(Scroll down for English version)

Natgate × POET：一个市场还没看懂的AI技术爆发点

//

最近在klsescreener留言区看到有人post出natgate和poet合作的新闻稿，引起了我的兴趣，因为之前在研究natgate的时候，ai就已经提醒过我natgate未来的爆发点很可能不是ai server组装而是光模块。因为有了这层认知，再看到最近natgate跟poet这个在光模块领域的破坏式创新者合作，就自然觉得肯定有点东西。于是，我就花了三小时（本来也不打算花这么久时间）研究他们这次的合作，主要针对poet在光模组上被动对齐的这个颠覆式创新技术如何让natgate一个ems公司也能承载以前只有osat才能做的制程工序，以及这对于natgate日后的估值会有多大影响，这两件事情上。natgate和poet的合作这件大事，各大新闻媒体都还没报道，各大投行分析师也许也还没注意到它的重要性。我想在这个时间点，除了公司的老板和员工以外，全网应该也只有我一人注意到这个光模组的技术突破对于natgate会有多大的利好。昨天，我延续一贯作风，直接把一大串我的ai对谈的原始记录张贴出来，无意中形成了一种筛选制度，只有懂行的人才愿意点进来看。在ai这么方便的年代，另外再叫ai写一篇像样的文章其实只是举手之劳，所以今天我决定重新post过。这次是有双语的ai文章，一大串的ai对谈记录（因字数太多，只能分开发布）还有你现在读着的，我亲手写的序，我想这应该就是日后我会呈现的格式。你只需要读以下ai文章即可，真的很感兴趣才去翻看ai对谈记录吧。

//

当市场谈论AI的时候，几乎所有焦点都落在 GPU，尤其是 NVIDIA。仿佛只要有更多算力，AI就会自动变得更强。

但现实没有那么简单。

AI系统真正的瓶颈，往往不是“算”，而是“传”。当成千上万台服务器一起训练模型时，问题变成：这些机器之间，如何以足够快的速度交换数据？

这正是光模块存在的意义。

被忽略的关键：光模块到底是什么？

如果用一个简单的比喻来理解，GPU是大脑，而光模块就是神经系统。

服务器内部处理的是电信号，但电信号在长距离和超高速下会遇到严重问题：发热、衰减、效率下降。因此，在大型AI数据中心中，数据必须转换成光信号，通过光纤传输，再在另一端转回电信号。

这个“电 ↔ 光”的转换过程，就是光模块完成的。

也就是说，没有光模块，再强的GPU也无法形成真正的算力集群。它不是一个可选组件，而是整个AI基础设施的必要条件。

为什么这个赛道突然变得重要？

过去几年，AI基础设施正在经历一个非常明显的升级周期：从400G到800G，再到现在正在推进的1.6T。

这个数字代表的是数据传输速度。随着AI模型规模爆炸，服务器之间需要交换的数据量也在指数级增长。如果传输速度跟不上，再多的GPU都会被“堵在门口”。

这带来一个很直接的结果：光模块的需求，几乎是跟着GPU一起增长的，而且是刚性需求。

也正是在这个背景下，像 POET Technologies 这样的公司开始进入视野。

POET真正改变了什么？

光模块并不是新东西，难点一直在“封装”。传统做法中，最复杂的步骤是光学对准：工程师需要让芯片发光，再一点一点调整位置，让光信号精准进入光纤。这种方法成本高、速度慢，而且非常依赖高端封测厂（OSAT）。

