Reference: Why scale-up AI networks demand scalable optical test

Share Elementary, Dear Watson

Never mind about the graduate level assumption - if you just read closely, the integrated photonics industry really needs people who can add value by doing the “integration”, solving the aforementioned integration challenges (well, opportunities), and truly codesigning on the system level. None of those aforementioned problems were necessarily restricted to a single domain. It should be clear that we need more senior engineers like or young curious, open-minded, T-shaped engineers with firm grasp of the fundamentals and willingness to go beyond comfort zone to solve system level problems that span across more than one domain. For one, if you can train, attract, and retain such talent within your company in the integrated photonics value chain, that can be your company’s incredible moat.

References: Learnings to Date from PhD Workforce Interviews (AIM Photonics), Photonics Forces A Chiplet Rethink (Semiengineering)

Share Elementary, Dear Watson

References: EE Times Interview of Mark Wade (Ayar Labs) and Congress testimony given by Pete Shadboldt (Psiquantum)

Share Elementary, Dear Watson

Share Elementary, Dear Watson

References: See “Silicon photonics foundries“ slides (a bit dated but has all of the useful fundamentals information) on this page from Joaquin Matres (GDSFactory), Form 20F for FY2025 from Tower Semiconductors

Share Elementary, Dear Watson

Insight: The dominant cost of a commercial optoelectronics system is assembly, testing, and packaging (see Fig. 1. of reference). At the product (system) level, the cost of the actual chips, photonic integrated circuits (PICs) and electrical integrated circuits (EIC), is but a small fraction of the overall bill of materials (BOM).

Reference: Advances in waveguide to waveguide couplers for 3D integrated photonic packaging

Share Elementary, Dear Watson

(EN)

Fig. 1. Disclaimers.

Fig. 2. Opening lines.

(ZH)

Fig. 3. Disclaimers.

Fig. 4. Opening lines.

Travel Plans in Coming Months

Tentatively, here are my travel plans in the coming months. I will be open to chat shop then over coffee with like minded subscribers and readers, time permitting.

• Boston - Nov. 14

• NYC - Nov. 14-16

• Austin - Nov. 16-19 (TomorrowX Summit)

• Chicago - Nov. 19

• Bay Area - TBD

• Taiwan - tentatively Jan. 2027 (trip around the island like in the Taiwanese movies while checking out the copackaged optics ecosystem there in person; below is the rare, free piece I wrote on CPO based on a talk from Winway from a while ago)

If any of you like to help me pay for my trips to Austin and Taiwan and thus receive paid posts with key takeaways afterwards, you are free to join the quickly growing number of paid subscribers. As for the planned Taiwan trip, I will appreciate it if any local reader can help make introductions or give tips on where and whom to visit as I check out the local ecosystem for copackaged optics there.

Share Elementary, Dear Watson

The primary reason you are reading this is because you are looking for an alternative source of information in English from a knowledgeable author who understands the fundamentals on the copackaged optics ecosystem. It is true that Taiwan has a tremendous head start in terms of the industrial capacity in terms of semiconductors (foundry, packaging, etc.). Ex ante, that does not necessarily translate to success in copackaged optics (CPO) or any of the variants in between CPO and traditional pluggables. If you don’t know the difference among all the different options, I recommend watching this video (sorry, it is in Chinese) from a Terahop (Innolight subsidiary) executive 于讓塵 and an overview piece on CPO I wrote with Omer Cheema a while back. Now, we will delve into the Taiwanese CPO ecosystem at a high level.

Share Elementary, Dear Watson

Read more

Share Elementary, Dear Watson

Image of Holmes and Sherlock from the popular BBC TV Series Sherlock (Source: Britannica).

The value proposition for this paid newsletter is simple - there is unmet demand for expository content underlying the fundamentals of highly technical topics such as copackaged optics, with respect to the underlying physics, engineering, economics, and cultural context. UNDERSTANDING THE FUNDAMENTALS GO A LONG WAY, AS A LARGAN EXECUTIVE FOUND OUT THE DIFFICULT WAY - see aforementioned first (free) piece of the newsletter! It should be understood that everything written here is based on public information and in principle content that the reader can get elsewhere in a much costlier way if you don’t have my unique, valuable background (see my explanation of Hamilton Helmer’s 7 Powers as applied to myself on the individual level for this newsletter).

