英伟达携手Coherent 扩产AI 光互联，黄仁勋称AI 是终极通用技术 - 股票，期货、外汇

家里蹲国仙 Lv.14

17-7-2026 09:37 AM 关注

支柱长老 | Credits 20605

・导体材料：磷化铟英文为Indium Phosphide
・黄仁勋在Sherman市的Coherent扩建厂房破土动工仪式上表示:“人工智能是终极通用技术。是根本,它影响着每一个行业”。获得5000 万美元CHIPS

赏

封装 , NPO

赞 0 踩 0

428阅读

家里蹲国仙楼主 2^#

17-7-2026 09:41 AM

支柱长老 | Credits 20605

Grok 4.5:

**中国主导磷化铟（InP）上游资源（铟原料），全球中游衬底生产高度集中于少数企业（日美为主），中国正加速国产化。**

### 1. 上游资源（核心原料：高纯铟 + 高纯磷）
磷化铟的核心是**7N级（99.99999%）高纯铟**（占InP质量约51%），铟是稀散金属，主要作为锡、锌冶炼副产物提取。中国是全球铟资源和生产绝对龙头：

- **中国掌控全球约70%精铟/原生铟产量**（储量占比也很高，主要分布在云南、广西、内蒙古等地）。2025年起中国对铟及InP相关出口实施严格许可/管制，进一步强化控制。

- **关键企业**（A股为主）：
  - **锡业股份（000960）**：全球原生铟龙头，储量/产量第一，稳定供应7N高纯铟，是磷化铟厂核心原料供应商。
  - **株冶集团（600961）**、**有研新材（600206）**：高纯铟产能强，7N级供应能力突出。
  - 其他：华锡有色、盛和资源等回收/伴生铟企业。
- 磷源：兴发集团等提供高纯红磷/磷烷。

**中国手握“源头护城河”**，海外（如日本住友、JX）高度依赖中国铟原料，出口管制已导致海外供应链收缩和价格上涨。

### 2. 中游：磷化铟衬底/外延/芯片（高壁垒环节）
全球高端InP衬底（尤其是大尺寸）高度集中，**90%以上产能由日美企业掌握**，但中国正在快速追赶（国产化率提升，6英寸突破中）。

- **全球格局**（衬底为主）：
  - **日本住友电工（Sumitomo）**：全球第一，份额约40-42%，内部消耗多。
  - **美国AXT（北京通美子公司）**：第二，份额约35-36%，主要在中国生产，受出口许可影响大。
  - **日本JX金属**：约10-13%。
  - 三家合计控制全球90%左右高端产能。

- **中国企业**（加速突破）：
  - **云南锗业（002423，通过鑫耀半导体）**：国内衬底绝对龙头，全球份额前三，已量产2-6英寸，扩产中（华为等大客户锁定）。
  - **三安光电（600703）**：IDM全产业链，衬底-外延-芯片一体化。
  - **先导科技/先导微电子**：垂直整合（原料+衬底），6英寸产能强。
  - 其他：有研新材、北京通美（AXT中国）、珠海鼎泰芯源等。

下游芯片/应用：源杰科技、三安、华为相关等，受益于800G/1.6T光模块、AI数据中心、激光雷达等需求爆发（供需缺口大，价格上涨）。

### 总结：资源主要“在中国手上”，但高端制造仍需突破
- **原料端**：中国强势垄断，战略优势明显（出口管制是重要筹码）。
- **制造端**：日美（住友+AXT）仍主导高端，但中国产能扩张快（订单排到2028年左右，缺口显著）。
- 整体产业链正向中国倾斜，尤其在AI/光通信驱动下。未来趋势是国产替代加速 + 价格维持高位。

信息基于公开市场报告和产业分析，实际情况可能随政策/扩产变化。

家里蹲国仙楼主 3^#

17-7-2026 09:50 AM

支柱长老 | Credits 20605

绝对是出大事了！！NVIDIA棄COUPE轉Tower NPO？

Claude Sonnet 5:

Got it — Indium Phosphide's role in photonics is genuinely central to modern semiconductor technology. Here's the detailed picture:

[size=150]Why InP works for photonics

InP has a direct bandgap (~1.35 eV), meaning electrons can recombine with holes and emit photons efficiently — unlike silicon, which has an indirect bandgap and is a poor light emitter. This single property is why InP (not silicon) dominates active photonic devices.

