供应链分析
产业链结构、上下游与瓶颈环节研究 · 共 1315 条 · 滚到底自动加载更多
-
Cantos因光子学在ASIC部署中的关键作用，估值获溢价。
很棒的主题选择。乍一看，Cantos 似乎特别走在风险投资(VC)投资曲线的前面。虽然不清楚这家初创公司的具体细节，但仅在过去几个月里，其估值倍数可能又获得了一次溢价。尤其是考虑到光子学(Photonics)在超大规模云服务商(Hyperscaler)专用集成电路(ASIC)部署中日益增长的作用。
英文原文
Great thematic choice. Just from first glance, Cantos seems especially ahead of VCs investment curve. Don't know the details of the startup but there's likely another premium multiple applied in the past few months alone. Especially given photonics growing role in Hyperscaler ASIC deployments.
-
解析AI供应链各环节标的风险收益特征，推荐TSM稳健、AXTI高弹性、NBIS高成长。
感谢关注！我来简单拆解一下：我看好金融科技(Fintech)和人工智能(AI)，这是我的两个主要成长型投资板块。像 $CRCL 这样的稳定币(Stablecoins)属于金融科技领域。至于 AI 领域，我专注于供应链投资： 新云厂商(Neoclouds) -> AI 顶层消费者/超大规模云提供商。$NBIS 是我的最爱。 顺着供应链往下，谁帮助制造 $NBIS 和 $IREN 购买的 GPU？ -> 来自 $NVDA 的 GPU，部分 $AMD，以及未来可能来自 $GOOGL、$MSFT、$AMZN 的 ASIC。 谁帮助生产这些 GPU？ -> $TSM 谁帮助在数据中心扩展这些 GPU 集群？ -> $LITE、$COHR、$AAOI 用于生产扩展 GPU 集群所需互连件的原材料是什么？ -> $AXTI 它们都主题性地搭乘 AI 资本支出(CapEx)浪潮，但敞口各不相同。 我通常不会说“1”，因为它们的风险特征不同。 如果必须为复利增长选择最安全的一个：$TSM 如果必须选择 6 个月内潜在上行空间最大的一个？$AXTI 可能因材料短缺而受益，但如果中国封锁出口也可能归零。 如果必须选择 2 年内最有可能实现 5-10 倍增长并成为超大规模云厂商的一个？$NBIS。但存在稀释风险，自动取款机(ATM)发行可能导致股价停滞数月，但从 70-90 亿 ARR 和 20% EBIT 的未来前景来看，长期非对称上行空间最大。 当然，如果 AI 资本支出浪潮放缓，它们都会下跌，因此分散投资于金融科技等其他板块是明智的。
英文原文
Thanks for following along! I'll just give you a breakdown: I like Fintech and AI, those are basically my two main growth investment baskets. Stablecoins like $CRCL are in the fintech domain. As for AI domain I am just supply chain investing: Neoclouds -> Top level consumer/hyperscaler cloud providers for AI. $NBIS is my favorite. Go down the supply chain, who helps manufacture the GPUs used that $NBIS and $IREN buy? -> GPUs from $NVDA, some $AMD, and maybe ASICs from $GOOGL, $MSFT, $AMZN down the road. Who helps create those GPUs -> $TSM Who helps scale up those GPU clusters in data centers? -> $LITE, $COHR, $AAOI What are those raw materials used to produce the interconnects to scale up those GPU clusters? -> $AXTI They're all ride the AI CapEx wave thematically, but the exposure is all different. I never typically never say "1" because they all have different risk-profiles. if I had to choose the safest one for just compounding: $TSM If I had to choose the one with the highest possible upside over 6 months? $AXTI from material shock, but it could also go to 0 if China blocks exports. If I had to choose the best chance to 5-10x and become a hyperscaler over 2 years? $NBIS. But there's risks from dilution, ATMs stalling out the stock for a few months, but highest asymmetrical upside over periods of time from 7-9B arr 20% ebit in the future But of course, if AI CapEx wave slows, then they all go down, which it's good to diversify in other segments like fintech.
-
对比住友与AXTI，指出前者营收占比低但无出口管制风险。
并非如此。我个人认为这是一家很棒的公司，但其中有些细微差别。 衬底(Substrate)生产仅占住友电工(Sumitomo)总收入的很小一部分，因为它已经是一家庞大的公司，所以你无法获得相同的敞口(Exposure)。 此外，在上游方面，他们采购磷化铟(InP)材料，而 $AXTI 部分拥有10多家原材料公司，并自行进行精炼和加工。 话虽如此，与 AXT 相比，住友电工的风险更低，因为没有出口管制核弹(Export Control Nuke)的风险，而且可能有更多的西方资金流入供应链多元化。
英文原文
Not quite. I personally think it's a great company but there's nuances. Substrate prod is only a small fraction of Sumitomo's total revenue since it's already a massive company so you're not getting the same exposure. Also upstream there's InP materials that they buy while $AXTI partially owns 10+ raw material companies and does the refining and processing themselves. That being said, Sumitomo is de-risked compared to AXT since theres no export control nuke and there's probably more western money being funneled into supply chain diversification.
-
COHR受上游InP原料瓶颈制约，超大规模数据中心或转向直接采购上游资源。
$COHR 表现惊人，凭借其新建的 6 英寸晶圆厂，可能是磷化铟(InP)衬底生产领域的美国领先企业。 两个细微差别： 1. 他们的产量远未达到超大规模数据中心(Hyperscalers)所需的规模，目前 $AXTI 和住友(Sumitomo)的产出量更大。 2. 他们受限于上游生长晶体所需的 InP 原料（由 Vital、$DOWA 和 $AXTI（自产）控制）。 CEO 在投资者关系(ER)会议中表示，其产能已满载，且受限于 InP 激光器（以及生产它们所需的上游一级 InP）。 我的思考过程是，超大规模数据中心可能不会继续向 $COHR 增加订单积压，而是向上游延伸。他们会直接从 Vital、$DOWA、$AXTI 等控制 InP 的企业采购 6n-7n 纯度原料，交给住友/$COHR 等晶圆厂，或者再高一级，直接收购 InP 衬底库存 -> 交给 $LITE 作为保险。 话虽如此，大量资金可能会注入西方供应链，但由于上游瓶颈，其产能仍远远不够。
英文原文
$COHR is amazing and probably the leading US company for InP substrate prod with their new 6in wafer fab. Two nuances: 1. Their production is nowhere close to the scale needed for hyperscalers that $AXTI and Sumitomo currently outputs 2. They are bottlenecked by upstream InP feedstock needed to grow crystals (owned by Vital, $DOWA, and $AXTI (captive)). CEO said in their ER that their capacity was maxed out and limited by InP lasers (and by 1 hop InP needed to produce them) My thought process is that instead of placing more order backlogs with $COHR, hyperscalers would go upstream. They would directly source 6n-7n from Vital, $DOWA, $AXTI and others that control InP, and give them to fabs like Sumitomo / $COHR or one level higher, buy out stock of InP substrates -> hand to $LITE as insurance. That being said a lot of money will likely be poured into Western supply chains but their capacity is nowhere enough since they're still bottlenecked upstream.
