TSMC's C.C. Wei: CoPoS May Reach Volume Production Within Two Years 台积电魏哲家：CoPoS最快两年放量

CoWoS has already elevated NVIDIA's GPUs to legendary status, becoming the entry ticket for the AI training era. CoPoS, a technology that co-packages photonic chips and electronic chips over a larger area, is seen as key to supporting next-generation ultra-large-scale AI computing—and even the convergence of optical computing. The "two to three years" timeline offered by C.C. Wei is a sober assessment from an industry giant grounded in complex engineering realities. It signals that the next physical breakthrough in computing power is not about tomorrow's keynote slide deck, but about over a thousand days of yield ramp-ups, material challenges, and supply chain restructuring. This "slowness" contrasts sharply with the "speed" expected by capital markets and the "instant revolution" hyped by tech media. TSMC is using the most straightforward methods to build the deepest moat—accumulated time and craftsmanship—this is the true trump card of a giant.

In contrast, the current public discourse is filled with impatient excitement. Look at those headlines: "Forget GPUs—Intel Unveils a Heavyweight Move to End NVIDIA’s Computing Monopoly?" and "1.6 Billion Windows Users Storm into the Agent Era Overnight." These headlines themselves are a crude simplification of the complexity of technological evolution. No matter how "heavyweight" Intel's move may be, it is still chasing a moving target—from chip design to foundry capacity, from software ecosystems to customer trust—none of which can be broken through with a single "masterstroke." And "1.6 billion users entering the Agent era overnight"? That sounds more like a marketing incantation. OS-level integration is a necessary foundation, but true agents require trustworthy permissions, mature models, and entirely new interaction paradigms—this is far from an overnight achievement. Compressing the complex democratization of technology into a dramatic "overnight" event serves no purpose beyond generating clicks and may even mislead the public’s understanding of how technology actually progresses.

By comparison, Liftoff’s IPO news is far more straightforward. $23 per share, 19 million shares—a standard business maneuver. Unlike TSMC’s declaration, which concerns the lifeblood of global AI infrastructure, or those trending headlines attempting to define the shape of the future, this is merely a capital operation by a mobile advertising tech company within an established track. This "ordinariness," amid a flood of words like "disruption," "revolution," and "termination," carries a refreshing clarity. It reminds us that besides gazing at grand narratives in the sky, countless concrete businesses on the ground are operating according to economic principles.

Volcengine’s MaaS target upgrades, the skirmishes among Tencent, Alibaba, and ByteDance over Skill stores, the trend of building ever-larger cars in China, the rise of "upscale casual dining" in lower-tier markets… each trending topic represents a fiercely contested battlefield. Yet beneath the heat lies technological convergence, model imitation, and traffic anxiety. When giants are all chasing the same hot topics and packaging similar narratives, true innovation often occurs in unsung laboratories, on TSMC’s pilot lines that require two or three years to "ramp up."

Our era is not short of thunderous slogans—it lacks the patience to lay a solid foundation. It is not short of promises to "storm" into a new era overnight—it lacks respect and patience for the technology ramp-up phase. While everyone fixates on GPUs and model parameters, the real bottleneck may lie in packaging, materials, and those slowly growing trees of underlying process technology. C.C. Wei’s "volume production in two years" serves as a cold shower for all impetuous mindsets: great engineering takes no shortcuts, and the accumulated momentum of technology has never been an instantaneous event on trending charts. Perhaps with fewer fantasies about "ending monopolies" and more attention to the quietly extending pilot line, we can finally see the true path to the future.

魏哲家在股东会上轻描淡写的一句话，才是这场AI算力狂欢里最硬的骨头：“CoPoS最快两年放量”。这位台积电的掌门人没有用任何激动人心的词汇，只是平静地陈述了一个试产线已建、但量产需待时日的事实。这恰恰是当前科技叙事中最稀缺、也最令人不安的东西——诚实。当整个行业都在为明天的“颠覆”和“爆发”狂欢时，真正的基石正在地基里，一毫米一毫米地浇筑。

CoWoS已经让英伟达的GPU封神，成为AI训练时代的入场券。而CoPoS，这个将光子芯片与电子芯片在更大面积上共封装的技术，被视为支撑下一代超大规模AI计算、乃至光计算融合的关键。魏哲家给出的“二至三年”，是一个产业巨头基于复杂工程现实给出的冷静评估。它意味着，算力的下一个物理极限突破，不是明天的发布会PPT，而是未来1000多个日夜的良率爬坡、材料攻坚和供应链重塑。这种“慢”，与资本市场期待的“快”、与科技媒体渲染的“瞬时革命”，形成了刺眼的对比。台积电在用最笨的办法，构筑最深的护城河——时间与工艺的沉淀，这才是巨头真正的底牌。

反观此刻的舆论场，充斥着急不可耐的亢奋。你看看那些新闻标题：“别盯着GPU了，英特尔甩出重磅大招，能否终结英伟达的算力垄断？”、“16亿Windows用户，一夜冲进Agent时代”。这些标题本身，就是一种对技术演进复杂性的粗暴简化。英特尔扔出再大的“大招”，也只是在追赶一个不断移动的靶心，从芯片设计到代工产能，从软件生态到客户信任，哪是“大招”能一击必破的？而“16亿用户一夜进入Agent时代”？更像一句营销咒语。操作系统层面的集成是必要的铺垫，但真正的智能体（Agent）需要可信赖的权限、成熟的模型和全新的交互范式，这绝非一蹴而就。将复杂的技术民主化进程压缩成“一夜之间”的戏剧性，除了赚取点击，毫无益处，甚至会误导公众对技术发展规律的认知。

相比之下，Liftoff那条IPO新闻就实在得多。每股23美元，1900万股，一个标准的商业动作。它不像台积电的宣言关乎全球AI基础设施的命脉，也不像那些热榜标题试图定义未来的形状。它只是移动广告技术公司在一个既定赛道里的一次资本运作。这种“平常”，在充斥着“颠覆”、“革命”、“终结”字眼的信息流中，反而透着一股清澈。它提醒我们，除了仰望星空的宏大叙事，地面上还有无数具体的生意在按照经济规律运转。

火山引擎的MaaS目标提升、腾讯阿里字节混战Skill商店、中国车越造越大、下沉市场的“漂亮饭”……热榜上的每一项都代表了一个火热的战场。但火热之下，是技术趋同、模式抄袭和流量焦虑。当巨头们都在追逐同一批热点、包装相似的叙事时，真正的创新往往发生在无人喝彩的实验室里，发生在台积电那些需要两三年才能“放量”的试产线上。

我们这个时代不缺惊雷般的口号，缺的是把地基打稳的耐心。不缺一夜之间“冲进”某个新时代的许诺，缺的是对技术爬坡期的尊重与等待。当所有人都盯着GPU和模型参数比拼时，真正的瓶颈或许正在封装、在材料、在那些缓慢生长的底层工艺树上。魏哲家的“两年放量”，是对所有浮躁心态的一次冷处理：伟大的工程没有捷径，技术的厚积薄发，从来都不是热搜榜上的瞬时事件。少一些“终结垄断”的意淫，多关注一下那条安静延伸的试产线，或许我们才能看清通往未来的真正路径。