Its bandgap also happens to correspond almost perfectly to the 1310 nm and 1550 nm wavelength windows — the two bands where optical fiber has minimum signal loss and dispersion. This isn't a coincidence InP was chosen for; it's why those telecom wavelengths became the industry standard in the first place.

[size=150]Core device types built on InP

Laser diodes

Distributed feedback (DFB) lasers — single-wavelength, narrow-linewidth sources used as the "light bulb" for fiber-optic transmitters
Vertical-cavity surface-emitting lasers (VCSELs) in InP variants for longer-reach applications
Widely tunable lasers, essential for reconfigurable optical networks (DWDM systems)

Photodetectors

PIN photodiodes and avalanche photodiodes (APDs) that convert incoming light back into electrical signal at the receiver end
InGaAs, grown lattice-matched on InP, is the actual light-absorbing layer in most of these — InP serves as substrate and cladding

Modulators

Electro-absorption modulators (EAMs) and Mach-Zehnder modulators that encode data onto a continuous laser beam at high speed (up to 100+ Gbps per channel)

[size=150]Photonic Integrated Circuits (PICs) — the big shift

This is where InP has become especially important over the last decade. Instead of packaging separate discrete lasers, modulators, and detectors, manufacturers now integrate all of them monolithically on a single InP chip — similar in philosophy to how silicon integrated circuits combined transistors.

Here's a look at how those components sit together on an integrated InP chip:

Laser ───► Modulator ───► Waveguide ───► Detector
│ │ │ │
└───────────── Shared Indium Phosphide Wafer ─────────────┘

All of these blocks — laser, modulator, waveguide, detector — sit on the same InP wafer, sharing the same crystal lattice, which avoids the loss and cost of coupling separate discrete chips together with lenses and alignment optics.

[size=150]Why this matters industrially

Data centers and hyperscalers (Google, Meta, Microsoft, Amazon) are major consumers of InP-based transceivers for interconnects between servers and switches — bandwidth demand from AI workloads has made this a hot growth area.
Companies actively building InP PICs: Infinera, Coherent (formerly II-VI/Finisar), Lumentum, NeoPhotonics, and Ciena are among the major players.
Silicon photonics vs. InP: Silicon is cheaper and leverages CMOS fab infrastructure, but can't natively emit light (indirect bandgap), so silicon photonics chips often still need an InP laser "bonded" or hybrid-integrated onto them as the light source. This makes InP somewhat irreplaceable even in silicon-dominant designs.

[size=150]Manufacturing challenges

InP wafers are smaller (typically 2–4 inch, moving to 6 inch) and more brittle than silicon's 12-inch wafers, making InP fabs lower-yield and more expensive per chip. This is a genuine bottleneck limiting how fast InP photonics can scale compared to silicon-based electronics.

Want me to go deeper into any part of this — like the physics of how the direct bandgap actually produces stimulated emission in the laser, or how InP compares against other photonic material platforms like silicon nitride or thin-film lithium niobate?

家里蹲国仙楼主 4^#

17-7-2026 10:00 AM

支柱长老 | Credits 20605

Claude Sonnet 5:

Good follow-up — this gets at a specific, current packaging trend. Let me lay out why FAU precision is such a critical differentiator for NPO specifically.

