English Report
Rain taps against the glass facade of Taipei’s Nangang Software Park, while just a few kilometers away in Hsinchu Science Park, wafer fabs hum through the night. Taiwan’s semiconductor industry posted a stunning 29% revenue surge in Q1 2026, with full-year projections exceeding $270 billion—a figure that sounds like a victory march. But veterans of this industry know better: this is merely the eerie calm at the eye of a gathering storm.
The AI compute tsunami is reshaping everything. Indium Phosphide (InP) compound semiconductors are quietly shattering the bandwidth and power barriers that have long constrained silicon. This isn’t incremental progress—it’s a materials revolution. While Nvidia and Broadcom scramble for CoWoS packaging capacity, a select few are already betting on photonic integration and heterogeneous integration as the next frontier. But who controls InP? The answer today lies not in the U.S. or South Korea, but in overlooked European labs and Taiwan’s tightly knit academia-industry ecosystem. And once this technology becomes the “new oil” of AI infrastructure, will Washington really stand aside?
The chairman of Etron wasn’t crying wolf: the emerging U.S.-South Korea semiconductor alliance isn’t just about tech synergy—it’s about exclusionary supply chains. Seoul denies plans to distribute semiconductor tax windfalls to citizens (clearly reserving funds for industrial reinvestment) while accelerating fab linkages with Arizona. Behind this lies Samsung and SK Hynix’s strategic encirclement of TSMC—not through pricing, but geopolitical leverage. If Taiwan fixates only on revenue figures, it risks being institutionally marginalized without even noticing.
Yet Taiwan isn’t defenseless. Nan Ya PCB is aggressively targeting high-end IC substrates—the unsung “nerve endings” of AI servers. Once dominated by Japan’s Ajinomoto and Korea’s LG Innotek, this market is now yielding to Taiwanese players thanks to rapid prototyping and localized service. What does this tell us? Even as equipment and materials face choke points, Taiwan retains resilience in the capillaries of its manufacturing ecosystem.
What about mainland China? Hua Hong is pouring $6 billion into expanding its Wuxi facility, focusing on power devices, MCUs, and automotive-grade chips—segments deemed “unsexy” but forming the bedrock of genuine domestic substitution. SMIC, despite sliding operating profits due to mature-node price wars, beat margin expectations, proving its cost discipline and customer stickiness have matured significantly. Meanwhile, Applied Materials hit a 25-year high in profitability, repeatedly citing “agentic AI” in its earnings call. It turns out AI doesn’t just consume chips—it’s transforming equipment economics. When AI agents autonomously optimize etch parameters or predict chamber maintenance cycles, equipment vendors evolve from “shovel sellers” into “intelligent factory partners.”
History repeats itself. In the 1980s, Japan’s DRAM rise triggered U.S. retaliation. In the 1990s, Korea leveraged state capital to dominate memory. In the 2010s, Taiwan defined global logic chip manufacturing through pure-play foundry. Now, under AI’s new cycle, the rules have changed: technological leadership no longer guarantees security. You might lead in InP for three years—but if your fab sits on the “wrong” island, you could be ejected from the alliance tomorrow.
I judge that the next two years will accelerate the semiconductor industry’s de-globalization—not a simple U.S.-China decoupling, but the formation of multipolar blocs: a U.S.-Korea-Japan “trusted supply chain,” a China-centric internal loop, and Taiwan caught in between, struggling to maintain technical neutrality amid mounting pressure. Taiwan’s $270 billion revenue miracle, without strategic depth or diplomatic leverage, may ultimately become a gilded cage.
So as we watch the blue glow of fabs late into the night, perhaps we should ask: whose arena is this AI-driven chip feast? The algorithm engineers? The equipment giants? Or the geopolitical puppeteers? The answer might lie hidden in the defect density of the next InP wafer—where there are no borders, yet lines of power have already been drawn.
中文报道
台北的夜雨敲打着南港软三园区的玻璃幕墙，而就在几公里外的新竹科学园区，晶圆厂彻夜不眠。2026年第一季度，台湾半导体产业营收暴涨29%，全年有望突破2700亿美元——这数字听着像胜利号角，但老半导体人心里清楚，这不过是风暴眼中的短暂平静。
AI算力需求如海啸般席卷全球，InP（磷化铟）化合物半导体正悄然打破传统硅基芯片在带宽与功耗上的物理极限。这不是渐进式改良，而是一场材料革命。当英伟达和博通还在为CoWoS封装产能焦头烂额时，少数玩家已押注于光子集成与异质集成的下一代路径。问题是：谁掌握InP？目前答案不在美国，也不在韩国，而在那些被忽视的欧洲实验室与中国台湾的产学研链条中。可一旦这项技术成为AI基础设施的“新石油”，华盛顿会坐视不管吗？
Etron董事长的警告并非危言耸听：美韩半导体联盟正在成型，其目标不只是技术协同，更是供应链排他。首尔一边否认将半导体税收红利分给民众（显然要留作产业再投资），一边加速与亚利桑那州的Fab联动。这背后是三星与SK海力士对台积电的围猎——不是靠价格，而是靠地缘政治杠杆。台湾若只盯着营收数字，恐怕会在不知不觉中被“制度性边缘化”。
但台湾并非束手无策。南亚塑胶（Nan Ya PCB）正全力切入高端IC载板市场，这是AI服务器不可或缺的“神经末梢”。过去这块市场被日本味之素、韩国LG Innotek垄断，如今台系厂商凭借快速打样与本地化服务撕开缺口。这说明什么？在设备与材料被卡脖子的同时，台湾仍在制造生态的毛细血管里保有韧性。
大陆呢？华虹半导体豪掷60亿美元扩建无锡基地，聚焦功率器件、MCU与车规级芯片——这些看似“不够性感”的领域，恰恰是国产替代最扎实的战场。SMIC虽因成熟制程价格战导致运营利润下滑，却意外录得超预期毛利率，证明其成本控制与客户黏性已今非昔比。更值得玩味的是，应用材料（Applied Materials）创下25年最高利润率，其财报中反复出现“agentic AI”一词——原来AI不仅消耗芯片，也在重塑设备端的商业模式。当AI代理能自主优化刻蚀参数、预测腔体维护周期，设备厂商就从“卖铲人”变成了“智能工厂合伙人”。
历史总在重演。1980年代，日本DRAM崛起引发美国反制；1990年代，韩国以国家资本砸出存储霸权；2010年代，台湾靠专业代工定义全球逻辑芯片格局。如今轮到AI驱动的新周期，但这次游戏规则变了：技术领先不再等于安全。你可以在InP上领先三年，但如果Fab建在“错误”的岛屿上，明天就可能被踢出联盟体系。
我判断，未来两年将是半导体产业“去全球化”的加速期。不是简单的中美脱钩，而是多极化阵营的形成：一个以美国-韩国-日本为核心的“可信供应链”，一个以中国大陆为主体的内循环体系，以及夹在中间、试图保持技术中立却日益艰难的台湾。台湾的2700亿美元营收神话，若缺乏战略纵深与外交筹码，终将成为精致的囚笼。
所以，当我们在深夜看着晶圆厂的蓝光闪烁，不妨自问：这场AI驱动的芯片盛宴，究竟是谁的主场？是算法工程师？设备巨头？还是地缘政治操盘手？或许答案藏在下一个InP晶圆的缺陷密度里——那里没有国界，却早已被权力划线。