· 供应链分析
观点中肯，但有两点误解： - 美国AI基础设施建设必须使用磷化铟(InP)，这是不可妥协的。我们近期看到“交易”案例：$NVDA GPU换取解除关键材料/原料/衬底禁令。美中虽为最大地缘政治对手，却不幸相互依赖。 - 尽管今日小幅回调，7N铟在SMM已创历史新高，而光子学(CPO)产能爬坡甚至尚未开始。你提到硅光(SiPh)替代，但无法替代InP。定制芯片可能采用硅+InP激光器的混合方案，但仍需InP。 - 需求呈指数级增长，$NVDA已锁定外延量子阱激光器(EML)产能，$MSFT Maia预计将占据两位数比例的InP供应链。我高度怀疑库存足以让所有人持续获取。 - 当然，或者他们可以通过涨价来弥补良率损失。 - 这不是垂直整合能解决的。整个供应链是物理/材料问题，因为大部分资源仅在中国可及。 超大规模云厂商可以垂直整合掉$LITE，但无法整合掉原始前驱体或材料。
Fair points and two misunderstandings: - US requires InP for AI buildout. This is non-negotiable. And we saw that "trade", $NVDA GPUs for unblocking critical materials/feedstock/substrates bans recently. US and China unfortunately are dependent on one another despite being the biggest geopolitical adversaries. - We're already seeing 7n indium reach ATHs on SMM (depsite recent dip today), and photonics ramp hasn't even started the curve yet. You say SiPh substitution, but you can't substitute away InP. Custom chips might use hybrid w/silicon + InP laser but you still need it. - It scales up exponentially, $NVDA secured EML capacity, $MSFT maia is expected to take up double digits of inp supply chain. Highly doubt there's enough inventory for everyone to keep it accessible. - Sure, or they can price hike to make up for yield - This is not something you can vertically integrate. The entire supply chain is a physics/materials problem because majority is only accessible in China. Hyperscalers can vertically integrate away $LITE but not the raw precursors or materials.