-
质疑SMM图表代表性，指出海外采购成本远高于国内现货价。
嗯，那张SMM图表不只显示现货国内价格吗？考虑到许可证、关税和配额溢价等因素，我相当确定超大规模数据中心（hyperscalers）无法以那些价格采购。我很好奇你是否有关于西方公司支付价格与国内价格对比的更多信息，以证明这种价格分化（bifurcation）不会发生。
英文原文
Hmm, isn’t that SMM chart just showing spot domestic price though with the 10% price increase? I’m pretty sure hyperscalers can’t buy at those prices accounting for stuff like permits, tariffs, and allocation premium. I’m curious if you have more info about what Western companies pay compared to domestic rates to show that bifurcation wouldn’t occur
-
澄清建模意图，探讨中美双层体系下国内产能对西方基板及原材料供需的影响。
噢，抱歉如果听起来我像是在按30倍建模，那只是一个理论示例，说明如果大幅增加成本，对物料清单(BOM)集群的影响只是轻微的。从西方角度来看，在供应冲击时期，超大规模云服务商可能会愿意支付当前价格数倍的保费作为保险。我真正想理解的是，既然看起来西方和中国之间形成了双层体系，如果国内产能提升，对于基板(substrate)级别和原材料(feedstock)的价格/供应，西方市场+超大规模云服务商会出现哪些细微差别？
英文原文
Oh, apologies if it sounded like I was modeling for 30x, it was just a theoretical example that if you increased costs by large amounts it would just be a dent in BOM clusters. In times of supply shock from a Western angle, hyperscalers would probably be willing to pay a few times current price for insurance. What I’m really trying to understand is that since it looks like a two tier system between the West and China, if domestic production ramps up, what nuances would come into play for prices/supply on Western markets + hyperscalers for substrate level + feedstock.
-
分析InP原料在中美市场的供需差异及西方供应瓶颈。
你认为短缺问题是否因双层体系而更复杂？看中国国内，磷化铟(InP)原料价格正常，可能因供应充足。而在西方市场，据我所知，企业已为6N、7N纯度支付300-500%溢价。除非中国想倾销廉价材料（不太可能），否则任何有出口业务的企业都将受益于西方供应短缺带来的瓶颈。因此，若看中国国内市场，瓶颈可能没那么严重；但从西方超大规模数据中心(Hyperscaler)视角看，情况可能更严峻？
英文原文
Do you think the shortage is a bit more nuanced since it’s a two tier system? If you look at Chinese domestic InP feedstock is priced normally, maybe bc of supply availability. In western markets from the top of my head, companies are already paying a 300-500% premium for 6n, 7n. And anyone with exports would benefit from the bottleneck from a dearth in western supply unless China wants to dump cheap materials on the market (probably unlikely). So if you’re looking at bottleneck China domestic markets it might not be as dire but from a Western hyperscaler perspective it might be more serious?
-
铟价持续上涨，预计2026年中出现瓶颈，$AXTI 受益最大。
更正一下上面关于铟(In)价格的内容。 6N级铟（2024年1月、2025年1月、2025年12月）~$680/公斤 -> ~$840/公斤 -> ~$1,110–$1,350/公斤 7N级铟 ~$1,150/公斤 -> ~$1,800/公斤 -> $3,450+/公斤 1.6T工艺需要7N级铟。 因此价格已经开始上涨，但我预计巨大的瓶颈(chokepoint)将出现在2026年年中。 不确定价格最终去向，但我预计会出现为确保材料供应的竞价战，且$AXTI 相比住友(Sumitomo)等其他公司，可能对此有最直接的敞口(exposure)。
英文原文
Just to make a correction to the Indium price above 6N grade Indium (Jan 2024, Jan 2025, Dec 2025) ~$680 / kg -> ~$840 / kg -> ~$1,110 – $1,350 / kg 7N grade Indium ~$1,150 / kg -> ~$1,800 / kg -> $3,450+ / kg 1.6T requires 7N. So prices have already started to ramp up, but I expect the huge chokepoint to come mid 2026. Not sure where the price goes, I just expect a bidding war to secure materials and $AXTI probably had the most direct exposure to this over Sumitomo and others.