What this newsletter does differently:

• Primary source analysis - Anything that is publicly available information is fair game - job posts, conferences, publications, public filings, etc. The author can read primary sources and give original, independent assessments directly in English, traditional Chinese, and standard German, and ask friends to translate for rarer ones.

• Target audience - The target audience is a highly sophisticated audience who have the need for high quality technical content in the densest form factor. This may include venture capitalists, AI engineers, hedge fund analysts, technology executives, etc.

• Target content - To be frank, a firm understanding of the fundamentals is the truly scarce resource in this day and age. What do I mean by that? For example, if you are working or investing in copackaged optics, not understanding the fundamentals (physics, engineering, economics, and culture) will likely cost you a fortune down the line, as I mentioned before.

• Value proposition - You should think of the newsletter as a convenient golden reference with highly technical content that saves you time to insights that are relevant for your business.

• Subscription price - I view the lowballed subscription price of 499.99 dollars/yr as a commitment device for me and my readers to take the content seriously. The current price is drastically undervalued, given how it is many orders of magnitude lower than the opportunity cost per hour for my target audience. I primarily view it as a way for me to reach like minded folks in all walks of life I otherwise would not typically talk to (拋磚引玉). If you are not getting at least 499.99 dollars/yr worth of value add after subscribing to the paid plan, I encourage you to cancel it and give me feedback - you will be refunded the part you paid me then (there may be some portion charged by Substack and Stripe that I have no control over).

Why me and this newsletter in particular? Why can’t you get this content elsewhere? Well, I will give you my individual variant of Hamilton Helmer’s 7 Powers:

1. Scale Economics

   1. Can I produce value while I sleep?

   2. This newsletter will become a golden reference for busy paid readers who don’t have the time to acquire the knowledge otherwise from someone who can speak your language, metaphorically and literally.

2. Network Economics

   1. Does my presence attract the right people?

   2. Given my unique background (cultural/educational/work), the newsletter will be informed by the conversations I had about the fundamentals with a broad cross section of contacts (engineers, executives, VCs, HF analysts, family office analysts, etc.) throughout the technology value chains spread across Europe, Asia, and North America.

3. Counterposition

   1. What can I do that others are too afraid to try?

   2. All of the valuable skills within my repertoire are not valued individually by the market, ex ante. For example, which American would study mathematics and physics at a liberal arts college back in his Midwestern hometown in the 2010s, given the difficulty and low return on investment? Similarly, who would teach himself proper standard German and traditional Chinese then, utterly useless outside of Germany (optics powerhouse) and Taiwan (semiconductors powerhouse), respectively? As for optics, as a millennial, I grew up in the aftermath of the dotcom bust when optics became largely a commodity and thus ignored industry - well, the Germans even have a noun for describing such an industry (“Scheissindustrie”) . Where else can you find such a rare American with the unique repertoire of skills useful for dissecting highly technical topics such as copackaged optics in terms you can understand?

4. Switching Cost

   1. Am I essential to the core operations of my environment?

   2. Once you subscribe to the newsletter, you will struggle to find an alternative with the same coverage of the fundamentals.

5. Branding

   1. Does my name reduce the "risk" for decision-makers?

   2. Given my unique, valuable background, you should have no trouble charging this subscription as an expense at your firm as the cost of acquiring highly valuable fundamentals knowledge underlying technologies that materially affect your firm’s bottom line whether as a supplier, customer, or investor.

6. Cornered Resource

   1. What do I know or possess that is genuinely rare?

   2. If you do a back of envelope calculation on how many people there are with my profile in the entire country (America), you would likely reach O(1) person. I am:

      1. A rare American with both theoretical training and practical experience actually relevant for covering highly technical topics such as copackaged optics - I studied physics and mathematics as an undergraduate at UChicago and worked on software used to design optical chips (say, transceivers) at Cadence, among other things,

      2. Neither a Taiwanese or mainland Chinese, but still deeply familiar with the technical and financial aspects and deeply familiar with the semiconductors and optics ecosystem in Asia - specifically, I am of Cantonese heritage, so like Hong Kongers, I am literate in both traditional/simplified Chinese and speak both Cantonese and Mandarin,

      3. A Germanophile who appreciates the culture, but more importantly, also has the knowhow to understand the commercial industrial landscape of Germany and Europe in general - Germany is the industrial powerhouse that the rest of the world will rely on for its world leading hidden champions in optics and other areas for the foreseeable future

7. Process Power

   1. Is my daily workflow faster and more effective than the industry standard?

   2. Of course, because this newsletter is the output from directly reading primary sources and truly understanding the fundamentals. By just glancing the cross section of a chip or package, I can already infer a lot about the underlying physics, engineering, and economics with high conviction from first principles. Moreover, I directly speak the languages of my audience, metaphorically and literally. Whether you are a hedge fund analyst/PM in NYC/Chicago/SG/etc., semicap engineer at Aachen, packaging engineer at Kaohsiung, laser engineer at Sunnyvale, venture capitalist at Menlo Park, etc., I can communicate the fundamentals of these highly technical topics in terms that you can understand and in your native tongue (if it is English, Chinese, or German).

