2025-12-23 · 供应链分析

解析微软Maia 300供应链，指出AAOI和MRVL为光模块核心受益者。

涉及标的： $AAOI $AMZN $COHR $LITE $MRVL $MSFT $NVDA $TSM

中文翻译

个人研究：$MSFT Maia 300 供应链。根据 $MRVL 的报告，瑞银(UBS)预计2027年前 Maia 出货量将超100万颗。预计 Maia ASIC 量产将带来超180亿美元流向供应商。资金流向如下： 1. $MRVL - 营收80-120亿美元（~30-40%利润率），作为主要受益者。他们设计数字底层架构（SerDes，互连），从 SK Hynix 购买 HBM4 内存，支付 TSMC 的 2nm 晶圆和 CoWoS 封装费用，并将成品模块卖给微软。当 Maia 在2026年底至2027年初开始量产时，鉴于其带来的巨大营收，Marvell 可能会迎来大幅重估。 2. 光网络架构：约30亿至40亿美元主要受益者：Applied Optoelectronics ($AAOI)，Coherent ($COHR)，Innolight Maia 集群需要1:1或更高的光收发器与GPU比例。对于1.6T速度（Maia 300所需），每个节点的光学内容估计超过3,000美元（每个加速器约6个800G/1.6T模块）。这里 AAOI 大获全胜：与市值700亿美元的 $MRVL 不同，$AAOI 是一家25亿美元的公司。即使只获得这块30亿美元蛋糕的20-30%（6-9亿美元），也将使其当前年化经常性收入(ARR)翻三倍，但具体份额取决于其德州磷化铟(InP)工厂的扩产情况（可能更多或更少）。因此，Applied Optoelectronics (AAOI) 和 Marvell 的光学 DSP 业务是隐形赢家，受益于连接这些巨型芯片所需的1.6T升级周期。像 $TSM 这样的公司会增加30-40亿美元营收，但相对于其1000多亿美元的营收来说，这只是中等体量。（仅说明 TSM 体量之大）硅基 BOM 拆解： 1. 计算逻辑 (ASIC) - TSM + Intel Foundry/18A Marvell (设计), TSMC, Intel, ARM/Synopsys? ($2,800 - $3,500) 2. 内存 (估计 HBM4? (6-8层堆叠), 估计值) SK Hynix, Samsung ($3,360 - $4,500) 3. I/O 与缓存 Die - TSMC N5/N4 Marvell, TSMC ($400 - $600) 4. 先进封装 - CoWoS-L / Foveros Direct TSMC, Intel ($900-$1300) 5. 散热方案 - 微流道冷板 / TIM ($200-400) 6. 杂项 (大电流电感, 电容) ($150-$200) 硅基 BOM 总计：$7,860 - $10,550 这非常粗略，未涉及无源器件/基板/盖板 (Ibiden, Unimicron)，测试与组装 (Advantest, Teradyne) 等。现在如果我们深入看光学 BOM 以及 $MSFT 在空心光纤和 OCS 方面的布局： DSP / SerDes IP: $MRVL ($800 - $1,200) SiPh 引擎 (CPO): Marvell / Intel ($600 - $900) 激光源 (ELS): $AAOI, $LITE ($300 - $600) 收发器组装: $AAOI, Innolight ($400 - $600) 光纤布线: 空心光纤 / MPO ($MSFT) ($200 - $300) 光学 BOM 总计~$2,300 - $3,600~占系统总成本的25% 我们已经确立 $LITE 和 Innolight 是光子学的两大主要玩家（在 $NVDA 和所有超大规模云厂商 ASIC 中）然而，$AAOI 的市值仅为 $MRVL 和 $LITE 的一小部分，作为小盘股，Maia 的机会对其股票更具变革性。另需注意，富邦研究(Fubon Research)估计 Maia300 芯片将于2026年底开始生产，约30万至40万颗，2027年增至120万至150万颗，但我们仅采用瑞银的~100万颗数据。（因此实际爬坡可能更多）但向1.6T的转变以及对空心光纤/OCS网络高功率激光器的特定需求是一个巨大的顺风。如果 AAOI 获得 Maia 光学 BOM 的一部分，他们到2027年仅从一个客户那里就能使 ARR 翻三倍或四倍（不包括 $AMZN）。风险：主要风险是资格认证失败。如果他们的激光器未能通过微软严格的可靠性测试，他们可能会被 $LITE 一夜之间取代。然而，Maia 300 的量产标志着超大规模云厂商 ASIC 交易的开始，关键供应商已就位。从2026-2027年底这一特定 ASIC 中受益最大的公司？看起来是 $AAOI, $MRVL, $TSM 和 SK Hynix。警告：这是基于公开材料+深度研究+映射的推测性研究+建模。没有公开的新闻证实这些信息，不要将其视为事实。