## What NPO is, in this context

Near-Packaged Optics sits between two extremes: pluggable transceivers (optics far from the switch ASIC, long lossy electrical traces) and Co-Packaged Optics/CPO (optics fused directly onto the same substrate as the ASIC, hard to service or replace). NPO places the optical engine physically close to the switching ASIC but keeps it as a separate, detachable module — this arrangement significantly increases data bandwidth and energy efficiency while maintaining system flexibility.

Because the optical engine sits so close to the switch, optical signals must achieve efficient coupling and stable transmission within extremely limited space — and that's exactly where FAU precision becomes the bottleneck.

## Why FAU precision matters more in this architecture

**Tolerances are unforgiving.** Fiber cores are tiny relative to the alignment budget: the V-groove formation accuracy is controlled at the nanometer level, requiring positioning accuracy on the order of sub-microns (0.1 µm) relative to the fiber's 125 µm outer diameter. Any deviation directly hits coupling efficiency to the waveguide.

**Higher channel density.** NPO-class designs push far more fiber channels into a compact footprint than legacy pluggables. Higher-fiber-count FAUs, such as those with 32 or 64 fibers, are needed to accommodate the higher densities in these near-chip architectures, and achieving low coupling losses requires tight positional tolerances and well-centered cores.

**Serviceability is the actual advantage over CPO.** This is the key structural edge: because NPO keeps the optical engine as a discrete, connectorized module rather than permanently bonding it to the ASIC package, the FAU can act as a real detachable interface. The fiber array serves as a precise alignment tool, ensuring the accurate alignment of multiple fibers into the waveguides of the PIC or other compact optical devices, while still connecting to a standard fiber optic connector like MPO on the other end. That lets a failed optical module be swapped without touching the switch board — something full CPO struggles with.

**Two coupling geometries, same precision demand.** Manufacturers now offer both approaches for this class of module: edge-coupled FAUs that route light through the edge of the PIC, and vertically coupled FAUs that enable surface-based coupling for higher density, advanced packaging, and next-generation designs.

**Permanent, stable bonding once aligned.** Since NPO modules are compact and thermally/mechanically constrained, the alignment has to survive years of thermal cycling in that tight space: after precise alignment is achieved between the FAU and the PIC, UV-curable resin is applied to secure the two together, ensuring long-term optical alignment stability.

## The net advantage, in short

NPO's edge over pluggables comes from shorter electrical reach (less power spent on SerDes/retimers, better signal integrity). Its edge over full CPO comes from keeping the optics field-replaceable. **Both of those advantages only exist if the FAU can deliver near-perfect fiber-to-waveguide alignment in a small, dense, detachable package** — which is why FAU manufacturing precision (sub-micron V-groove tolerances, high channel counts, low insertion/return loss) has become one of the actual gating technologies for NPO adoption, not just a supporting detail.

Want me to go deeper into edge-coupling vs. vertical-coupling tradeoffs, or into who's leading FAU manufacturing for this space (Corning, Orbray, Focuslight, etc.)?

家里蹲国仙楼主 5^#

17-7-2026 10:09 AM

支柱长老 | Credits 20605

Claude Sonnet 5:

Here's the current competitive landscape across those companies, based on recent industry moves:

## Corning (US) — the dominant US anchor

Corning has moved from being just a fiber supplier to building a full CPO ecosystem. Its FAU lineup includes both edge-coupled units, designed for today's high-performance photonic systems routing light through the edge of the PIC, and vertically coupled units for surface-based coupling in higher-density, next-generation designs. It also makes an ultra-compact FAU using reduced-clad, bend-insensitive fiber specifically for 90-degree, surface-coupling designs where vertical space is limited to just a few millimeters.

The bigger story is Corning's new **Glass Bridge** technology, unveiled recently: it uses ion-exchange waveguides formed inside glass to directly bridge the huge size mismatch between on-chip waveguides (hundreds of nanometers wide) and fiber cores (several micrometers wide) using wafer-based ion-exchange waveguide technology to create optical pathways within the glass. This sent Corning's stock up sharply and triggered rallies in Taiwan optical-supply-chain stocks, though Taiwanese FAU makers noted Glass Bridge's edge-coupling approach differs from the surface grating-coupling ecosystem Nvidia favors, so the two are expected to coexist rather than compete directly.