-
分析InP材料供应瓶颈，通过历史案例说明关键材料博弈论，预判超大规模云商将抢购InP衬底产能。
当我发WSB帖子时，我半开玩笑地提到$AXTI。但对于极端瓶颈和InP供应冲击，我可不是开玩笑的。这是一个关于关键材料博弈论的呼唤。 如果我们回顾历史上许多瓶颈案例： 1. 镝(Dysprosium)：约2300%涨幅（约$100/kg -> $2400/kg，2010-11年） 2. 氖气(Neon Gas)：约1000%-2000%涨幅（2022年） 3. 铑(Rhodium)：约$3,000/盎司 -> 约$29,000/盎司（约860%） 当前铟(In)的价格涨幅与其他材料相比如何？ 89%：（约$440/kg -> 约$832/kg）。 用于半导体的关键材料（如用于光刻激光器的氖气），因为芯片制造商如英特尔和三星在乌克兰冲突期间进行了储备，价格暴涨超过2000%。他们吸收了成本，因为氖气只占芯片价值的一小部分。 与氖气一样，磷化铟(InP)是一种关键的"赋能"材料。晶圆成本相对于AI集群成本（$100的晶圆对比$30,000的GPU）来说很小。超大规模云商宁愿吸收巨大的价格上涨，也不愿延迟部署。InP目前还没有出现这种情况，但我们很可能会看到。 传统分析师和许多AI模型不知道如何建模关键瓶颈，因为历史上很少有这样的案例：一种极其廉价的商品（如InP，TAM只有几亿美元）突然成为价值数万亿美元的AI建设中最关键的材料。 目前，超大规模云商可能已经向下游供应商Coherent/$LITE下了订单，但可能没有意识到激光器生产和产能受到上游材料短缺的制约而排满。 InP衬底（$AXTI + Sumitomo + JX生产）和磷化铟原料（$AXTI、Vital、$DOWA）都存在供应储备，需求超过供应数倍。 $NVDA最初通过锁定EML大量产能加剧了供应短缺。但短缺早在超大规模云商扩产之前就已经存在，大部分需求将在2026年底至2027年Trainium或Maia出现时涌入。 但如何绕过原始瓶颈并确保AI部署？ 直接从（$AXTI、Sumitomo）购买衬底产能，或者更进一层，购买InP原料并将其交给衬底生产商，以确保谷歌的TPU项目不会停滞。 与其他关键材料（如氖气用于光刻激光器）类似，磷化铟和InP衬底对整个AI建设更为关键： 1. Nvidia：InfiniBand / NVLink（光学 - 800G/1.6T EML光模块）- 对InP高度依赖 2. Google：Jupiter / Apollo（光学 - OCS），高度依赖 3. Microsoft：Azure Maia：高（800G DR4/FR4） 4. Meta：F16 / Artemis（800G）高 这还不包括其他Mag7公司如$AMZN或使用光子技术扩展ASIC的其他公司。 目前，分析师基于TAM对InP公司进行建模。例如，2024年InP晶圆市场总规模为$183M-$205M（Straits Research, Mordor Intelligence，2024年） - 6英寸（150mm）细分市场TAM：$65M - 预计2032年InP总TAM：$580M-$700M 现在看激光级磷化铟原料的TAM： - 多晶InP原料TAM：约$3亿 - 高纯度铟（6N/7N）仅：约$4亿 衬底数据：Mordor Intelligence（InP晶圆市场2025-2030）和Straits Research。 原料数据：Market Report Analytics（磷化铟多晶行业分析2025）。 这是一种极其便宜的电信产品，现在成为整个供应链中最珍贵的材料之一。 价格上涨3000%只会占$GOOGL TPU部署或$MSFT MAIA部署BOM值的低个位数。而仅微软2027年的扩产就会占用全部InP产能的两位数百分比。 就像2022年的英特尔或半导体行业一样，超大规模云商可能会像当时囤积氖气一样，直接购买InP衬底+InP作为"保险"。 如果"TAM"实际上只有几亿美元，但每个超大规模云商如果不确保这种材料就会面临部署延迟，那么这就变成了一场博弈论竞标战。 在一个因$100衬底而停滞的价值$200亿的TPU部署中，超大规模云商会毫不犹豫地支付$10,000（比当前价格高出100倍）。 这就是为什么当你看到InP衬底公司有$5000万订单积压（来自2024年），同样的$5000万产能对未来来说在极端情况下可能价值$5亿、$50亿甚至更多。 这是一个"尾部风险"投资，如果超大规模云商被这种材料卡住，回报是极高的，但并非没有重大的地缘政治风险和稀释。 我不知道会发生什么，但我预计2026年InP衬底和InP原料都会出现巨大瓶颈。 那么，如果像$AXTI这样的公司拥有整个InP供应链40%的份额（来源：AXT CEO），而估值只有$7亿，你会给它什么估值？
英文原文
When I made my WSB post, I'm half joking about $AXTI. But not about the extreme bottleneck and InP supply shock. This is a call on critical materials game theory. If we see many historically at bottlenecks: 1. Dysprosium: ~2,300% increase (~$100/kg -> $2400/kg 2010-11) 2. Neon Gas: ~1000%-2,000% increase (2022) 3. Rhodium: ~$3,000/oz -> ~$29,000/oz (~860%) The current Indium spike compared to others? 89% : (~$440 / kg -> ~$832 / kg). Critical materials used for semiconductors like Neon Gas (semi-grade lithography lasers), spiked over 2000% because chipmakers like Intel and Samsung had stockpiled during Ukraine conflicts. They absorbed the cost because neon is a small fraction of the chip's value. Like Neon, InP is a critical "enabler" material. The cost of the wafer is small relative to the cost of the AI cluster ($100 wafer vs $30,000 GPU). Hyperscalers would happily absorb huge price increases rather than delay deployment. We haven't seen that yet with InP, but we likely will. Traditional analysts and many AI Models do not understand how to model critical bottlenecks because there's not many examples in history where an extremely cheap commodity like InP with few hundred million TAM suddenly became the most critical material for the multi-trillion+ AI buildout. Currently, hyperscalers likely placed backlogs on downstream providers Coherent / $LITE, but likely didn't realize that laser production and capacity are backlogged because of materials shortage upstream. There is a supply storage both for InP substrates ( $AXTI + Sumitomo + JX production) as well as the Indium Phosphide feedstock ( $AXTI, Vital, $DOWA) where demand exceeds supply by multiple factors. $NVDA originally contributed to the supply shortage by locking in a large percent of capacity of EML. But shortage is already pre-hyperscaler ramp, where majority of the demand will come in late 2026 into 2027 when Trainium or Maia show up. But how do you get around original chokepoint and securing your AI deployment? Buying out substrate capacity directly from ( $AXTI, Sumitomo) or going one step lower and buying InP feedstock and passing them to substrate producers so Google's TPU program doesn't stall. And similar to how other critical materials were needed for lithography lasers. Indium Phopshide and InP substates are even more critical to the entire AI buildout: 1. Nvidia: InfiniBand / NVLink (optical - 800G/1.6T EML Transceivers) - Extreme dependency on InP 2. Google: Jupiter / Apollo (optical - OCS), Extreme dependency 3. Microsoft: Azure Maia: High (800G DR4/FR4) 4. Meta: F16 / Artemis (800G) High and this is not considering other Mag7 companies like $AMZN or other companies scaling out ASICs with photonics. Currently, analysts are modeling InP companies based on TAM. For example, the total InP Wafer Market from 2024 is $183M – $205M Straits Research, Mordor Intelligence (2024) - 6-inch (150mm) Segment TAM: $65m - Projected Total InP TAM (2032): $580M – $700M Now looking at TAM for Laser-Grade Indium Phosphide Feedstock: - Polycrystalline InP Feedstock TAM: ~$300 Million - High-Purity Indium (6N/7N) Only: ~$400 Million Substrate Data: Mordor Intelligence (InP Wafer Market 2025-2030) and Straits Research. Feedstock Data: Market Report Analytics (Indium Phosphide Polycrystalline Industry Analysis 2025). This was an extremely cheap telecom product and now it's one of the most precious materials in the entire supply chain. Increasing prices 3000% would only be low single digits of BOM value to $GOOGL TPU deployments or $MSFT MAIA deployments. And the project Microsoft ramp in 2027 alone would take up double digits of all InP capacity. Like Intel or semis back in 2022, hyperscalers will likely buy insurance directly with InP substrates + InP, similar to how semis stockpiled Neon Gas. If the "TAM" really was a few hundred million, but every single hyperscaler would face deployment delays if they don't secure this material. Then this becomes a game theory bidding war. In a $20B TPU deployment is stalled because of a $100 substrate, a hyperscaler would easily pay $10,000 (100 times current prices) for it without hesitation. That's why when you look at InP substrate companies and see $50m backlog (from 2024), that same $50m worth of capacity for future ones could be valued at $500m, $5B or more in an extreme scenario. This is a "tail-risk" investment, extremely reward if hyperscalers get bottlenecked by this material, but not without significant geopolitical risk and dilution. I don't know what's going to happen, but I do expect there to be an immense bottleneck in 2026 both in InP substrates and InP feedstock. So, if companies like $AXTI owns 40% of the entire InP supply chain (source: AXT CEO) and is valued at $700m, what value would you place on it?