Pricing — effective June 10 (11:59 PM, Chicago Time):

Monthly: $49.99/month

Annual: $499.99/year

Annual (Founding Member): $599.99/year

If you’ve been considering subscribing, now is the time. Lock in the current rate before I consider raising it as I create content with more value add over time.

Content offerings, format, and frequency may evolve over time as Elementary, Dear Watson grows.

— Jack Huang

Share Elementary, Dear Watson

This post will be the first and last free post of this new newsletter Elementary, Dear Watson, which you can thank my dear friend for. As he liked to remind me in our native tongue (一份材料一份話，沒有材料不說話), I only ever speak to matters within my circle of competence and for which I have both firm practical and theoretical understanding. Because of a certain principle, I can’t and won’t refute the words of many readers (whether on the sell side, buy side, or within a copackaged optics company) one by one, at least not for free. Copackaged optics is a highly technical area that requires rare, valuable expertise in understanding the underlying physics, engineering, and economics. Not understanding the fundamentals when investing in copackaged optics will cost you as an executive of Largan (大立光) found out belatedly - oops, it turned out that CPO is a high precision business that requires significant investment and specialized knowhow. Just knowing the fundamentals can be a key competitive advantage!

Article title page and relevant passage from this 商周 (local Taiwanese business magazine) piece.

Share Elementary, Dear Watson

Unique Background

I recognize that I am one of a few Americans who have the technical, financial, and cultural chops to give you a down-to-earth take on different aspects of copackaged optics (CPO), unlike most of the folks opining on the subject. As a Midwesterner, I don’t think pedigree (education/work) matter much in terms of understanding the heart of objective matters, but I did study physics and mathematics in a small liberal art college of sorts in my small Midwestern hometown and work on electronic-photonic design automation software at an EDA company, if that counts for anything. Specifically, I am a Germanophile (many high precision optics and glass vendors are Mittelstand/Nebenwerte companies based in DACH countries) and of Cantonese heritage, so neither Taiwanese or mainland Chinese. My beloved Hong Kong is a non-player in semiconductors except for a small presence of ASMPT (another player worth paying attention to for its small but growing exposure to CPO) in New Territories (香港新界青衣). I will be in the Bay Area this weekend - specifically, South Bay (Palo Alto, San Jose, Santa Clara, etc.) on this Saturday 6/6 and San Francisco on this Sunday 6/7 (feel free to DM me for coffee or something if like minded readers are in the area). As an aside, I recommend reading , , , , , on Substack for more in depth pieces in this area.

Winway Copackaged Optics Talk

Winway is a testing equipment company based in Taiwan and it held a talk a while back to educate investors on CPO. You can watch the talk here. While the talk is delivered in Mandarin Chinese, the slides are all in English. That said, some of the nuances might be lost if you are not familiar with Mandarin Chinese with Taiwan specific lingo (例如，矽vs.硅), copackaged optics, or testing. So I will explain them with more context here after each slide if the slide itself is not obvious enough.

It is certainly an exciting time to be working or investing in the optics industry, given that the telecom boom and bust occurred a generation ago, during my childhood when I looked up to a certain Hong Konger called Charles Kao. Optics R and D takes a long time so it is almost always preferable to do mergers and acquisitions to acquire both the IP and experienced people to accelerate time to market.

The big picture is that photonics is small but growing very fast relative to electronics, given the system level gains photonics unlocks.

Yes, there are many engineering and economics challenges, some of which Collins Sun mentioned here. Challenges mean opportunities (危機-有危才有機) for those who can genuinely add unique value in this emerging supply chain.

Those are the two relevant paradigms with the tradeoffs elucidated clearly. He skipped the half-steps (such as NPO, LPO, etc.) in the gradient from traditional pluggables to copackaged optics.

This should be obvious for anyone who studied electromagnetism at high school (Gymnasium/高中) or university level.