The clearest signal of Corning's position: Nvidia is investing up to $3.2 billion in Corning as part of a deal to build three new factories focused on AI, explicitly aimed at scaling U.S.-based manufacturing of the optical connectivity needed for next-generation AI infrastructure.

## US Conec (US, Hickory, NC) — the connector/ferrule backbone

Less famous than Corning but arguably just as load-bearing. US Conec designed the MT ferrule at the core of the industry-standard MTP/MPO connector format, and has pushed precision further with its PRIZM expanded-beam ferrules — monolithic ferrules combining fiber microholes, lenses, and mechanical alignment features for robust connections that need little to no cleaning or inspection.

For the CPO to NPO transition specifically, US Conec is now co-developing the next-generation **MMC connector** and **TMT ferrule** format alongside Corning, Fujikura, and Sumitomo Electric, and has separately partnered with Japan's Hakusan and Sanwa Technologies to multi-source that same VSFF (very small form factor) ecosystem — a sign the industry is deliberately avoiding single-vendor lock-in for this critical component. Corning's own MMC connectors use US Conec's PRIZM TMT ferrule under license, showing how tightly these two US firms are working together on the same standard.

## Orbray (Japan) — precision-machining pedigree

Orbray's edge comes from decades in ultra-precision jewel machining (diamond, sapphire, ruby components), which translates directly into fiber-array tolerance work. Its V-groove fiber arrays are held to sub-micron connection-loss tolerances, and its "optical lead frame" product goes a step further by integrating the waveguide element and fiber array into one structure, keeping connector loss below 0.1dB.

Notably, Orbray just signed a supply-chain partnership (announced this April) specifically to scale up fiber-array production capacity for AI infrastructure demand, citing its proprietary alignment and bonding processes as the core competitive asset customers are buying into.

## Focuslight (China) — volume manufacturing of the supporting micro-optics

Focuslight's role is slightly different: rather than the full FAU assembly, it supplies the engineered V-groove arrays and protective lids that go into FAUs, plus the microlenses and microlens arrays used elsewhere in NPO modules to shape and couple beams between fiber and optical engine. Its pitch is high-volume, automated manufacturing at high channel counts rather than boutique precision.

## Other notable players worth knowing

- **Teramount** (Israel) — pushing a detachable fiber-to-chip connector supporting 32+ mixed single-mode/polarization-maintaining fibers at 127µm pitch, aimed squarely at the density problem next-gen CPO switches will hit.

- **GlobalFoundries** — not a FAU maker itself, but showcased a complete CPO ecosystem with Corning and test-equipment maker EXFO, and separately launched its own SCALE optical-engine platform.

## The bottom line on US positioning

The US side of this supply chain really rests on two companies working in tandem: **Corning** for the glass/fiber/FAU precision components (now backed by direct hyperscaler capital), and **US Conec** for the connector and ferrule standards that make those FAUs field-serviceable. Japan (Orbray) holds an edge in raw ultra-precision machining heritage, while China (Focuslight) compete more on manufacturing scale and cost.

家里蹲国仙楼主 6^#

17-7-2026 10:16 AM

支柱长老 | Credits 20605

Orbray Co., Ltd. is a privately held company and its stock is not currently traded on any public stock exchange.