-
探讨InP原料与衬底瓶颈差异，指出原料冲击利好AXT而非Sumitomo。
@zephyr_z9 我很好奇。鉴于磷化铟(InP)衬底控制和磷化铟(InP)材料是两个不同的瓶颈，你会更看重哪一个？ 如果上游的磷化铟(InP)原料供应出现冲击，主要的受益者将是像AXT/DOWA这类公司，而不是无法捕获上行收益的住友电工(Sumitomo)。
英文原文
@zephyr_z9 I’m curious. Would you place more value on InP substrate control or indium phosphide since they’re two different bottlenecks? If InP feedstock has a supply shock upstream, the main beneficiaries would be your AXT/DOWA type companies instead of Sumitomo that don’t capture upside
-
InP涨价对集群成本影响微，长期担忧超大规模云服务商转向定制ASIC。
好问题！关于 $CRWV、$NBIS 和 $IREN，我尚未对 B300 集群进行物料清单(BOM)分析，但光网络(Optical Networking)资本支出占比可能在 10-20% 左右。 我记得做过测算，即使磷化铟(InP)价格上涨 30 倍，集群价格仅边际增加 3%，这对 Google 来说微不足道，仅对内部收益率(IRR)造成极微小的影响。 话虽如此，我不确定 InP 成本是仅边际增加，还是因总可寻址市场(TAM)极小且突然用于 AI 导致供应受限而上涨 50 倍。 总之，短中期影响甚微。长期我更担心超大规模云服务商(Hyperscaler)转向定制 ASIC 并将英伟达作为过渡期。 我们需要观察 Neoclouds 是否能在拥有极大杠杆的 2-3 年窗口期内，基本转变为超大规模云服务商。
英文原文
Good question! As for $CRWV, $NBIS, and $IREN, I haven’t done a BOM on B300 clusters, but optical networking is probably around 10-20% of capex. I remember doing calculations and increasing InP prices by 30 times would only create a 3% marginal increase on cluster prices, which is extremely negligible to Google but hurts IRR just a tiny tiny bit. That being said I don’t know if InP costs just marginally increases or it goes up 50 times because it’s incredibly low TAM and suddenly it’s used for AI/supply constrained. So basically little effect short-medium term. Longer term I’d be more worried about hyperscaler shift to custom ASICs and using nvidia as a transition period. We’d need to see if Neoclouds can basically become hyperscalers themselves in the 2-3 year window where they have extreme leverage
-
AXTI掌控磷化铟关键瓶颈，制约AI光互联扩展。
$AXTI CEO：“我们占据了磷化铟(Indium Phosphide)供应链的40%份额。” 是否曾有一家如此小规模（市值7亿美元）的公司，掌控着一个数万亿美元产业的关键瓶颈？
英文原文
$AXTI CEO - “We are 40% of the Indium Phosphide supply chain” Has there ever been a company this small ($700M) that controlled a critical bottleneck for a multi-trillion dollar industry? https://t.co/tn7EOMPgas
-
磷化铟供应链集中度高，是光子器件未来关键瓶颈。
@brandy117 声称控制了全球 40% 的磷化铟(Indium Phosphide)供应链，这令人震惊。特别是当光子器件依赖该材料时，它将成为未来巨大的瓶颈。 https://t.co/3u29jRnTT8
英文原文
@brandy117 The claims were mindblowing controlling 40% of the world's Indium Phosphide supply chain. Especially when photonic components depend on the material, it's a huge future chokepoint. https://t.co/3u29jRnTT8
-
分析InP供应链瓶颈，指出AXT受益于垂直整合，需求总体仍超供应。
我理解你的观点是基于仅看中国国内市场，而未考虑西方AI的扩张。两点如下： 一是整体需求的极端增长。即使有6英寸磷化铟(InP)晶圆，超大规模云服务商(Hyperscaler)的需求仍将超过供应。因此，无论是否存在竞争性替代品，需求不会消失，但可能会分散到中国国内市场的竞争者中。 二是垂直整合。你提到了西方阵营的$DOWA，特别是磷化铟(InP)领域。$AXTI控制了全球绝大部分的InP供应。如果超大规模云服务商为保险起见直接高价采购，再转交给$COHR，那么这对住友电工(Sumitomo)来说问题更大。而AXT则从这两层瓶颈中受益。新耀/JFS进入供应链只是增加了参与者，并非末日景象。例如存储领域（SK海力士；MU，三星）。 当然，如果中国实施出口管制，我们仅看中国国内市场，上述观点可能失效。但总体而言，即使增加更多参与者，需求仍超过供应（例如COHR已满产）。
英文原文
I can see that if you look at China domestic markets only but not western AI ramp. So two things: One is extreme ramp in general. Hyperscaler demand will always outstrip supply even with 6-inch InP wafers. So competitive alternative or not, demand doesn’t disappear but might get spread out through competitive alternatives in China’s domestic market. The other is vertical integration. You mentioned $DOWA on the western side and especially InP. $AXTI controls a huge percent of the world’s InP supply and if those prices gets jacked up from hyperscaler buying direct for insurance-> handing them to $COHR, then it’s more of an issue for Sumitomo. And AXT benefits from both layers of the bottlenecks. Xinyao/JFS entering the supply chain just makes adds one more player to the equation so it’s not doom and gloom. Eg memory (sk Hynix; MU, Samsung) Of course these points might get thrown out the drain if China export controls and we’re looking at China domestic market only. But generally demand outstrips supply (eg. COHR is at max) even if we add in more players.
