2026-01-09 · 供应链分析

HBM 瓶颈导致传统建模失效，磷化铟成新瓶颈，$AXTI 因卡位 AI 建设具极高价值。

中文翻译

当出现像 $MU 或 SK 海力士的 HBM（高带宽内存）这样极不常规的瓶颈时，诸如 NTM（下一财年）这样的标准建模就会失效。目前 HBM 的需求极度缺乏弹性，买家不再询问价格，即使在价格飙升数百%后，他们仍在寻求配额分配。在这种情况下，基于 $NVDA 2025 年上半年 EML（边缘机器学习）的现有短缺，以及 $MSFT 等公司仅其项目就需要全球两位数的磷化铟(InP)产量（针对 2026 年下半年至 2027 年），$AXTI 很可能在接下来几个月成为下一个瓶颈。因此，当 $GOOGL 等超大规模云服务商为避免 TPU 项目停滞而迫切寻求配额时，原本数亿 TAM（总可寻址市场）可能迅速膨胀至数十亿甚至上百亿，这使得对瓶颈的建模变得不可能。这本质上是买入“比特分配”、原料控制和瓶颈环节。一家市值 13 亿美元的公司卡住了整个 AI 建设的脖子，在我看来很便宜。我不知道这会走向何方，但我预计会出现极大的价格挤压（我们在 SMM 上已看到迹象），这将增加 $AXTI 的底线利润。许可证波动一直是一个风险因素。但同样，任何延迟都会因为这是单点故障而卡住整个 AI 建设，前提是如果中国继续对日本实施出口管制。

英文原文

Standard modeling like NTM breaks when there's extremely unconventional bottlenecks like HBM with $MU or Sk Hynix. HBM demand is so inelastic rn that buyers aren't asking for price, they're asking for allocation even after hundreds of percent increases. And in this case InP and $AXTI will likely be the next bottleneck in the upcoming months, just based on existing shortages from $NVDA EML H1 2025 and how others like $MSFT require the world's double digit inp output just for their program for H2 2026 into 2027. So it's impossible to model bottlenecks where few hundred million TAM previously might squeeze to few billion or tens of billions when hyperscalers like $GOOGL are desperate for allocation so their TPU program doesn't stall out. This is buying into bit-allocation, feedstock control, and bottlenecks. A $1.3B company bottlenecking the entire AI buildout looks cheap to me. I don't know where this is heading, but I do expect an extremely large price squeeze (we're seeing it now on SMM), which just increases the bottom line of $AXTI. License volatility was always a risk factor. But again, any delays would just bottleneck the entire AI buildout since this is single point of failure if China continues their export controls on Japan.

在 X 上查看原推 ↗