家里蹲国仙楼主 7^#

17-7-2026 10:22 AM

支柱长老 | Credits 20605

从1亿飙到390亿，NPO光互连赛道升温，产业链最新进展一览

港股IPO观察
07-10 14:17 来自北京市

你有没有想过，训练一个万亿参数的大模型，光靠一颗芯片是搞不定的。成千上万颗芯片组成集群，24小时不间断交换数据。芯片算得再快，如果数据在集群里传不动，整座算力工厂的效率照样被拖垮。

这个数据传输的环节，就是光互连技术。而就在昨天，这个赛道发生了一件大事。

一、华为牵头，国内首个NPO光互连MSA正式启动

7月9日，北京。由全球计算联盟（GCC）指导、Open AI Infra社区主办的"超节点与GW级AIDC技术论坛暨Open AI Infra社区半年工作会议"上，华为联合中国移动研究院、京东云、百度、中国电子技术标准化研究院等20余家产业链伙伴，共同启动OPEN NPO项目，发起国内首个NPO光互连MSA。什么叫MSA？就是多源协议（Multi-Source Agreement），说人话就是产业链上下游一起定规矩——机械尺寸怎么设计、电气接口怎么接、测试标准怎么测，让不同厂商的产品能互相兼容。这是新技术从实验室走向产业化的关键一步。

这次发起单位覆盖了算力设备、光模块、电连接器、光电芯片等关键环节。华为计算公众号公布的首批发起单位名单里，光迅科技、华工正源、华丰科技、立讯技术、纳真科技、曦智科技、傲科光电子、集益威半导体等国内企业都在列。

值得注意的是，这不是孤立事件。今年5月21日，中国信通院联合华为、腾讯、阿里云等，在光互联论坛（OIF）成功立项了《12.8Tb/s光电近封装模块》国际标准。现在国内MSA和国际标准"两条腿走路"，说明我国在NPO领域已经从跟跑变成了并跑。

按照规划，OPEN NPO项目预计2026年三季度发布首版技术规范，并完成NPO及电连接器产品样品适配和全场景测试；2027年上半年推动NPO技术规模化商用。

二、NPO到底是什么？为什么偏偏是现在？

很多老铁可能还没搞懂，NPO和之前炒得火热的CPO有什么区别。

简单打个比方：传统可插拔光模块，就像外接U盘，插在交换机面板上，数据要走"芯片→PCB板铜走线→连接器→光模块"这么一长串路。距离越远，电信号损耗越大，功耗越高，也越不稳定。

NPO（近封装光学），就是把光引擎搬到离交换芯片更近的位置，把高速电信号的传输距离从几十厘米压缩到几厘米。这样做的好处很明显：功耗降下来，信号质量提上去，同时保留了光引擎可以单独更换的灵活性。

CPO（共封装光学）走得更远，直接把光学引擎和交换芯片封装在一起，密度更高、功耗更低，但维修难度大，对制造工艺的要求也更高。

XPO（扩展封装光学）则是另一种思路，通过特殊封装把密度提得更高。华工正源在OFC 2026上首发的12.8T XPO模块，密度达到了传统OSFP封装形式的四倍。业内普遍把NPO看作近中期的过渡方案——比传统可插拔光模块更省电更快，比CPO更容易量产和维护，供应商也更丰富。TrendForce集邦咨询的数据显示，CPO与NPO合计市场规模将从2025年约1亿美元，增长到2030年的390亿美元以上。

华福证券研报指出，NPO制造工艺贴近现有产线，支持芯片与光引擎解耦设计，降低了供应链协同门槛。

三、13家A股公司，在NPO赛道上各自做了什么？

以下信息均来自各公司官方公告、财报、业绩说明会及权威行业媒体报道。

1. 华工科技：3.2T NPO与阿里云联合点亮

华工科技子公司华工正源，是这次OPEN NPO项目的发起单位之一。今年3月的OFC 2026上，华工正源与阿里云联合完成了全球首款3.2T NPO模块的成功点亮，并且已率先应用于行业头部客户。

同月，华工正源还加入了XPO MSA并成为创始成员，同步全球首发了12.8T XPO光模块。

财务数据方面，华工科技2026年一季度营收42.66亿元，同比增长27.13%；归母净利润6.38亿元，同比增长55.76%。不过扣非净利润为3.73亿元，同比增长20.51%，与净利润增速差距明显，差异主要来自持有金融资产的公允价值变动收益。

