Mr. Vinyl 美股投资之路

The most asymmetric AI supply chain setup atm is $AEHR. A $1B company with a literal monopoly. Seen it gain some traction on X these past few weeks/months, but not really seen a proper deep-dive. They make burn-in test systems — equipment that stress-tests chips at full operating power before they ship. Every GPU, every custom ASIC, every AI accelerator inside a hyperscaler data center has to be tested. You cannot ship a $10K chip that fails after 200 hrs. $AEHR tests entire 300mm wafers simultaneously — at up to 3,500 watts per wafer, thousands of amps of current. No other company on earth does this. $TER and Advantest are the giants of chip testing. They test chips individually or in small batches. $AEHR tests entire wafers before packaging. It's a completely different point in the manufacturing process — and it means defects caught before expensive packaging. They also just became the first company to ship wafer-level burn-in systems for AI processors. That's an actual $14M order delivered in Feb'26. The two platforms driving the AI thesis FOX-XP (Wafer-Level Burn-In) → Up to 9 × 300mm wafers in parallel → 3,500W per wafer — unmatched → $14M order received Feb 26 for AI processors → Silicon photonics follow-on order early March → Used for: AI processors, SiC power, optical chips Sonoma (Package-Level Burn-In) → Acquired via Incal Technology → Up to 2,000W per device — for packaged AI ASICs → First hyperscaler production win: Feb 11, 2026 → 10 companies visited HQ to evaluate → "Very large expansion" of orders guided for H2 2026 Three orders in 30 days. Two platforms. Both now confirmed in AI production. But the consumables angle is what most people miss Every system $AEHR sells locks in a recurring revenue stream. WaferPak contactors, DiePak carriers, and BIM consumables are proprietary, device-specific, and must be replaced regularly. This is razors-and-blades applied to semiconductor test equipment. Each FOX-XP system sold = years of consumable revenue at high margins. The installed base compounds. FY22–FY24: SiC/EV drove explosive growth. Revenue went from $22M → $65M. Gross margins hit 50%. Operating margins 20%+. FY25: EV programs delayed across the board. OEMs pulled back on SiC orders. Revenue fell to ~$60M then kept falling. Q2 FY26: $9.9M revenue. Gross margin: 29.8%. The stock got obliterated. Down 80%+ from peak. And then the AI orders started coming in. Q2 FY26 revenue: $9.9M ← confirmed trough H1 FY26 revenue: $20.9M (−21% YoY) H2 FY26 guidance: $25–30M revenue H2 FY26 bookings guidance: $60–80M That bookings number is the key figure. Management was explicit: the $60–80M is based on specific customer forecasts provided to $AEHR — not internal projections, not aspirational targets. Customers told them what they plan to order. Bookings in H2 FY26 ship and recognize as revenue in FY27 (which starts June 2026). FY27 is the inflection year. Management has stated the AI processor TAM is 3–5× larger than the SiC/EV market that drove the FY22–24 peak. Think about what that means: - Peak SiC revenue: ~$65M - If AI TAM is 3–5×: $195–325M revenue potential - Current market cap: ~$400M You're paying ~1.2–2× peak revenue for a company at trough, with a confirmed monopoly position, zero debt, and three orders in the last 30 days. The balance sheet is a genuine differentiator Zero long-term debt. ~$31M cash. No equity raises since the downturn. Compare that to $ENTG with $3.4B net debt from the CMC acquisition, or $LWLG burning $21M/year pre-revenue. $AEHR can execute without diluting you. But it's not a layup 1. Extreme lumpiness. Systems are $3–5M+ each. One delayed order swings quarterly revenue 30%+. This has happened multiple times. 2. Customer concentration. A handful of hyperscalers likely represent 80%+ of near-term bookings. Any program cancellation is devastating. 3. Gross margins still compressed. 29.8% last quarter vs. 50% at peak. Recovery requires volume ramp to arrive on schedule. 4. SiC is still dead weight. Legacy EV exposure hasn't recovered and occupies capacity. TLDR: $AEHR has a monopoly on a process that is physically required to ship AI processors at scale. Revenue is at a confirmed trough. Three orders have arrived in 30 days. Management has reinstated specific bookings guidance ($60–80M H2) based on customer forecasts. FY27 is the inflection. The balance sheet is clean. The TAM is 3–5× the peak that drove the last cycle.