-
梳理AI光子学及InP衬底供应链核心标的，涵盖从底层材料到芯片的全链条。
呃，根据我个人的研究，为了全面覆盖我看好的一切： 1. $AXTI 位于整个供应链的最底端，占据 1/3 的磷化铟(InP)衬底和 1/4 的磷化铟(InP)份额。AI 供应链竟然与这只市值 7 亿美元的股票紧密相连，这完全出乎意料。 2. $DOWA + 住友电气，用于在衬底和材料方面进行西方对冲。 3. $AAOI, $LITE 针对超大规模云服务商的芯片，一家服务于 $AMZN/$MSFT，另一家服务于其他客户但更侧重于 $TPU。 4. $MRVL + $AVGO 我认为这就足以捕捉光子学物料清单(BOM)的价值以及磷化铟(InP)供应冲击。 还有一些其他有趣的标的，比如 Landmark Opto，我正在关注。 $POET 没进这个名单，反正你可以通过 $MRVL 获得敞口，除非我漏掉了什么。
英文原文
Uhh, from my own personal research so far, to cover all grounds I’ve liked: 1. $AXTI at the very bottom of the entire supply chain + 1/3rd of InP substrates 1/4th of InP. Completely unheard of that AI supply chain is tethered to this $700m stock. 2. $DOWA + Sumitomo Electric for Western Hedge on substrate + materials. 3. $AAOI, $LITE for hyperscaler chips, one for $AMZN/ $MSFT, the other for everything but more levered to $TPU 4. $MRVL + $AVGO I think that’s kind of all you need to capture photonics BOM value + InP supply shock. There’s some other interesting ones like Landmark Opto that I’m looking at now. $POET didn’t quite make that list and you’d be getting exposure through $MRVL anyway, unless there’s something I missed
-
对比住友与AXTI在磷化铟供应链中的地位及风险。
住友电气(Sumitomo)受磷化铟(InP)价格上涨影响，但如果$GOOGL直接向$DOWA或其他供应商采购原料，其影响将减弱。 $AXTI拥有完整的供应链，并占据激光级磷化铟的巨大份额。但明显的下行风险是中国出口管制。 不过住友电气体量巨大，因此你获得的并非最直接的敞口。
英文原文
Sumitomo is subject to hikes in InP prices though if $GOOGL goes direct to $DOWA or others for feedstock. $AXTI owns the whole supply chain, and a huge portion of laser grade inp. But downside is obviously anuke from China exports. Sumitomo Electric is huge though so you’re not exactly getting the most direct exposure.
-
分析AXTI/中国磷化铟瓶颈对AI供应链的致命影响及关键材料估值逻辑。
这是一份关于 $AXTI/中国 在衬底生产和激光级磷化铟方面导致整个 AI 基础设施建设出现单点故障的研究笔记。我并未对个别公司进行估值分析。但如果这一流程中断，$LITE、$COHR、$AVGO 乃至 $GOOGL、$MSFT 都将面临极端的延迟。如果你们感兴趣，我会另写一篇文章关于如何对关键材料瓶颈进行估值。基本上，如果 2024 年的积压订单为 5000 万美元，而总可寻址市场 (TAM) 为 1.5 亿美元。而现在，由于它是关键材料，如果谷歌愿意为此支付 50 倍的溢价（仅为保持生存而增加 BOM 成本的百分之几），那么原本 5000 万美元积压订单的估值逻辑就会转变为未来产能的 25 亿美元。
英文原文
This is a research note on point failure of the entire AI buildout in $AXTI/China from both substrate production and laser-grade Indium Phosphide. I’m not doing valuation analysis on individual companies But if this flow stops, $LITE, $COHR, $AVGO to $GOOGL, $MSFT faces extreme delays. I’ll write something else about valuing critical material bottlenecks if you want. But basically if a backlog was $50m from 2024 and TAM was $150m. And now because it’s a critical material, if Google would be willing to pay 50 times that (only a few percent increase on BOM cost to stay alive), then the valuation math of that original $50m backlog -> $2.5B for future capacity changes.
-
Coherent虽为西方希望，但受制于磷化铟原料垄断，面临供应链瓶颈风险。
我认为超大规模云计算厂商尚未意识到其中的危险。$COHR 是领先的“近乎垂直整合”的美国公司。但风险在于，它仍受制于磷化铟(InP)原料的瓶颈，并从现货市场购买6N级磷化铟多晶（原料）。 如果上述公司停止销售6N多晶，Coherent 的激光和收发器业务将随着超大规模云计算厂商的建设而有效受阻。 Coherent 并不拥有像 Vital Materials 或 $AXTI 那样规模的主要铟矿或大规模6N合成工厂。 简单来说：Coherent 是打破住友与 $AXTI 之间衬底双寡头垄断的“西方希望”，但他们目前却是磷化铟原料垄断的人质。
英文原文
Don’t think hyperscalers understand the danger yet. $COHR is the leading “virtually integrated” American company. But the risk is it’s still stuck in the bottleneck for InP feedstock and buys 6N InP polycrystal (feedstock) from the merchant market. If the companies above stop selling 6N polycrystal, Coherent’s laser and transceiver business effectively hits a wall alongside the hyperscaler buildout. Coherent does not own major Indium mines or high-volume 6N synthesis plants on the scale of Vital Materials or $AXTI. To put it more simply: Coherent is the "Western Hope" for breaking the substrate duopoly between Sumitomo and $AXTI, but they are currently a hostage to the InP feedstok monopoly.
