English Report
Reading through earnings reports and press releases late at night, I can’t shake the feeling that the semiconductor industry is staging a meticulously choreographed “AI frenzy play”—while the audience in the front rows has quietly changed seats.
Apple once dictated memory pricing. With its annual procurement of hundreds of millions of devices, it could squeeze Samsung and Micron until they gasped for air. But now? Korean media puts it bluntly: amid the AI-driven memory supply crunch, Cupertino’s bargaining power is evaporating. This isn’t a technical issue—it’s a structural shift in power. When AI servers demand HBM3E instead of LPDDR5, and when training cluster procurement cycles dwarf consumer electronics’ quarterly rhythms, Apple gets pushed from the center of the table to the periphery.
Meanwhile, China’s OSATs—JCET, TFME, and others—are sharpening their knives. Once relegated to the back end of the supply chain, they’re now stepping into the spotlight thanks to AI chips’ insatiable hunger for advanced packaging. TSMC’s CoWoS capacity is maxed out; its expansion can’t keep pace with Nvidia’s order surge. Chinese players smell opportunity. They may not replicate TSMC’s process overnight, but even capturing 10% of overflow demand could send their revenue curves soaring. It reminds me of how TSMC pulled ahead of UMC after the 2008 financial crisis—disruption has always been the catalyst for realignment.
CXMT’s recent profit surge appears, on the surface, as a triumph of China’s DRAM self-reliance drive. But look closer: this is largely a cyclical rebound amplified by patriotic sentiment. The company remains locked out of mainstream markets, unable to enter high-end servers, let alone challenge Micron or SK Hynix’s technological moats. In my view, CXMT’s “success” precisely reveals the ceiling of China’s memory ambitions: without ecosystem synergy, isolated breakthroughs remain islands.
Then there’s Nvidia. Jensen Huang’s H200 may be the last high-performance AI chip legally sellable in China. U.S. export controls tighten by the month—Elon Musk himself has publicly complained about insufficient compute access. The irony? This very “chokehold” accelerates the fragmentation of the global supply chain. Chinese customers pivot to Ascend and Cambricon, while Nvidia funnels more capacity to North American cloud giants. I predict that within 18 months, we’ll see two parallel AI chip ecosystems: one CUDA-centric, the other built around Huawei’s CANN. Incompatible, yet equally expensive.
Intel’s move is telling. It’s abandoning mid- and low-end CPUs to double down on high-end Xeons. On paper, it makes sense—AI inference and training do require robust general-purpose compute. But here’s the rub: as AMD eats into its datacenter share with Zen architectures, and as ARM chips like Ampere and AWS Graviton quietly infiltrate, could Intel’s “premium focus” become a desperate all-in bet?
Don’t overlook Phison. This Taiwanese controller specialist just raised $800 million in overseas bonds to expand its storage footprint. Why? Because it understands: AI isn’t just a GPU story—it’s a data-movement story. SSD controllers, UFS, enterprise NVMe—every bit of training data needs a high-speed highway. Phison’s gamble is actually a precise bet on the “second layer” of AI infrastructure.
Seoul’s anxiety isn’t unfounded. South Korea is sounding alarms over manufacturing hollowing-out. While Samsung and SK Hynix dominate global memory markets, AI server assembly, advanced packaging, and even segments of logic chips are increasingly concentrated in Taiwan. TSMC, ASE, Foxconn—they form the new AI hardware triangle. Korea retains strength in materials and equipment, but it’s falling behind in system integration. It echoes Japan’s semiconductor decline in the 1990s: being a category champion doesn’t guarantee systemic victory.
So, back to the original question: who’s really paying for this AI frenzy?
Not Nvidia—its margins keep climbing. Not TSMC—its fabs run at full throttle. Not even the Chinese government—it’s willing to pay a premium for technological sovereignty.
The real bill-payers are the second-tier players squeezed out of the supply chain, the traditional IDMs that misread the tech trajectory, and the consumer electronics giants who assumed scale alone would protect their throne. AI isn’t a democratizing force—it’s a brutal filter. In this selection process, speed, ecosystem cohesion, and geopolitical agility matter more than transistor density.
