2025-12-27 · 供应链分析

InP衬底成AI光互连瓶颈，$AXTI/$SMTOY掌控命脉，2026年或致供应链危机。

涉及标的： $AMZN $AVGO $AXTI $COHR $GOOGL $LITE $META $MRVL $MSFT $NVDA $SMTOY

中文翻译

“磷化铟(InP)瓶颈”：AI基础设施建设的关键瓶颈解析： $NVDA Blackwell、$META MTAI、$GOOGL TPU 和 $MSFT Maia 的产能爬坡，其未来取决于两家市值仅7亿美元的小盘股 $AXTI 和 $SMTOY。如果无法解决 InP 问题，AI“增长”故事将在2026年终结。原因如下： AI行业正开始向光子学迁移，以部署未来的 ASIC/GPU，因为铜互连已触及物理极限。然而，超大规模云服务商在这样做时，却将命运押注在了 InP（磷化铟）这一通用材料上，而全球仅有少数工厂能以激光器所需的纯度生产6英寸 InP 晶圆。以 Google 及其 TPU v7 Ironwood 项目为例： Google 使用光电路交换(Optical Circuit Switching, OCS)，简单来说就是由光构成的交换机。Pod 中的每一个 TPU 进行通信都需要基于 InP 的激光器。与 Google 合作的 $LITE 在很大程度上依赖 InP 衬底（如 AXT/住友）来制造这些激光器。如果没有这些衬底，Google 的整个 Ironwood 项目不仅仅是“放缓”，而是会直接撞墙。来自 $NVDA GB 系列、$AMZN Trainium、$MSFT Maia、$META MTAI 的现代 ASIC/GPU 都做出了同样的押注：光是未来的方向。现在，问题出现了。整个西方 AI 路线图目前都系于一家市值7亿美元的小盘股和一家单一的日本公司，它们生产了全球光子学所需的大部分 InP 衬底。目前这是一个双寡头格局（粗略估计 AXT + 住友供应约60%），最近的估计显示约70%+的供应来自住友电气、AXT、Freiberger、JX 和 Visual Photonics Epitaxy（填补剩余缺口）。无论如何，整个未来的 AI 供应链细如针尖： - Moomoo 研究：InP 市场处于“全球争夺”和“严重供应短缺”状态，NVIDIA GB200 的推出（扩展仍需大量 InP，不仅仅是机架内的 NVL72 通信）。 - 高速收发器的需求目前可能超过供应近两倍（LightCounting） - 看到创纪录的预订，但明确“受限于 InP 激光器的供应”（$COHR CEO Q3 财报电话会） - 麦肯锡：800G 模块存在 40% 到 60% 的缺口，1.6T 模块存在 30% 到 40% 的缺口。鉴于需求的激增，这些报告可能理解得过于保守。即使按照微软对 Maia 爬坡的预测（据 UBS $MRVL 笔记，2027年 Maia 估计超过100万台），加上未来一年超过200万台 1.6T 收发器，这一体量如此之大，代表了全球衬底产量的双位数百分比。 AI 预期的“指数级增长”即将与关键材料生产的现实发生碰撞。因此，“Ironwood”、“MTIA”和“Maia”的爬坡不仅雄心勃勃，在当前材料限制下可能根本不可能实现。即使 $COHR、JX 日本、住友、$AXTI 等以最大产能扩产（例如 $COHR / JX -> 6英寸 InP 晶圆产能提升4倍），他们可能仍无法满足超大规模云服务商日益增长的需求。特别是考虑到需求激增，例如仅 $NVDA 自身（GB200/GB300 修订版）的需求。硅光子学等技术解决方案可以弥合差距，但这仍然主要需要外置 InP 激光器作为光源。TFLN 或量子点激光器还需要很多年才能成熟。未来几年可能无法摆脱对 InP 的需求。因此，芯片设计与材料可用性之间的错配创造了一个战略瓶颈，在供应链的最底层，极少数公司控制了大部分配额、定价和供应。当叠加美中关系紧张和出口管制的地缘政治风险时，这尤其危险。话虽如此，以下是可能发生的事情： - 价格飙升：$AXTI、JX、住友的价格将大幅飙升 -> $LITE、$COHR、中际旭创（也会通过下游传导提高价格） - 超大规模云服务商将直接囤积材料，绕过传统组件采购，直接从 $AXTI、JX 日本、住友购买 InP 衬底库存，并直接委托给像 $LITE 这样的收发器制造商。（例如，Meta 将绕过收发器公司，直接去找 AXT 或住友） - 超大规模云服务商将买断生产配额（就像 $NVDA 已经积极“锁定”EML 产能（在 InP 衬底上制造）那样）。购买衬底制造商或生产配额将成为一种必要，以免被 $NVDA 或 $GOOGL 等竞争对手饿死。随着 TPU v7 和其他超大规模云服务商在 2026-2027 年爬坡，我们可能会进入衬底的“饥饿游戏”阶段，每个超大规模云服务商都将为了资源配额而相互吞噬对方的增长。像 $NVDA（拥有创纪录的预分配量）这样的公司可能暂时没事，但其他项目可能会面临重大延误。思考： 1. 一些超大规模云服务商可能没事（$NVDA）。其他如 $GOOGL 和 $MSFT 需要在其他人之前买断材料和配额。 2. 行业需要加倍投入工程转型，如延长铜的使用寿命和更节省材料的方式如硅光子学(SiPh)。 3. 转向6英寸晶圆以提高良率（能缓解情况，但仍不足以满足需求）因此，现状是，数万亿美元的 AI 扩展系于一些不起眼的7亿美元公司 $AXTI 和 $SMTOY。除非架构改变，否则 AI 似乎不可避免地会因 InP 衬底产能而触及物理天花板。 2024年的瓶颈是 GPU。2025年是 HBM。2026年，主要约束很可能是光互连，特别是驱动它们的 InP 衬底。这已成为 AI 基础设施建设中隐藏的瓶颈。