-
AI光互连依赖磷化铟，其原料78%在中国，AXTI成关键瓶颈。
“瓶颈中的瓶颈”：磷化铟(InP) $AXTI 和 $SMTOY 构成双寡头。它们控制了全球 60% 以上的磷化铟(InP)衬底，是 AI 建设的关键瓶颈。 更令人警觉的是：这依赖于磷化铟(InP)。 磷化铟(InP)是垄断市场。78% 由 $AXTI 和中国控制。 这就是“瓶颈中的瓶颈”： 磷化铟(InP)衬底市场已是日本和中国实体之间的双寡头。然而，上游激光级原料（如住友所需）的生产主要由 $AXTI（再次出现）和中国公司主导。 主要问题在于： AI 行业极度依赖光子学来实现规模化。无论收发器是“基于磷化铟”还是“基于硅”，光源始终是：磷化铟(InP)。 随着 AI 集群从 1 万颗 GPU 扩展到 10 万颗以上，“铜缆瓶颈”迫使全面转向光互连。这使得磷化铟(InP)对 AI 规模化的重要性超过几乎任何其他单一材料。 此外，激光级原料纯度必须至少达到 6N（99.9999%）。即使是微量杂质（硅、硫、锌）也不像 LED 等其他领域那样可接受。大多数研究可能混淆了激光级原料与普通磷化铟产出。 住友是磷化铟衬底市场最著名的名字。他们在高端激光衬底领域占据巨大市场份额。但贸易数据显示，他们越来越依赖来自 $DOWA 和中国公司的外部磷化铟原料。（由于自有回收和储备，他们可能仍被计入原料所有权列表中） 所以，如果认为整个西方 AI 建设因有一家日本公司作为双寡头之一而感到安全，请三思。 瓶颈的源头被 $AXTI 和中国供应商所瓶颈。 如果我们看生产，以下是深度研究估算的控制权分布： Vital Materials - 35%（现货市场） - Vital 在 2020 年收购凡亚金属交易所库存（3600 吨铟），使其掌控全球铟市场。他们利用这一优势成为最大的磷化铟多晶生产商。主导成本领先者；控制原生铟供应（凡亚库存）。 $AXTI - 25%（北京同美晶体技术） - 主要自给自足。其大部分多晶生产被自家晶体生长炉消耗以制造晶圆。但他们也向现货市场销售原材料和化合物，是混合型玩家。 他们是全球唯一从矿山、精炼厂到磷化铟衬底产出的公司，鉴于他们同时拥有原材料和磷化铟衬底控制权，这是一个绝对疯狂的瓶颈。供应自家晶圆生产；关键的非日本来源。 株洲科能新材料 - 18% - 科能实现了“磷化铟多晶量产”并拥有“高度集成的自动多晶生产能力”。他们是中国铟出口配额的关键接收方。纯合成专家；主要出口商。 然后外部： $DOWA - 12% - 西方公司中最大的供应商（是的，日本）。在采矿（秋田锌）方面根基深厚，因此是高纯度金属有机化合物和化合物的领先供应商。 “日本企业”（住友、JX）的“来源” Wafer Technology Ltd - 5% - IQE plc 的子公司。对于需要非中国原料的北约联盟供应链来说，这是关键战略资产，尽管他们目前正在进行资产清算。 InPACT（法国）- 3% - InPACT 称他们通过高压合成的水平梯度凝固(HGF)生产多晶锭，以满足（诺基亚、爱立信）和一些欧洲国防部门的要求。 世界其他地区 - 2% 这包括主要衬底制造商（Coherent、住友）、小规模研究生产以及试图打破垄断的韩国或台湾新兴玩家。 现实是，全球 78% 的激光级磷化铟原料产能位于中国。 光组件的需求已经以数倍超过供应/生产。这还是在上量之前。整个 AI 建设被两家公司瓶颈。 但 1/4 的材料来自：$AXTI。而约 1/3 的磷化铟衬底产出再次来自激光级原料。 整个 AI 建设依赖于一家随机的 7 亿美元公司，它既是瓶颈中的瓶颈，也是瓶颈本身。 扩大磷化铟衬底和激光级原料的生产既是国家安全问题。
英文原文
"Bottleneck within a Bottleneck": Indium Phosphide $AXTI | $SMTOY is a duopoly. They control 60%+ of the world's InP substrates as the bottleneck of the AI buildout. What's even more alarming? This is dependent on: Indium Phosphide. InP is a Monopoly. 78% control by $AXTI / China. Here's the bottleneck within a bottleneck: The InP substrates is already a duopoly between Japanese and Chinese entities. However, upstream production of laser-grade feedstock required by companies like Sumitomo is dominated by $AXTI (again) and Chinese companies. Here's the major issue: The AI industry is extremely reliant on photonics to scale. Whether, transceivers are "InP-based" or "Silicon-based" the photon source remains: Indium Phosphide. As AI clusters move from 10,000 to 100,000+ GPUs, the "copper bottleneck" is forcing a total shift to optical interconnects. This makes InP more critical to AI scaling than almost any other single material. On top of that, laser-grade feedstock must be at least 6N (99.9999%) pure. Even trace amounts of impurities (Silicon, Sulfur, Zinc), cannot be used like they do in other segments like LED. And most research probably confuses laser-grade feedstock with regular InP outputs. Sumitomo is arguably the most famous name in the InP substrate market. They hold a massive market share in the high-end laser substrate segment. Yet trade data suggest they are increasingly reliant on external feedstock for Indium Phosphide from companies like $DOWA and from China. (They're likely still included in feedstock ownership lists bc of captive recycling and reserves) So if the entire Western AI buildout feels safe regarding a Japanese company as a duopoly, think again. The source of the bottleneck is bottlenecked by Chinese suppliers like $AXTI and Chinese companies. If we look at production there's are estimates from deep research breakdown of estimate controls: Vital Materials - 35% (Merchant) - Vital acquired Fanya Metal Exchange inventory (3,600 tonnes of Indium) in 2020 gave them large control over the global Indium market. They have leveraged this to become the largest producer of InP polycrystal. Dominant cost leader; controls raw Indium supply (Fanya stocks). $AXTI - 25% (Beijing Tongmei Xtal Technology) - Primarily Captive. Most of their polycrystal production is consumed by their own crystal growth furnaces to make wafers. However, they also sell raw materials and compounds to the merchant market, making them a hybrid player. They're the only company in the world that goes from mines, refineries -> InP substrate output, which is an absolutely crazy bottleneck given they own both the raw materials into InP substrate controls. Feeds own wafer production; key non-Japan source. Zhuzhou Keneng New Material - 18% - Keneng has achieved "mass production of polycrystalline indium phosphide" and possesses "highly integrated automatic polycrystalline production capacity". They are a key recipient of China's export quotas for Indium. Pure-play synthesis specialist; major exporter. Then Outside: $DOWA - 12% - The biggest supplier out of any Western company (yeah Japanese). Deep roots in mining (Akita Zinc), so they're the leading supplier of high-purity metal organics and compounds. The "source" for Japan Inc. (Sumitomo, JX) Wafer Technology Ltd - 5% - A subsidiary of IQE plc. Critical strategic asset for NATO-aligned supply chains that require non-Chinese feedstock, even though they're going through asset liquidations right now. InPACT (France) - 3% - InPACT states they produce polycrystalline ingots via Horizontal Gradient Freeze (HGF) using high-pressure synthesis so they meet requirements from (Nokia, Ericsson) and some European defense sectors. Rest of World - 2% This includes major substrate makers (Coherent, Sumitomo), small-scale research production, and emerging players in Korea or Taiwan who are attempting to break the monopoly. The reality is that 78% of the global capacity for laser-grade InP feedstock is located in China. The demand for optical components already exceeds supply/production by multiple factors. This is pre-ramp too. The entire AI buildout is bottlenecked by two companies. But 1/4th of the materials comes from: $AXTI. And ~1/3rd of InP Substrate outputs again comes from laser-grade feedstock. The entire AI buildout is dependent on some random $700m company that serves as both the bottleneck of a bottleneck and the bottleneck. It's both national security issue to scale up InP production both for substrate and laser-grade feedstock.
