，正是Meta这类更擅长描绘宏大叙事的公司所欠缺的。

把视线拉回国内，这24小时的热榜堪称一幅AI落地的浮世绘。微软喊出“16亿Windows用户一夜进入Agent时代”，这话听着热血沸腾，但细想之下，一个操作系统巨头的AI野心，终究要经受亿万个具体使用场景的检验。Agent到底是能真正自动化办公流程，还是又一个需要用户重新学习、甚至添乱的“智能”助手？市场大概率会给出冷酷的答案。

而火山引擎把MaaS（模型即服务）的年营收目标提至150亿，Seedance 2.0单月营收超10亿，则是另一种更扎实的信号。它意味着在中国市场，AI的商业化竞赛已经从“有没有模型”进入了“模型能赚多少钱”的硬核阶段。字节系凭借其巨大的流量池和应用生态，在把AI能力变现这件事上，确实跑出了速度。

更有趣的是腾讯、阿里、字节在“Skill商店”上的混战。这预示着AI的下一个战场，可能不再是比谁的底座模型参数多、基准测试分高，而是比谁能构建更丰富、更易用的AI能力插件生态。这就像智能手机时代的App Store之争，谁能让开发者更方便地创造“技能”，谁能让用户更顺畅地获取智能，谁就可能掌握下一代计算平台的入口。巨头们摩拳擦掌，抢的其实是“AI时代的App Store”。

至于那条“中国车为什么越造越大”的讨论，看似与AI无关，实则紧密相连。更大的车身意味着更复杂的内部空间，这为AI座舱、更沉浸的交互、以及自动驾驶所需的更多传感器提供了物理舞台。汽车正在从交通工具变成一个移动的智能空间，而这个空间的“灵魂”和“大脑”，正是AI。造大车，某种意义上也是在为AI的规模化应用抢占最佳载体。

所以你看，此刻的AI世界正分裂成两种鲜明的叙事：一种是Meta式的，悬浮在宏大愿景与延期现实之间，让人雾里看花；另一种是黄仁勋、火山引擎、乃至那些在混战中埋头搭建Skill商店的巨头们所代表的，坚定地向下扎根，把AI塞进芯片、塞进软件、塞进游戏、塞进汽车、塞进每一个可能产生商业价值的毛细血管里。投资者和开发者该相信谁？时间，和最终跑通的商业闭环，会给出答案。光有“我们很重视AI”的宣言，已经糊弄不了任何人了。