英文原文

Personal Research: $MSFT Maia 300 supply chain. On a $MRVL report, UBS est. 1M+ Maia by 2027. An est. of $18B+ is estimated to flow down to suppliers from Maia ASIC ramp. Here's where it goes: 1. $MRVL - $8-12B revenue (~30-40% margin) as the prime beneficary. They design the digital plumbing (SerDes, interconnects), buy the HBM4 memory from SK Hynix, pay TSMC for the 2nm wafers and CoWoS packaging, and sell the finished module to Microsoft. We'll likely see a large re-rating in Marvell when it comes time for Maia ramp late 2026, early 2027, especially given how much revenue this brings in. 2. The Optical Fabric: ~$3 Billion to $4 Billion Primary Beneficiaries: Applied Optoelectronics ( $AAOI ), Coherent ( $COHR ), Innolight Maia clusters require a 1:1 or higher ratio of transceivers to GPUs. For 1.6T speeds (required for Maia 300), the optical content per node is estimated at $3,000+ (approx. 6x 800G/1.6T modules per accelerator). Here AAOI Wins Big: Unlike $MRVL (which is $70B), $AAOI is a $2.5B company. Capturing even 20-30% of this $3B pie ($600M-$900M) would triple + their current ARR, but the amount captured is dependent on their InP Texas Fab scaling (could be a lot more or less). So Applied Optoelectronics (AAOI) and Marvell's optical DSP business are the silent winners, riding the 1.6T upgrade cycle required to interconnect these massive chips. Companies like $TSM gain another $3-4B revenue, but that's medium volume compared to their $100B+ revenue. (just speaking to how much of a giant TSM is) For the silicon BOM breakdown: 1. Compute Logic (ASIC) - TSM + Intel Foundry/18A Marvell (Design), TSMC, Intel, ARM/Synopsys? ($2,800 - $3,500) 2. Memory (est. HBM4? (6-8 stacks), estimate). SK Hynix, Samsung ($3,360 - $4,500) 3. I/O & Cache Die - TSMC N5/N4 Marvell, TSMC ($400 - $600) 4. Advanced Packaging - CoWoS-L / Foveros Direct TSMC, Intel ($900-$1300) 5. Thermal Solution - Microfluidic Cold Plate / TIM ($200-400) 6. Misc (High-Current Inductors, Caps) ($150-$200) Total Silicon BOM: $7,860 - $10,550 This is really high level doesn't go into Passive/Substrate/Lid (Ibiden, Unimicron), Test & Assembly (Advantest, Teradyne), etc. Now if we look deeper at optical BOM and what $MSFT is doing with hollow core fiber and OCS: DSP / SerDes IP: $MRVL ($800 - $1,200) SiPh Engine (CPO): Marvell / Intel ($600 - $900) Laser Source (ELS): $AAOI, $LITE ($300 - $600) Transceiver Assembly: $AAOI, Innolight ($400 - $600) Fiber Cabling: Hollow Core Fiber / MPO ( $MSFT) ($200 - $300) Total Optical BOM~$2,300 - $3,600~25% of Total System Cost We've established $LITE and Innolight as the two major players in photonics already (within $NVDA and every single hyperscaler ASIC) However, $AAOI MC is a fraction of $MRVL and $LITE as a microcap, making the Maia opportunity far more transformative for its stock. Someting also to note is Fubon Research est Maia300 chip is set to start prod in late 2026 with about 300,000 to 400,000 units, increasing to 1.2 to 1.5 million units in 2027, but we're just going with the ~1M figure from UBS. (so might ramp up more) But the shift to 1.6T and the specific requirement for high-power lasers for hollow-core/OCS networks is a big tailwind. If AAOI captures a portion Maia optical BOM, they triple or quadruple their ARR from one client alone by 2027 (not including $AMZN). Risk: The primary risk is qualification failure. If their lasers fail Microsoft’s rigorous reliability testing, they could be replaced by $LITE overnight. However, Maia 300 ramp signals the start of the hyperscaler ASIC trade and key suppliers are there. The companies that benefit the most from this specific ASIC late 2026-2027? Looks like $AAOI, $MRVL, $TSM, and SK Hynix. Warning: This is speculative research + modeling given public materials + deep research + mapping. There is no public breakdown and news of this information, do not treat it as a fact.

在 X 上查看原推 ↗