-
三星遭中方窃取存储芯片机密，恐被逐出中国市场，处罚力度引争议。
哇，中国知识产权间谍活动并非新鲜事，但这一事件对三星以及美国/韩国而言简直是灾难性的。窃取全球最先进存储芯片的秘密，然后交给中国国有背景的公司……随后很可能将三星/美光切断出中国国内市场。对于这10名被起诉者，即使只有丝毫牵连，终身监禁作为威慑都不足以构成惩罚。如果他们10年后出狱还能拿到5000万美元的薪水，那这就足以激励这些人继续盗窃。
英文原文
Wow, Chinese IP espionage is not new but this one just catastrophic for Samsung and US/Korea. Stealing secrets of the world's most advanced memory chips then handing them over China’s state backed company…. Then likely cut off Samsung/Micron from China’s domestic market. Maximum life in prison for deterrence isn’t even enough of a punishment for this, from the 10 indicted, even with an iota of involvement. If they get out of jail 10 years later with a $50M paycheck, then that’s incentive enough for these people to continue theft
-
构建模型对比磷化铟激光级原料的垂直整合与公开市场供应。
@dr_lmfao_ 是的，绝对如此。我正在构建一个模型，以标准化磷化铟(Indium Phosphide)激光级原料的拥有权，对比像 $AXTI 这样的私有垂直整合玩家，与像 $DOWA 或 Vital Materials 这样在公开市场销售的其他公司。 这极其复杂 lol
英文原文
@dr_lmfao_ Yeah absolutely, I’m working on a model to normalize ownership of indium phosphide for just laser grade feedstock from private vertically integrated players like $AXTI vs others like $DOWA or Vital Materials that sell on open markets. It’s extremely complex lol
-
AI供应链瓶颈类似当年锂和HBM，衬底及关键组件面临短缺与估值难题。
同意，对于交易者而言，这应该是一个有趣的机会，就像2021年电动车兴起时的锂，或者当前SK海力士和美光的高带宽内存(HBM)，如果人们能早期识别瓶颈的话。 机会众多（$AAOI尽管获得新的超大规模客户订单，年初至今仅上涨2.9%。这与Maia/Tranium等超大规模客户的部署挂钩）。 $MRVL需求极度旺盛，但主要来自Maia在2026年第四季度至2027年的需求，年初至今下跌22%。$POET作为$MRVL扩产的间接受益者。 磷化铟(InP)衬底在整个供应链中普遍短缺。$GOOGL加速TPU部署，利好$LITE，以及$COHR的垂直整合+创纪录积压订单。 像$AXTI这样极其荒谬的玩家，市值仅7亿美元，却几乎是未来AI扩产的单点故障(SPOF)。真不知道该如何给这个估值，哈哈。 $COHR、$LITE和衬底供应商现在正触及产能极限。我们将看到这需要多长时间才能达到像内存那样的极端程度。
英文原文
Agreed, for traders, this should be an interesting opportunity like Lithium in 2021 when EVs came about or HBM currently with Sk Hynix and Micron if people recognized the bottlenecks early. Tons of opportunity ($AAOI only up 2.9% YTD despite new hyperscaler orders. Tied to hyperscaler deployments like Maia/Tranium). $MRVL with extreme demand but mainly in q4 2026, into 2027 from Maia, down 22% YTD. And $POET as an indirect beneficiary of $MRVL ramp too. Shortages throughout the entire supply chain with InP substrates. $GOOGL ramping up TPU which benefits $LITE and vertical integration + record backlog from $COHR Extremely absurd players like $AXTI with a $700m marketcap basically being the single point of failure for the future AI ramp. Not quite sure how you place a value on that lol. $COHR, $LITE and substrate providers are hitting the capacity limits now. We’ll see how much time it takes for his to go to the extremes like memory.
-
AXTI估值低反映其风险，AI供应链对其光子学衬底依赖加剧脆弱性。
$AXTI 确实存在很大风险。这就是为什么它的估值仅为 7 亿美元，像是一只仙股，而不是像 $SMTOY 那样超过 300 亿美元，尽管两者都是垂直整合企业且都是最重要的供应商之一。此外，AI 供应链极度依赖该公司提供的用于光子学的 inp 衬底（inp substrates）原材料以实现产能爬坡，这一事实进一步加剧了本已脆弱的 AI 基础设施建设所面临的风险。
英文原文
There is a lot risk with $AXTI for sure. That’s why it’s valued like a penny stock at $700m and not $30B+ like $SMTOY despite being both vertically integrated and one of the most important suppliers. And when you combine with the fact the AI supply chain is extremely reliant on this company for raw materials -> inp substrates they provide for photonics in order to ramp, that’s even a greater risk to the already fragile AI buildout.
-
澄清Grok误解，指出AVGO依赖AXTI等InP衬底，AXTI具备全产业链优势。
Grok 可能混淆了晶圆制造(wafer fab)与磷化铟(InP)衬底制造。$AVGO 的 Penn 工厂仍需从住友、$AXTI 或 JX 采购 InP 衬底来制造激光器（这就是瓶颈所在），且它们很可能是客户。 而那家市值仅 7 亿美元的化合物半导体公司 $AXTI 可能是其中唯一一家实现端到端控制的企业，从镓、铟等初始原材料精炼开始就拥有上游资源。 因此，“主要不进口 InP 晶圆/衬底”这一说法可能是错误的。
英文原文
Grok is likely conflating wafer fab with InP substrate manufacturing. $AVGO's penn fab still requires InP substrates from Sumitomo, $AXTI, or JX to build the lasers (hence the bottleneck) and they're likely customers. And that small $700M MC company $AXTI is probably the only one of the bunch that does this end-to-end by owning Gallium, Indium, etc. on from the initial raw material refining. So the "does not primarily import InP wafers/substrates” part is probably wrong.
-
光子供应链瓶颈或现类似HBM的投资机会
是的，$LITE 和 $COHR 表现惊人，自拥有强大定价权以来涨幅显著。 若深入挖掘 $AXTI（垂直整合）、住友 -> 同和等更深层环节，集中度风险与瓶颈会发出警示信号。 在 HBM（高带宽内存）领域，当人们发现瓶颈时便出现了投资机会。如今在光子供应链中，随着超大规模云厂商囤积材料并收购产能，很可能存在同样的被忽视的机会。
英文原文
Yep, $LITE and $COHR are amazing and are up quite a bit since they tons pricing power. Go few levels deeper into $AXTI (vertically integrated), Sumitomo -> Dowa, etc. the concentration risk + bottleneck flashes warning signs. With HBM, it was an investment opportunity when people found that bottleneck. There's likely that same overlooked opportunity with photonic supply chains now when hyperscalers hoard materials and acquire capacity.
