2026-01-09 · 个股论点

$AXTI因出口管制成InP垄断者，控制AI光子学双瓶颈，作者买入。

涉及标的： $AXTI $DOWA $GOOGL $META $MSFT $MU $NVDA $SNDK

中文翻译

为了提供一些背景：存在两个不同的瓶颈，$AXTI 正处于这两个瓶颈之间。磷化铟(InP) 原料——双寡头垄断（至关重要，$AXTI 与中国控制着约 78%-80% 的供应）。像 $DOWA 这样的公司相对产量极低。磷化铟(InP) 衬底——垄断（$AXTI）。此前是双寡头（$AXTI 和住友电气 Sumitomo Electric），占据 60-70% 的 InP 衬底产量，其他日本公司如 JX 贡献的量微乎其微。然而，两天前实施的出口管制限制了向住友及生产 InP 衬底的日本公司供应原料和关键材料，情况发生了改变。他们虽有库存，但随着出口管制持续，他们将缺乏大规模生产衬底所需的材料，$AXTI 将在 3-6 个月内成为垄断者，这就是上述框架的由来。这不是数量游戏。这是基于有限材料的价格挤压博弈论。$AXTI 拥有从矿产开采、精炼到衬底生产的全链条供应链及成本结构。这将转化为利润率。你之前看到的是数亿美元的市场规模(TAM)，现在仅因 $GOOGL TPU、$MSFT Maia、$NVDA、$META 等光子学 AI 建设所需的纯粹必要需求，就可能膨胀至 100 亿美元以上。这家市值 10 亿美元的公司控制了 AI 建设的两个不同瓶颈。我像早期在 HBM 领域买入 $MU 和 $SNDK 那样买入瓶颈环节，我不试图说服你也这样做，只是分享我的投资论点。

英文原文

Just to give some perspective: There's two different bottlenecks. $AXTI sits in the middle of two. InP Feedstock - Duopoly (Vital, $AXTI with China controlling ~78%-80% of supply). Companies like $DOWA produce extremely low amounts relatively. InP Substrates - Monopoly ( $AXTI ). Formerly it was a duopoly ( $AXTI, Sumitomo Electric) with 60-70% of InP substrate production. With misc Japanese companies like JX contributing much tinier amounts. However this changed with export controls 2 days ago restricting feedstock and critical material supply to Sumitomo and Japanese companies that produce InP substrates. They have inventory, but with the export controls continuing, they will lack the materials to produce substrates at scale and $AXTI becomes the monopoly in 3-6 months, hence the framing. This is not a quantity game. This is game theory on price squeezes based on limited materials. $AXTI owns the mineral mines all the way to refining -> substrate production and the supply chain + cost structure. This translates into margins. You were looking at few hundred million TAM that could now inflate to $10B+ just due to pure necessary required for $GOOGL TPUs, $MSFT maia, $NVDA, $META, and so no AI buildout with photonics. This $1B company controls two different bottleneck for the AI buildout. I'm buying into the bottleneck like what happened at the start with HBM with $MU and $SNDK, I'm not trying to convince you to as well, just sharing my thesis.

在 X 上查看原推 ↗