英文原文

The "InP Chokepoint": The Bottleneck of the AI Buildout explanation: The future of $NVDA Blackwell, $META MTAI, $GOOGL TPU, and $MSFT Maia ramp is tied to: A $700M small cap $AXTI and $SMTOY. The AI "Growth" story ends in 2026 if there's no solution to InP. Here's why: The AI industry started its migrating to photonics for future ASIC/GPU deployments, because copper is hitting a physical limit. However, in doing so, hyperscalers traded the common material for InP (Indium Phosphide), when there's only a few factories capable of producing 6-inch InP wafers at the purity levels required for lasers. Let's take for example Google and their TPU v7 Ironwood program: Google uses Optical Circuit Switching (OCS), in simpler terms, switchboards made of light. For every one of those TPUs in the pod to talk, they require InP-based lasers. $LITE, which works with Google on this, largely depends on InP substrate (eg. AXT/Sumitomo) to make them. If they don't have it Google's entire Ironwood program doesn't just "slow down", it hits the wall. Modern ASICs/GPUs from $NVDA GB series, $AMZN Trainium, $MSFT Maia, $META MTAI have all made the same bet: Light is the way forward. Now, here's the issue. The entire Western AI roadmap is currently tethered to a $700M small-cap and a single Japanese company that produce majority of the world's InP substrates required for photonics. It's currently a duopoly (rough estimates majority supply ~60% between AXT + Sumitomo), with recent estimates of ~70%+ coming from Sumitomo Electric, AXT, Freiberger, JX, and Visual Photonics Epitaxy (filling in the gaps). Regardless, the entire future AI supply chain is thinner than a needle: - Moomoo Research: InP market is in a state of "global scramble" and "serious supply shortage" NVIDIA GB200 rollout (scale-out still requires tons of InP, not NVL72 within-the-rack comm). - Demand for high-speed transceivers today probably exceeds the supply by almost a factor of two (LightCounting) - Seeing record booking, but explicitly "supply-constrained by InP lasers" ( $COHR CEO Q3 ER) - McKinsey: 40% to 60% shortfall for 800G modules and a 30% to 40% shortfall for 1.6T modules. And these reports are likely understanding + very conserative given the demand ramp. Even going off Microsoft's projections on Maia ramp, ( est. 1M+ Maia by 2027 on UBS $MRVL note), with 2 million+ units of 1.6T transceivers over the next year, this volume is so large it represents a double-digit percentage of global substrate output. The projected "exponential growth" of AI is about to collide with the reality of critical material production. So, the "Ironwood", "MTIA" and "Maia" ramps aren't just ambitious, they may be impossible under current material constraints. Even if $COHR, JX Nippon, Sumitomo, $AXTI, and others, ramp up at maximum capacity (eg. $COHR / JX -> 6-inch InP wafers for 4x capacity), they still might not be able to meet the increasing demand from hyperscalers. Especailly with demand spikes occurring, eg. just for $NVDA alone (GB200/GB300 revisions). There are technical solutions like silicon photonics is one solution to bridge the gap, but this still largely requires an external InP laser as the light source. TFLN or quantum dot lasers are many many years away. There's probably no escaping the InP requirements for the next few years. So, the mismatch between chip design and material availability has created a strategic chokepoint, where if you go to the very bottom of the supply chain, very few companies control a majority of allocations, pricing, and supply. This is especially dangerous when compounded with geopolitical risks on US/China relations + export controls. That being said, here's what's probably what's going to happen: - Price Spikes: Prices from $AXTI, JX, Sumitomo will spike significantly -> $LITE, $COHR, Innolight (also increases prices from pass down) - Hyperscalers will directly stockpile materials, bypassing traditional component procurement and buying InP substrate inventory from $AXTI, JX Nippon, Sumitomo, and directly to consign to transceiver manufacturers like $LITE. (eg. Meta would bypass transceiver companies and go directly to AXT or Sumitomo) - Hyperscalers will buy out production allocation ( like $NVDA that has already aggressively "locked in" EML capacity (manufactured on InP substrates). Buying a substrate manufacturer or production allocation would become a necessity to so others like $NVDA or $GOOGL doesn't starve them out. As TPU v7 and and as other hyperscalers ramp up in 2026-2027, we will likely enter a "hunger games" phase for substrates where only each hyperscaler will be cannibalizing each other's growth for resource allocation. Companies like $NVDA (with record amounts of pre-allocation), might ramp be okay for the time being, but others programs would likely face major delays. Thoughts: 1. Some hyperscalers might be fine ( $NVDA). Others like $GOOGL and $MSFT will need to buy out materials + allocation before others do. 2. Industry needs to double down on engineering shifts like copper life extension and more material efficient ways like SiPh. 3. Move to 6-inch wafers for yields (eases things, but still not enough to meet demand) So the way things are now, the multi-trillion dollar AI scaling are tethered to some obscure $700m company $AXTI and $SMTOY. It seems inevitable that AI will hit the physical ceiling because of InP substrate capacity unless architectures change. In 2024, the bottleneck was GPUs. In 2025, it was HBM. In 2026, the primary constraint will likely be the optical interconnect, and specifically, the InP substrates that power them. This has now become the hidden bottleneck of the AI buildout.

在 X 上查看原推 ↗