The general gist is that you can get way lower loss, higher frequency, longer reach, and EMI immunity with optics in principle.

That aside, although they are both different manifestations of the same electromagnetic field, the actual engineering and economics differ significantly.

In principle, CPO brings system-level benefits for the next few generations of technologies relevant for the AI semis ecosystem. As I always reminded my aforementioned dear friend, what do you see in the idealized cross section of this CPO system?

Notice where the laser source is? Why? This is an exercise left for the reader.

The general trend is for ever smaller form factor of the overall system which enable large improvements in bandwidth, power use, latency, and so on. Of course, the ones past today are just projections.

This is a very useful, but likely not comprehensive map of the silicon photonics supply chain. It looked a bit familiar from the ones from and . In any case, I am working on my own map of the silicon photonics supply chain - stay tuned.

Here is a highly stylized schematic of what goes into a photonic integrated circuit - you can think of it as a chip for photons. DO NOT FORGET - there is always an electrical integrated circuit (EIC) that goes along with the photonic integrated circuit (PIC)! This sounds obvious, but lost on many ordinary folks. Of course, the overall insertion loss have to do with what materials platform you make the PIC out of and what fiber-to-chip couplers you use (most common are grating and edge couplers with pros and cons). For passive PICs, you want to use silicon nitride (specifically, deposited via low pressure chemical vapour deposition) waveguides for low loss, as may have mentioned before somewhere.

Yeah, that neatly summarises the tradeoffs among the different kinds of laser sources that are of course downstream of the underlying device physics and engineering choices which affect the overall economics. Again, just as before, I will like to remind my dear friend and other readers - just look at the cross section (especially of the relevant laser cavity).

Here is a summary of the different kinds of modulators. Real products made in high volume production amount to some point chosen on a tradeoff curve dictated by the underlying economics and engineering, which sit downstream of the device physics. wrote about this at some point, but you can look back at past Hot Chips Conference slides for similar information provided by the respective vendors. The more useful exercise, in my humble opinion, is to only look at revealed preferences - again, just look at the cross section of commercial PICs already produced in high volume and understand the fundamental reason why a modulator is used.

Yeah, this is fairly straightforward passive integrated photonics.

The idea here is that you can send more data at optical wavelengths by encoding them on light of different “colous” (wavelengths). In reality, each signal of a certain length has some finite spread, so the goal here is to minimise crosstalk between adjacent optical channels.

This slide tells you the gist of what you need to know. Grating couplers are the most scalable couplers if you can take the hit on insertion loss and bandwidth tolerance elsewhere in the system. They enable wafer level testing and packaging, which improve the overall manufacturability of the end CPO system. Of course, when overall insertion loss and polarisation insensitivity matter, edge couplers are preferred.

Yeah, there are different optical engines and optical engine connectors. Again, the engineering and economics differ significantly among the different integration schemes between EIC and PIC. The scalability of each EIC-PIC integration scheme is left as an exercise for the reader - again, what do you see in the cross section?

That is right - we are still in the early days with lots of spec diversity and customized solutions. Useful standards will only get determined by the vendor who ships first with the highest volume of quality products.

The reader should note that the silicon photonics related part of each PIC foundry/IDM is only a small (but vastly growing) part of each business. If any of you spent any time working or investing in a foundry or read Morris Chang’s memoir, you would know how important it is to be going up the learning curve as soon as possible. That said, I don’t know why the speaker left out Tower on this slide - go read who already wrote about it.

Yes, testing is a bottleneck in CPO production. I am just stating that this is a bottleneck solely on technical and economic considerations - nothing more and nothing less. Just stare at the cross section of a CPO system - how do you even know any of it works at all if you don’t test it at each step? Here is an excellent PIC Magazine piece from Andrew Yick (from Marvell) explaining how this works:

Here is a good summary slide on visualising the CPO test flow in addition to the one I have shown you earlier from Andrew Yick. Here is an exercise for the reader - what is the value of the inputs and the output of each step when constructing the CPO system? Again, just stare at the cross section…

From a practical system level perspective, you want to use single mode fiber (SMF), the most famous of which is Corning SMF-28 (but of course, there are other variants).

Yes, as someone used to rigor, I am delighted to see error bars because the engineers actually thought about manufacturability (design for manufacturability), testability (design for test), and scalability in the presence of variations that can be expected in high volume production settings. These quoted values are plausible values - for context, I have made V-grooves in silicon and used fiber arrays personally.