POET的突破，在于它把这个过程“预先解决了”。

它的核心技术叫 Optical Interposer，可以理解为一种预先设计好的光学底座。光路已经被刻在芯片结构中，当其他组件被放上去时，会自动对准。

这带来一个极其重要的变化：
原本属于“封装”的复杂过程，被简化成了高精度组装。
换句话说，原本需要昂贵设备和长时间调校的步骤，现在更像是一次精密的“贴装”。

Natgate的角色为什么因此发生变化？

这就引出了这次合作的另一方：NationGate Holdings。

传统上，像Natgate这样的EMS公司主要做的是电子组装，例如服务器、PCB板等。这类业务的特点很明显：规模大、毛利低、可替代性高。

而在POET体系下，Natgate做的事情开始变得不一样。

在生产流程中，Globetronics 仍然负责晶圆级处理，也就是把原始芯片从晶圆切割、处理成可以使用的“核心部件”。这一部分仍然属于典型的半导体封测领域。

但当这些核心部件进入Natgate的生产线之后，事情发生了变化。

Natgate不只是把零件拼在一起，而是在完成两个关键动作：第一是高精度贴装（类似倒装芯片），第二是光纤连接（FAU耦合）。尤其是后者，是整个光模块中最容易出问题、最影响良率的步骤。

在传统技术体系下，这些工作本来是由OSAT或者顶级光模块厂完成的。但因为POET的“被动对准”设计，Natgate能够用自身擅长的精密自动化设备完成这些步骤。

这意味着什么？

意味着Natgate正在从“组装厂”，逐渐变成一个具备“光电集成能力”的制造商。

为什么利润看起来不大，却很重要？

如果只看数字，这个故事很容易被低估。

假设Natgate原本的AI服务器业务带来100单位利润，而未来光模块业务贡献20到30，看起来只是增加了20%-30%。

但这个比较忽略了一个关键问题：利润的“质量”。

服务器组装业务，本质上是低毛利、高周转。利润很容易受到汇率、零件价格甚至客户订单波动的影响，稳定性很差。

而光模块相关业务，属于技术驱动型。毛利率更高，而且一旦通过客户认证，很难被替换。

更重要的是资本市场的看法。

一个传统EMS公司，可能只能获得10到15倍的市盈率。但一旦被市场认定为光通信或半导体相关企业，估值区间可能上升到25甚至40倍。

于是就出现了一个非常关键的变化：
利润只是小幅增长，但估值体系发生了改变。

这也是为什么，在某些情况下，利润增加20%，股价却可能翻倍甚至更多。

那会不会很快被别人复制？

这是一个很合理的担心。如果POET的技术让封装变得更简单，那其他EMS公司是不是也可以快速进入？

短期来看，难度比想象中大。

首先，POET并不是把技术开放给所有人，而是通过“代管”（consignment）模式，把核心设备和组件放在合作伙伴的产线上。这意味着即使其他公司有设备，也拿不到关键部件。

其次，光纤连接（FAU）并不是标准化零件，而是为特定设计定制的。没有原厂设计支持，很难复制。

最后，也是最现实的一点：认证周期。进入大型数据中心供应链，通常需要半年到一年以上的验证时间。一旦某个厂商通过认证并稳定供货，客户不会轻易更换。

换句话说，这不是一个“买机器就能复制”的行业。

真正需要关注的风险

如果说这个故事有风险，那核心并不在竞争对手，而在技术路径本身。

例如，未来如果行业从“可插拔光模块”转向更先进的CPO（共封装光学），整个产业链结构可能发生变化。如果POET不能跟上这一步，Natgate的优势也会受到影响。

另外一个关键变量是客户验证。如果POET的产品无法进入主流AI数据中心供应链，那么所有的产能和技术都很难转化为稳定利润。

一个更本质的理解

如果把整个故事压缩成一句话，它其实不是关于订单，也不是关于短期盈利。

它真正的核心是：
Natgate正在尝试从一个“可替代的组装厂”，转型为一个“嵌入AI基础设施中的关键制造环节”。

这个转变如果成立，带来的不仅是利润增长，更是公司在资本市场中的“身份改变”。

而市场，往往对这种变化反应很慢。

//

Natgate × POET: An AI Breakout Opportunity the Market Hasn’t Understood Yet

//

Recently, I came across a post on the KLSEScreener comment section sharing news about the collaboration between NationGate Holdings (Natgate) and POET Technologies. It immediately caught my attention.