7月9日，华工科技获融资买入14.86亿元，融资余额111.89亿元，占流通市值比例7.00%，超过近一年90%分位水平。

光芯片方面，华工科技具备硅光芯片从设计到模块集成的自研能力，推出了用于1.6T光模块的单波200G自研硅光芯片。

2. 光迅科技：业界首家完成3.2T NPO全系统验证

光迅科技也是OPEN NPO项目的发起单位之一。今年3月的OFC 2026上，公司推出了全球首款3.2T硅光单模NPO模块，并已完成送样测试。更关键的是，公司在国内头部CSP厂商完成了3.2T NPO全系统验证，成为业界首家实现这一突破的光模块厂商。

光迅科技是国内唯一同时拥有InP、GaAs、SiP三大光芯片平台的企业，打通了芯片-器件-模块-子系统全产业链。

在CPO领域，光迅科技2023年发布了可插拔CPO ELSFP自研光源模块。2025年提出了35亿元定增预案，用于算力中心光连接及高速光传输产品生产建设，投产后将形成武汉基地高速光模块年产能499.2万只（含1.6T光模块）。

业绩方面，光迅科技2025年前三季度营收85.32亿元，同比增长58.65%；净利润7.19亿元，同比增长54.95%。

3. 中际旭创：全球光模块龙头，800G规模化出货

中际旭创2025年上半年营收147.89亿元，同比增长36.95%；归母净利润39.95亿元，同比增长69.40%。公司800G光模块已经规模化出货，1.6T产品紧跟行业节奏。

公司布局了硅光、ZR相干光模块、LPO、OCS等多项光互连技术，始终保持与AI和云数据中心等领域头部厂商的深度密切合作。

4. 新易盛：6.4T NPO方案亮相，但Q1净利润环比下滑

新易盛在OFC 2026期间推出了高密度6.4T NPO解决方案，采用硅光技术实现32条200Gbps通道聚合带宽。公司在互动平台表示，产品技术体系覆盖传统可插拔光模块、LPO/LRO、NPO及CPO等多种互联形态。

财务数据方面，新易盛2025年上半年营收104.37亿元，同比增长282.64%；归母净利润39.42亿元，同比增长355.68%。但财新报道指出，2026年一季度新易盛净利润同比增长76.8%，环比却下滑超过10%，这是时隔九个季度以来首次出现环比下滑。同时，公司上半年存货从41.32亿元增至59.44亿元，增幅明显，主要因备货增长。境外销售占比高达94.5%。

5. 仕佳光子：12.65亿加码高速光芯片

仕佳光子是国内少数掌握DFB激光器芯片全产业链生产能力的企业，数据中心硅光用CW DFB光源及器件已实现小批量供货。2025年开发出数据中心用O波段CWDM-4的100G EML激光器，正在客户送样验证中。

2026年4月，公司公告拟投资12.65亿元建设高速光芯片与器件开发及产业化项目。光芯片是NPO/CPO产业链里最关键的上游环节之一。

6. 光库科技：纳入MSCI中国指数

光库科技主要产品包括MPO光纤连接器、铌酸锂调制器等。MPO连接器是NPO/CPO架构下光纤管理的核心部件。2026年5月13日，MSCI公布指数季度调整结果，光库科技被新纳入MSCI中国指数，5月29日生效。

7. 炬光科技：激光元器件供应商

炬光科技专注于高功率半导体激光元器件、激光光学元器件，产品广泛应用于光通信、激光雷达等领域。在NPO/CPO产业链中，高功率CW激光器是核心光源之一。

8. 华丰科技：光连接器仍处市场布局阶段

华丰科技是OPEN NPO项目的发起单位之一，高速背板连接器已量产并广泛应用于数据中心服务器、交换机。公司在研的特种光连接器产品，可应用于高速光模块核心器件及光互连系统传输。不过公司在2026年5月的互动平台回复中明确表示，目前在光连接器领域仍处于市场布局阶段，相关产品的整体市场份额暂未形成规模。448Gbps产品正在预研中。

