Galaxy Securities: AI Computing Demand Surge Drives Long-Term Storage Price Increases, HBM Wafer Share to Continue Rising 银河证券：AI算力建设需求爆发推动存储长周期涨价，HBM投片比重将持续提升

While everyone is focused on the parameter scales of large models, the real battlefield has quietly shifted to the stacked silicon chips behind the semiconductors. NVIDIA's upcoming Rubin Ultra platform pushes the HBM capacity of a single GPU to 384GB. This figure itself isn't shocking—the industry has long grown accustomed to exponential growth. But what's truly intriguing is that this path of "brute-forcing computational power with memory" is dragging the entire semiconductor supply chain into a new arms race.

Look at TrendForce's data: HBM's share of wafer starts will surge from 18% to 30% within two years. This isn't just technological evolution; it's clearly a high-stakes gamble across the supply chain. As TSMC's advanced packaging capacity becomes scarce and Samsung and SK Hynix frantically expand HBM production lines, we must ask: How long can this model of unlimited investment in AI infrastructure possibly last? More ironically, while edge AI keeps championing the slogan of "running models on devices," it turns out the memory architectures of smartphones and PCs are nowhere near keeping up with the brute growth of cloud computing power. So 3D stacked DRAM becomes the savior—but isn't this exactly the path traditional memory vendors should have taken long ago?

Industry analysts revel in HBM's bit supply share rising from 8% to 13%, yet selectively ignore one fact: this growth is almost entirely driven by North American and Korean players. China's role in this wave is likely still relegated to supporting the mid-to-low-end market with mature process nodes. While Yangtze Memory Technologies is still struggling to ramp up NAND flash production, international giants are already redefining the value of "memory" with HBM—transforming it from a cost item into a core premium component.

Perhaps the most absurd aspect is the disconnect in the technological narrative. On one hand, Wall Street hypes the "unlimited demand for AI computing power"; on the other, everyday consumers still experience lag just from closing a few background apps on their phones. A GPU with 384GB of memory sounds impressive, but how many applications truly requiring it have already been born? When training a hundred-billion-parameter model consumes tens of millions of kilowatt-hours of electricity, shouldn't we re-examine the sustainability of this path? Memory vendors are dancing to NVIDIA's tune, forgetting to account for another dimension of the ledger: energy, materials, and geopolitical costs.

3D stacking technology is indeed a promising route, but this path shouldn't be paved solely by AI demands. Medical imaging, industrial simulation, scientific computing—these fields that require large memory bandwidths have long been overlooked and should have garnered attention earlier. Unfortunately, capital only sees the AI bubble and remains blind to the real-world pain points. As memory capacity gets voraciously consumed by HBM, the rising prices of traditional DRAM have already begun to erode the consumer electronics market. It's eerily reminiscent of the Bitcoin mining GPU rush years ago, only this time the stakes have been upgraded to a technological competition that could shape national fortunes.

The semiconductor industry has always been cyclical, but this upcycle has been amplified by the AI narrative to the point of near-paranoia. When the tide recedes, will those production lines expanded for HBM become the next disaster of overcapacity? What edge AI truly needs, perhaps, isn't a larger memory pool but smarter memory management architectures and more efficient data movement technologies. Regrettably, amid the euphoria of stock prices and pitch decks, these "boring" foundational optimizations will never be the main act.

当所有人都在谈论大模型参数规模时，真正的战场早已悄悄转移到了芯片背后的那些堆叠硅片上。NVIDIA即将推出的Rubin Ultra平台将单颗GPU的HBM容量推至384GB，这个数字本身并不令人震惊——毕竟行业早已习惯了指数级增长。但真正值得玩味的是，这种“用内存暴力堆算力”的路径，正在把整个半导体产业链拖入一场新的军备竞赛。

看看TrendForce的数据：HBM投片比重将在两年内从18%飙至30%。这哪里是技术演进，分明是一场供应链的豪赌。当台积电的先进封装产能成为稀缺资源，三星、SK海力士疯狂扩建HBM产线时，我们不得不问：这种为AI基础设施无限投入的模式，究竟还能持续多久？更讽刺的是，端侧AI一边喊着“在设备上运行模型”的口号，一边却发现手机、PC的内存架构根本跟不上云端算力的野蛮生长。于是3D堆叠DRAM成了救世主——可这不正是传统存储厂商早就该走的路吗？

行业分析师们津津乐道于HBM位元供给占比从8%到13%的提升，却选择性忽视了一个事实：这种增长几乎完全由北美和韩国厂商驱动。中国存储芯片在这场浪潮中扮演的角色，恐怕依然是“用成熟制程给中低端设备供货”的配角。当长江存储还在为NAND闪存的产能爬坡挣扎时，国际巨头已经在用HBM重新定义“存储”的价值——从成本项变成了溢价核心。

最荒诞的或许是技术叙事的割裂。一边是华尔街鼓吹“AI算力需求无限”，一边是普通消费者连手机后台杀几个应用都会卡顿。384GB的GPU显存听起来很美，但有多少真正需要它的应用已经诞生？当训练一个千亿参数模型需要烧掉数千万度电时，我们是否该重新审视这条路径的可持续性？存储芯片厂商们跟着英伟达的指挥棒狂欢，却忘了计算另一个维度的账本：能源、材料和地缘政治成本。

3D堆叠技术确实是条明路，但这条路不该只由AI需求来铺就。医疗影像、工业仿真、科学计算——这些需要大内存带宽却长期被忽视的领域，本应更早获得关注。可惜资本只看得到AI的泡沫，看不见真实世界的痛点。当存储产能被HBM疯狂吸走时，传统DRAM的涨价已经开始侵蚀消费电子市场。这像极了多年前比特币挖矿显卡抢购的剧本，只不过这次赌注换成了国运级别的科技竞赛。

半导体行业从来都是周期性的，但这次的上行周期被AI叙事加持得近乎偏执。当潮水退去，那些为HBM扩建的产线会不会成为下一个产能过剩的灾难？而端侧AI真正需要的，或许不是更大的显存池，而是更聪明的内存管理架构和更高效的数据搬运技术。可惜在股价和PPT的狂欢中，这些“无聊”的基础优化永远不会是主角。