When I was previously researching Natgate, AI had already pointed out to me that its future breakout might not come from AI server assembly, but rather from optical modules. With that perspective in mind, seeing Natgate partner with POET—a disruptive innovator in the optical module space—made it clear that there was likely something significant here.

So I ended up spending about three hours (which I honestly didn’t plan to) digging into this collaboration. My focus was on two things: first, how POET’s disruptive passive alignment technology in optical modules allows an EMS company like Natgate to take on manufacturing processes that were traditionally only handled by OSAT players; and second, what kind of impact this could have on Natgate’s future valuation.

What’s interesting is that this collaboration hasn’t really been picked up by major media outlets yet, and it seems that most investment banks and analysts haven’t fully recognized its significance either. At this point, aside from the company’s management and employees, I suspect I might be one of the very few people online who has noticed just how impactful this technological breakthrough in optical modules could be for Natgate.

Yesterday, in line with my usual style, I posted a long, unedited transcript of my AI discussions. It unintentionally created a kind of filtering mechanism—only those who truly understand the topic would bother to go through it. In today’s world, where AI is so accessible, getting it to generate a polished article is almost effortless. So today, I decided to repost this in a more structured format.

This time, it includes a bilingual AI-generated article, the full AI conversation transcript (posted separately due to words limit), and the introduction you are reading now, written by me. This is likely the format I will use going forward.

You really only need to read the AI article below. If you’re genuinely interested, then you can go deeper and explore the full conversation transcript.

//

When the market talks about AI, almost all the attention is focused on GPUs—especially NVIDIA. It often feels as if simply having more compute power will automatically make AI better.

But reality is not that simple.

The real bottleneck in AI systems is often not “compute,” but “communication.” When thousands of servers are training models together, the key question becomes: how do these machines exchange data fast enough?

That is precisely where optical modules come in.

The Overlooked Piece: What Exactly Is an Optical Module?
A simple way to think about it is this: GPUs are the brain, and optical modules are the nervous system.

Inside a server, everything runs on electrical signals. But over long distances and at ultra-high speeds, electrical signals run into serious limitations—heat, signal loss, and inefficiency. As a result, in large-scale AI data centers, data must be converted into light, transmitted through optical fiber, and then converted back into electrical signals on the receiving end.

This “electrical ↔ optical” conversion is exactly what optical modules do.

In other words, no matter how powerful your GPUs are, without optical modules, you cannot form a true compute cluster. They are not optional components—they are fundamental to the entire AI infrastructure.

Why Is This Sector Suddenly So Important?

Over the past few years, AI infrastructure has been going through a clear upgrade cycle: from 400G to 800G, and now moving toward 1.6T.

These numbers represent data transmission speed. As AI models scale exponentially, the amount of data exchanged between servers grows at the same pace. If transmission cannot keep up, even the most powerful GPUs will end up “waiting at the door.”

This leads to a very direct conclusion: demand for optical modules grows alongside GPUs—and it is a non-negotiable demand.

It is in this context that companies like POET Technologies start to stand out.

What Did POET Actually Change?

Optical modules are not new. The real challenge has always been packaging.

In traditional manufacturing, the most complex step is optical alignment. Engineers must power up the chip, make it emit light, and then carefully adjust its position until the light is perfectly aligned with the fiber. This process is slow, expensive, and heavily dependent on advanced OSAT (Outsourced Semiconductor Assembly and Test) capabilities.
POET’s breakthrough lies in solving this problem in advance.
Its core technology, the Optical Interposer, can be thought of as a pre-designed optical platform. The optical pathways are already embedded into the structure. When components are placed onto it, they naturally align.

This leads to a profound shift:

A process that used to belong to “packaging” is now simplified into high-precision assembly.

In other words, what once required expensive equipment and time-consuming calibration now looks more like a precise “placement” process.

Why Does This Change Natgate’s Role?