While everyone obsesses over the compute arms race, perhaps we should ask: if the AI bubble bursts tomorrow, whose balance sheet collapses first?
中文报道
深夜翻看财报和新闻稿，总有一种错觉：整个半导体行业正在上演一场精心编排的“AI狂欢剧”，而台下观众却早已悄悄换人。
苹果曾经是内存市场的定价者。靠着每年上亿台设备的采购量，它能压得三星、美光喘不过气。可现在呢？韩国媒体直言不讳：在AI驱动的存储供应紧缩面前，库比蒂诺的议价权正在蒸发。这不是技术问题，而是结构性权力转移——当AI服务器需要的是HBM3E而不是LPDDR5，当训练集群的采购节奏碾压消费电子的季度周期，苹果就从牌桌中央被推到了边缘。
与此同时，中国封测厂（OSAT）正摩拳擦掌。长电、通富微电这些名字过去只出现在产业链后段，如今却因AI芯片对先进封装的渴求而跃居舞台中央。CoWoS产能吃紧，台积电扩产速度赶不上英伟达的订单增长，于是大陆厂商嗅到了机会。他们未必能立刻复制TSMC的工艺，但哪怕分到10%的溢出需求，也足以让营收曲线陡峭向上。这让我想起2008年金融危机后，台积电如何借机甩开联电——危机从来都是重构格局的催化剂。
合肥长鑫（CXMT）最近利润暴涨，表面看是中国DRAM自主化的胜利。但细看数据就知道，这更多是周期性反弹叠加国产替代情绪的结果。它仍被锁在主流市场之外，无法进入高端服务器领域，更别说挑战美光或SK海力士的技术护城河。我认为，CXMT的“成功”恰恰暴露了中国存储产业的天花板：没有生态协同，单点突破终究是孤岛。
再看Nvidia。黄仁勋的H200成了在中国能合法销售的最后一代高性能AI芯片。美国出口管制步步紧逼，特斯拉的马斯克甚至公开抱怨买不到足够算力。讽刺的是，正是这种“卡脖子”加速了全球供应链的分裂——中国客户被迫转向昇腾、寒武纪，而英伟达则把更多产能留给北美云巨头。我判断，未来18个月，我们将看到两个平行的AI芯片生态：一个以CUDA为中心，另一个以华为CANN为内核。它们互不兼容，却同样昂贵。
英特尔的选择耐人寻味。它放弃中低端CPU，全力押注高端至强处理器。这看似明智——AI推理和训练确实需要更强的通用算力支撑。但问题是，当AMD凭借Zen架构蚕食其数据中心份额，当ARM架构通过Ampere和AWS Graviton悄然渗透，英特尔的“高端聚焦”会不会变成孤注一掷？
别忘了Phison（群联）。这家台湾主控芯片厂竟跑到海外发8亿美元债券，只为加码存储布局。为什么？因为它看懂了：AI不只是GPU的故事，更是数据搬运的故事。SSD控制器、UFS、企业级NVMe——每一比特训练数据都需要高速通道。Phison的豪赌，其实是对“AI基础设施第二层”的精准押注。
而首尔的焦虑并非空穴来风。韩国制造业正面临“空心化”警报。三星和SK海力士虽在全球存储市场呼风唤雨，但AI服务器整机制造、先进封装、甚至部分逻辑芯片，却越来越集中在台湾。台积电、日月光、鸿海——这些名字构成了AI硬件的新三角。韩国拥有材料和设备优势，却在系统整合能力上掉队。这让人想起90年代日本半导体的衰落：单项冠军不等于体系赢家。
所以，回到最初的问题：谁在为这场AI狂潮买单？
不是英伟达——它的毛利率还在攀升；不是台积电——它的产能永远满载；甚至不是中国政府——它愿意为技术自主支付溢价。
真正埋单的，是那些被挤出供应链的二线玩家，是那些误判技术路线的传统IDM，是那些以为靠规模就能守住地位的消费电子巨头。AI不是普惠技术，它是一场残酷的筛选机制。在这场筛选中，速度、生态、地缘政治敏感度，比晶体管密度更重要。
当所有人都在谈论算力军备竞赛时，或许该问一句：如果明天AI泡沫破裂，谁的资产负债表最先崩盘？