如果你想真的了解光通讯产业链，那以下的内容不容错过——一图了解全产业链。 文章稍长，但是把光通讯的每个环节都讲的清楚明白需要费一些笔墨。 光通讯相比较电信号本质上还是一种信息传输的方式。只是光通讯更加高效，可以传递的信息量/速度都远优于传统的电缆。现在数据中心不可逆的趋势，就是需要传递的数据量的指数级增长。这个背景下，光通讯的流行和应用是势之必然。 —--------------------------- 整条链我分成了6个板块，我们一个一个来看。 1️⃣原材料，基本中国是龙头 铟（Indium）是一种稀有金属，主要产在中国，全球产量里中国占60%以上。铟本身没什么特别用处，但和磷结合之后，就变成了磷化铟（InP）——这个组合有一个神奇的物理特性：通电之后会发出激光。 镓（Gallium）也是稀有金属，中国2023年宣布出口限制，和砷结合变成砷化镓（GaAs），也能发光，用在短距离的场景。 做一个类比：原材料就像是"面粉"。面粉本身不是食物，但没有面粉就做不出面包。铟和镓就是制造激光器的"面粉"。 2️⃣衬底层 把铟和磷在极高温度下熔融、结晶、切片，就得到一片圆形的薄晶圆，叫做磷化铟衬底（InP Substrate）。直径约4-6英寸，看起来就像一片黑色的圆玻璃。 它是所有后续加工的基础材料，就像做芯片要先有硅晶圆一样。没有这片衬底，后面的激光器根本没有东西可以"长"在上面。 衬底就是"白纸"。在上面可以画出激光器的电路结构。 $AXTI 是全球最大的白纸供应商，占全球60-70%的份额。另一家是日本住友。 3️⃣外延/芯片 拿到InP衬底之后，用一种叫外延生长（Epitaxy）的技术，在这片晶圆上一层一层地沉积几纳米厚的半导体薄膜，形成精密的量子结构。这个过程就像在白纸上精密印刷电路。 这步做完之后，晶圆上就有了激光器的"基因"——一个可以在通电后发光的微小结构。然后把晶圆切成一个个极小的芯片颗粒，每一颗就是一个激光芯片（Laser Die）。 外延相当于"在面团上发酵"，让基础材料产生功能性结构。这一步技术门槛极高，良率只有15-50%，大量芯片在这步就报废了。 4️⃣激光器——让“信号”发光 激光器就是把上一步做好的激光芯片封装成一个可以使用的元件。通电之后，它会发出特定波长的红外激光——肉眼看不见，但光纤最爱传这个波段。 激光器是灯泡，光模块是台灯。没有灯泡，台灯就是一个空壳。 5️⃣光模块——光电世界里的翻译官 光模块（Optical Transceiver）是一个大约手指大小的小盒子，直接插进服务器或交换机的插槽里。它里面装着：激光器（发光）、光探测器（接收光）、驱动芯片、镜头和光纤接口。 它的作用是双向翻译：服务器发出电信号 → 光模块把它变成光信号打进光纤；光纤传来光信号 → 光模块把它变回电信号送给服务器。 光模块是连接两个世界的翻译官，他把光世界的语言和电世界的语言来回翻译。 6️⃣交换机层 ——指挥交通、分拣信息 交换机（Switch）是一台专门负责"分发数据"的设备，长得像一个扁平的大铁箱子，正面插满了几十上百个光模块插槽。 在AI数据中心里，成千上万台GPU服务器要互相通信——比如一个大模型训练任务，可能同时用到1万块GPU，它们之间每时每刻都在交换中间计算结果。交换机就是负责把这些数据包准确、快速地送到对的GPU。 一台顶级AI交换机（比如博通的Tomahawk系列）上面可以插128个800G光模块，每秒总吞吐量超过100Tb。 交换机很像我们现实世界的快递中转站。每个包裹（数据包）进来，它看一眼地址，立刻分拣到正确的出口。速度极快，几纳秒就完成一次分拣。 —--------------------------- 最后的最后，就是大客户们买单了。 看完这这链条，再炒CPO光通讯，心里是不是有谱多了？

昨晚的四巨头财报总结了一下： $GOOGL Cloud 增长 63%，Pichai 说收入本可以更高，只是受算力约束。 $MSFT Azure 重回 40% 增速，Copilot 付费席位继续增加。 $AMZN AWS 增长 28%，自研芯片年化收入突破 200 亿美元。 $META 广告变现已经验证，但 AI 独立变现还是比较模糊的。 Pichai 说算力供给跟不上需求，微软管理层Nadella 说 Azure 需求比产能增长还快。 这些话从产业最前线说出来，比任何分析师报告都更有说服力，现在的算力是供不应求的 谁能拿到更多的算力芯片、谁能搞定更多的数据中心能源配额，比如OpenAI在AWS锁定的2吉瓦算力，谁就能吃下更多的市场份额。 算力即权力的逻辑在2026年依然坚不可摧。 芯片、HBM、光互连、电力、液冷、数据中心设备，仍然是最硬的主线。