Again, reminder, there is little room in this field for amateurs who don’t appreciate the tolerances required to realize CPO systems.

Active alignment is largely a necessary evil, but it contributes to cost and throughput. Ideally, you don’t want to do that for a high volume commercial product if you can.

Here is the tradeoff between the two solutions.

Do not be fooled by how easy it looks as if it is like Legos - just no, these FAUs need to be very precisely aligned with the optical couplers on the PIC. Moreover, would they stay in place (post alignment) and not move that much after you apply (dispense and cure) some epoxy (basically glue)?

Yes, next time someone tells me about CPO, I will show them this slide. Again, my perspective is simple - more problems equal more opportunities for those who can add unique value for the end customers of CPO systems!

Just think about the opportunities here (imagine the testing intensity at each step) at each of the testing steps and levels - die, wafer, package, and module. Of course, this flow already exists for electronics, but the optics and optoelectronics testing market is still a blue ocean - go read the relevant posts written by my friends , , and .

Ideally, you want to do the test on the wafer level for high throughput in both the electrical and optical domains. As before, stare at the cross section!

Here is a slide on the die level test.

You still have to test more after you package the PIC and EIC.

Here is what package level test can look like.

Here is what module level test can look like. Again, I need to emphasize that it is not easy as it looks given the tolerances involved in CPO systems. I won’t comment on specific commercial products - do your own due diligence.

Here is a cross section of the COUPE platform from TSMC. As mentioned before, I like seeing grating couplers - why? Just stare at the cross section and this is left as an exercise for the reader.

Here is a projected CPO roadmap that shows what kinds of advances are needed to enable future technologies in the future.

That is right - it is tough to engineer the electro-optical interface. The future will require very close co-design by multidisciplinary folks with non-traditional backgrounds - say, T-shaped engineers like , , and .

Don’t dismiss copper yet - optics and copper will coexist for a long time in different niches (reaches, etc.) and copackaged copper has a role to play. Again, stare at the cross section and see what pops out at you.

CPO and CPC have different vendors in their ecosystems.

Okay, he is now talking about what Winway has to offer.

I am always excited to see future technical challenges - that means more value will be accrued to companies who can actually execute and solve these valuable technical challenges for their customers. I won’t comment more other than saying solving these problems will require rare folks with firm understanding of the underlying fundamentals.

There are mechanical considerations to ever larger interposers (from TSMC) with demand driven by AI chips.

There are mechanical considerations to ever larger interposers (from Intel) with demand driven by AI chips.

Here are some of the tradeoffs among near term technologies in terms of electrical performance. Cleaner “eyes” are preferred in the context of optical communications.

Why can’t the industry just keep using copper? A few reasons listed here…

Pay attention to the SNR penalty and DSP complexity.

Just imagine the heat density for these next generation GPUs - what a exciting time to live if you are a thermal engineer!

Here are some of the considerations why some may consider glass as opposed to organic substrates (just go read any serious Taiwanese local media pieces on advanced packaging technologies from TSMC).

Amazing opportunity for those who can execute and deliver those advanced package test solutions for such packages (complexity grow as a function of size, pin count, speed, and heat density)!

These should look familiar to those already familiar with semiconductors testing - if not, go read my friend ’s pieces.

He is talking about the product now, so I have no comment. The only thing that matters is whether their customers qualify their products for use on their products.

Perhaps I should go read the dry patents (some of which are probably in traditional Chinese) - had I understood the importance of traditional Chinese at the time in undergrad, I should have spent some time in Taiwan or my beloved HK (my alma mater UChicago has a center at Mt Davis 摩星嶺) then. On that note, I am currently reading this patent from Himax.

The talk was so long, but fine, it was a great two hour long talk.

This is a projected timeline - what an exciting time to live if you work or invest in photonics!

Again, just stay at the cross section!

Share Elementary, Dear Watson

Conclusions

In a sea of AI slop about such an important technology, I hope that this short piece can clarify a few things here and there. I can assert that none of this is written by AI and solely written by myself. Just as my younger self, I have always preferred to read only primary sources in the original language (English, Hochdeutsch, and 繁中) and make my own independent judgments about technologies actually from first principles in terms of the underlying physics, engineering, and economics. As a reminder, I will be in the Bay Area this weekend - specifically, South Bay (Palo Alto, San Jose, Santa Clara, etc.) on this Saturday 6/6 and San Francisco on this Sunday 6/7 (feel free to DM me for coffee or something if like minded readers are in the area).