9. 立讯精密：昨天刚在港股敲钟

立讯精密子公司立讯技术是OPEN NPO项目的发起单位之一。2026年5月26日，立讯技术与Marvell签署战略合作备忘录，聚焦高端光模块及铜缆高速互连，明确光铜并进策略。公司800G系列光模块已实现成熟量产，1.6T系列进入规模化商用初期。

就在昨天（7月9日），立讯精密正式在港交所主板挂牌上市，成为2026年以来港股募资规模最大的IPO，募资约242.66亿港元。发行价63.28港元，收盘报62.3港元，跌幅1.55%。基石投资者包括淡马锡、GIC、阿布扎比投资局、高瓴、腾讯等20余家顶级机构，合计认购15亿美元。

10. 中天科技：光模块+空芯光纤双线布局

中天科技在OFC 2026上首次公开展出了基于单通道200G技术的1.6T硅光模块，800G产品系列涵盖DSP和LPO线性模块及AOC。公司自2024年开始研发空芯光纤，2026年5月业绩说明会上表示已实现核心性能突破，衰减最低可达0.08-0.2 dB/km，并实现了批量供货。

11. 长盈通：特种光纤产业链一体化

长盈通专注于特种光纤和光纤环器件，打通了环-纤-胶-模块-设备一体化微型产业链。产品主要用于军工配套，特种光纤技术在数据中心高速互连等领域有应用前景。

12. 兴森科技：1.6T光模块PCB量产爬坡

兴森科技子公司北京兴斐专注于高阶HDI、类载板（SLP）和CSP封装基板。2026年5月8日业绩说明会上，公司表示北京兴斐1.6T光模块产品板目前处于量产爬坡阶段，并在同步开展多家客户的验证导入工作。

13. 长飞光纤：光互连组件收入增长48%

长飞光纤2025年营收142.52亿元。2026年一季度净利润同比增长226%，扣非净利润同比增长966%。公司光互连组件（包括光器件与模块）收入同比增长48%，毛利率接近40%。子公司博创科技向多家国内外互联网客户批量供货25G至400G速率的中短距光模块。

四、NPO赛道的一些冷思考

看完这13家公司的布局，有几点值得冷静看待。

第一，NPO不是新概念，而是技术演进的中间态。TrendForce预计，CPO/NPO市场在2028到2029年才会明显加速，可插拔光模块仍将长期并存。短期别指望一夜爆发。

第二，各家进展差异很大。有的已经量产出货，有的还在样品验证，有的仅处于预研阶段。光模块行业技术迭代快，今天的领先不代表明天的优势。

第三，上游核心环节仍存瓶颈。NPO/CPO大规模商用面临良率、可维修性、光纤连接器标准和InP激光供给等挑战。

第四，业绩分化明显。有的公司净利润暴增几倍，有的扣非增速远低于净利润增速（靠非经常性损益撑场面），有的环比已经出现下滑，有的存货大幅增长。看财报不能只看 headline 数字。

第五，行业竞争激烈。光模块赛道已经涌入大量玩家，价格战风险始终存在。技术路线选择失误或产能扩张过快，都可能导致盈利能力下滑。

五、结语

光互连正在从数据中心的配套环节，变成AI算力基础设施的核心讨论。昨天OPEN NPO项目的启动，加上5月OIF国际标准的立项，意味着NPO技术从各自为战走向统一规范，这对产业链是好事。

对中国企业来说，参与标准制定，意味着在光互连领域从跟着别人走，变成了一起定规矩。但这只是起点，能不能把标准优势转化成产品和订单优势，还得看各家的真功夫。

最后问大家一个问题：你觉得NPO、CPO、XPO这三条技术路线，哪条会先大规模商用？欢迎在评论区聊聊你的看法。