This brings us to the other key player in the story: NationGate Holdings.

Traditionally, EMS companies like Natgate focus on electronics assembly—servers, PCBs, and so on. These businesses are characterized by large scale, low margins, and high substitutability.

But under POET’s system, Natgate’s role begins to evolve.
In the production chain, Globetronics still handles wafer-level processing—cutting and preparing the raw chips into usable core components. This remains firmly within the traditional semiconductor packaging domain.

However, once these core components reach Natgate’s production line, the nature of the work changes.

Natgate is no longer simply assembling parts. It performs two critical functions: high-precision placement (similar to flip-chip processes) and fiber attachment (FAU coupling). The latter, in particular, is one of the most sensitive steps in the entire optical module process, with a direct impact on yield.

In the traditional manufacturing paradigm, these tasks would be handled by OSAT players or top-tier optical module companies. But because of POET’s passive alignment design, Natgate can now execute them using its expertise in precision automation.

What does this mean?

It means Natgate is gradually transitioning from a “simple assembler” into a manufacturer with genuine optoelectronic integration capabilities.

Why Do the Profits Look Small—but Matter So Much?
At first glance, this story can be easily underestimated.
Suppose Natgate’s AI server business generates 100 units of profit, while future optical module contributions add another 20 to 30 units. On paper, that looks like just a 20–30% increase.
But this comparison misses a crucial point: the quality of earnings.

Server assembly is fundamentally a low-margin, high-volume business. Its profits are highly sensitive to exchange rates, component costs, and customer demand fluctuations.
In contrast, optical module-related work is technology-driven. Margins are higher, and once a supplier is qualified, it is difficult to replace.

More importantly, the capital market views these businesses differently.

A traditional EMS company might trade at 10–15 times earnings. But once recognized as part of the optical communications or semiconductor ecosystem, that valuation could rise to 25–40 times.

This creates a critical shift:
Profits may only increase modestly, but the valuation framework changes entirely.
That is why, in some cases, a 20% increase in profit can lead to a doubling—or more—of the stock price.

Can This Be Easily Replicated?

A natural concern is whether this advantage can be quickly copied. If POET’s technology simplifies packaging, wouldn’t other EMS players be able to enter the space?
In the short term, it is much harder than it appears.

First, POET does not openly license its technology. Instead, it uses a consignment model, placing critical equipment and components within partner facilities. Without access to these, competitors cannot replicate the process.

Second, fiber attachment units (FAUs) are not standardized parts. They are custom-designed for specific architectures, making them difficult to reproduce without direct design support.

Finally, there is the issue of certification. Entering the supply chain of large data centers typically requires six to twelve months of validation. Once a supplier is qualified and stable, customers are reluctant to switch.

In other words, this is not an industry where you can simply “buy machines and compete.”

The Real Risks to Watch

If there are risks in this story, they are not primarily about competition, but about technological direction.

For example, if the industry shifts from pluggable optical modules to more advanced solutions like Co-Packaged Optics (CPO), the structure of the supply chain could change significantly. If POET fails to keep up, Natgate’s advantage could diminish.

Another key variable is customer validation. If POET’s products fail to enter mainstream AI data center supply chains, then production capacity and technical capability may not translate into sustainable profits.

A More Fundamental Perspective

If we reduce this entire story to a single idea, it is not really about orders or short-term earnings.

At its core, it is about this:
Natgate is attempting to transform from a “replaceable assembly manufacturer” into a “critical manufacturing node embedded within AI infrastructure.”

If this transformation succeeds, it will not only drive profit growth, but also fundamentally change how the company is valued by the market.

And the market is often slow to recognize such shifts.

Trump bans semiconductor sales to China. Follow Uncle for the latest insights and understanding on the market it's free forever Uncle promise as Uncle earn from the market Warren Buffett would never charge you because for a true superinvestor the returns from the market exceeds all... unless the so called "stock guru or sifu" has no skills and relies on fees and subscriptions from water fish...