-
InP衬底成AI光互连瓶颈，$AXTI/$SMTOY掌控命脉，2026年或致供应链危机。
“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 $SMTOY. It seems inevitable that AI will hit the physical ceiling because of InP substrate capacity unless architectures change. In 2024, the bottleneck was GPUs. In 2025, it was HBM. In 2026, the primary constraint will likely be the optical interconnect, and specifically, the InP substrates that power them. This has now become the hidden bottleneck of the AI buildout.
-
InP衬底短缺成AI瓶颈，AXTI与SMTOY具关键战略价值。
没错。这正是我向 @sama、@sundarpichai、@satyanadella 及西方 AI 领袖敲响警钟的原因：磷化铟(InP)衬底严重短缺/瓶颈将限制未来涉及光子学的 TPU/ASIC 部署。基于 InP 的收发器/激光器需求远超供应数倍（来源：LightCounting + 麦肯锡）。我们才刚开始 AI 扩张，TPU 相关的 InP 衬底行业已遇瓶颈 -> 被两家企业扼住咽喉：$AXTI、$SMTOY。若贸易战影响 $AXTI，AI 集群扩展将面临材料真空。这两家公司的重要性无与伦比，暂且不论股价。
英文原文
Exactly. You just out why I'm sounding all the alarm bells to @sama, @sundarpichai, @satyanadella and Western AI leaders about huge InP substrate shortages/bottlenecks limiting future TPU/ASIC deployment involving photonics. Demand FAR exceeds supply for InP-based transceivers/lasers by multiple factors (source: LightCounting + McKinsey). We've only started the AI ramp with the TPU and the InP substrate industry is already bottlenecked -> held at chokepoint by two players: $AXTI, $SMTOY If trade wars affect $AXTI, there's just a complete void of available materials to scale AI clusters. There's just incredible importance on these two companies, stock price aside.
-
AXTI和SMTOY垄断InP衬底，成西方AI光互连供应链致命瓶颈。
这令人极度警惕的是，西方AI基础设施建设可能会被像$AXTI和$SMTOY这样市值仅7亿美元的小公司扼住咽喉。 磷化铟(InP)衬底曾是电信领域的极小众材料，但随着其成为全球激光器和光互连(Photonics Interconnects)的关键材料，这突然变成了国家安全紧急事件。 我不确定顶级供应商是否意识到了这一点，或者至少没有意识到问题的规模（例如，与$LITE签约时并未深入考察Lite的材料供应商），因为供应链正被两家公司严重卡脖子。 以Google TPU v7的量产为例，由于他们的芯片组使用共封装光学(CPO)，几乎完全依赖光子学，这对InP供应链造成了前所未有的压力/依赖，而这种依赖在几个月前根本不存在。 随着铜互连触及上限，AI行业对AXTI的依赖本质上变成了对某家供应所有材料的小盘股的“单点故障”。 我看的不是营收数据，这种依赖关系简直荒谬，1.6T网络架构直接瘫痪。 我不确定企业接下来该怎么办，我确信他们会尝试囤积材料或签订多年协议，但无论如何，这两家公司掌握着所有主动权。
英文原文
That’s what’s so incredibly alarming, that the Western AI buildout might be held at choke point by an obscure $700m company like $AXTI and $SMTOY. InP substrates were incredibly niche for telecom, but it’s suddenly a national security emergency as it became the material used for the world’s lasers and optional interconnects. I’m not quite sure the top level vendors realized this or at least didn’t realize the scale of this issue (eg. Signing agreements with $LITE without going deeper into the material suppliers for Lite), as the supply chain is incredibly becoming choke pointed by two companies. Google’s TPU v7 ramp for example placed an unprecedented strain/dependency on the InP supply chain that simply didn't exist a few months ago, since they use OCS for their chip pods, which is almost all in on photonics. As we’ve hit the upper limits with copper, the AI industry's reliance on AXTI is essentially a "single point of failure” on some small cap that supplies the materials for everything. I’m not looking at revenue numbers here, this dependency is just absurd, the 1.6T networking just breaks. I’m not sure where companies go from here, I’m sure they’ll try and hoard materials or secure multi-year agreements but either way two companies hold all the cards.
-
磷化铟供应高度集中，成为AI基建关键瓶颈。
这是多方面的。然而，磷化铟(InP)的情况极其极端，因为 $AXTI 和 $SMTOY 控制了全球约 60%以上的这一关键供应。 行业从铜向光子学的转变，在整个建设过程中造成了这种结构性的单点故障。 超大规模数据中心中的每一个光收发器/互连设备，无论是 Google 的 TPU、NVIDIA 的共封装光学(CPO)路线图，还是 AWS/Meta 的网络，都需要基于磷化铟(InP)的组件。 这就是瓶颈的定义。 整个万亿美元 AI 建设的未来，取决于某只市值 7 亿美元的随机仙股和日本工业巨头的稳定性。
英文原文
It's multifaceted. However, InP is EXTREME because $AXTI and $SMTOY control roughly 60%+ of this critical global supply. The industry's shift from copper to photonics has created this structural single point of failure on the entire buildout. Every optical transceiver/interconnects in hyperscaler data centers from Google TPUs, NVIDIA's CPO roadmap, AWS/Meta networks requires InP-based components. This is the definition of a bottleneck. The future of the entire trillion dollar AI buildout depends on the stability of some random $700m penny stock and a Japanese industrial giant.
-
InP衬底成AI供应链瓶颈，$AXTI等公司议价权极大。
这由你自己去判断。我只是指出，两家公司（其中一家市值仅6亿美元）恰好是近期整个AI交易的关键瓶颈。历史上，磷化铟(InP)衬底是一种廉价的大宗商品。而现在……它们不再是了。如果每个花费数千亿美元购买GPU/ASIC的超大规模数据中心(hyperscaler)都无法在没有价值2000万美元的InP的情况下运转，那么像$AXTI这样的公司理论上可以将价格提高5倍，而这对于整体支出来说只是零头。你还需要考虑中国带来的巨大地缘政治风险，但如果整个建设进程被一两家公司的瓶颈所限制，60亿美元的市值通常只是微不足道的零钱。
英文原文
That’s for you to figure out. I’m just highlighting how two companies (one of which is worth $600m) happens to be the bottleneck of the entire AI trade in the near future. Historically InP substrates were a cheap commodity. And now… they’re not. If every single hyperscaler spending hundreds of billions of GPUs/ASICs can’t function without $20m of InP, companies like $AXTI can theoretically raise prices 5x and it would be a rounding error. There is a massive geopolitical risk with China for you to consider too but $600m MC is typically hilarious pocket change if the entire buildout is bottlenecked by one or two companies.