Going to be a busy few days.

All of this content will be available for paid subs.

Subscribe now

Here is what I am working through and planning to publish over the next week or so:

1. China oversupply risk and the impact on companies like Lumentum, Coherent, Applied Optoelectronics, and others.

2. Deep dive into testing/packaging for optics + company map across Keysight, Aehr, Nokia, Intel, FormFactor, and more.

3. Thesis on my newest Physical AI position, where I have allocated around 12% of my target capital for the portfolio I am building in real time.

4. Layer 4 of my Physical AI thematic research: Edge Compute + company map.

5. Valuation models for Semtech, Credo, Astera Labs, MaxLinear, and MACOM. The goal is to identify which names offer the best risk/reward at current levels relative to their growth outlooks.

6. Thoughts on Ciena’s price action and a full stress test of the thesis.

7. Updated Optics Mock ETF with revised weightings.

8. Analysis of the upstream tool providers Aixtron, Veeco, and Riber.

This is my full-time job now.

I get to spend many hours each day digging through filings, transcripts, company presentations, industry data, and market structure.

It is a lot of work, but it is also exactly what I want to be doing.

Can’t wait to share all of this.

Thank you for everyone that is here!

The optical engine is the largest value pool. The switch ASIC is obviously critical.

But this BoM screenshot shows something the market may still be underappreciating.

In the GS estimate, single-mode fiber plus FAUs are roughly $15.9k. Add MPO connectors and cables, and the passive optical connectivity bucket moves above $21k.

That is the $GLW read-through. I am not saying that this whole bucket is Corning revenue in this exact architecture because the supplier split is undisclosed publicly to my knowledge. The illustration I am makign is that Corning has disclosed products, partnerships, and manufacturing expansion across the passive optical layer these systems require.

We can talk about lasers all we want, but CPO still needs fiber, fiber arrays, connectors, cables, trays, routing, and physical optical infrastructure to actually work.

That is where Corning sits, and that is one reason I have spent so much time on this name.

The rest of this post is for paid subscribers. Below the paywall I walk through why CPO changes the physical challenge in ways that go well beyond fiber demand, why the recent customer agreement sequence matters, and why the passive optical layer may be more valuable than the market currently prices in.

This is general market commentary, not individualized financial advice, tax advice, or legal advice, and should not be treated as a recommendation to buy, sell, hold, or short any security.

Read more

$NVDA already moved first. It put $2 billion into $LITE and $2 billion into $COHR, with purchase commitments and future capacity access rights tied to advanced lasers and optical networking products.

Then TrendForce just reported that AMD has accelerated procurement of external laser solutions and is reportedly negotiating major purchase agreements for high-power CW laser chips, with the stated goal of securing future capacity away from supply chains tied to NVIDIA’s ecosystem.

So if NVIDIA has already tied itself strategically to Lumentum and Coherent, and AMD wants high-power CW laser supply outside those capacity pools, the next question becomes obvious: who else can supply the light?

That is where $AAOI gets interesting.

Rosenblatt channel checks also recently suggested AAOI was believed to be in talks with AMD about becoming a customer.

That is below company-confirmed disclosure and I am treating it exactly that way. But it lines up with the idea that AOI is trying to position itself to capture as much of future architectures as possible. And it adds a supporting theoretical angle to the Trendforce statement.

At OFC 2026, they showed a 25dBm ultra-high-power ELSFP built around a 400mW CW laser, plus a live 6.4T on-board optics demo powered by that ELSFP, alongside in-house high-power laser manufacturing, Texas capacity expansion, and a target of roughly 400,000 ELSFP units per month of capacity by the end of 2027.

And importantly, management has already confirmed that it is working with large unnamed customers around ELSFP. On the Q1 call, when asked about ELSFP customer engagements, CFO Stefan Murry said, “We have a couple of large customers that we’re working with. We haven’t said who they are.”

But, the AMD deal has not been confirmed.

The major question we are currently navigating is whether they could ramp 800G and 1.6T.

But another question now is whether AAOI can become one of the alternative high-power laser and ELSFP suppliers for the AI accelerator ecosystem. That is a very valuable question to be able to answer.

And if AMD is involved in a meaningful way, that will be a major catalyst for the stock if/when announced.

The rest of this post is for paid subscribers. Below the paywall I walk through why AMD needs a separate optical supply path, what AAOI is actually building in Texas, what a supply relationship could look like at each stage, and the exact milestones that would make this thesis real.

None of this is financial advice or a recommendation. This is general market commentary. This company is high risk and extremely volatile. Readers should perform their own due diligence.

Read more

• Optical and IP are growing faster than the rest of the company.

• Infinera added optical scale at exactly the right time

• AI and Cloud orders are finally showing up in the numbers

• The San Jose indium phosphide ramp gives Nokia more control over its optical supply chain

• The 2027 coherent roadmap gives us a product cycle we can underwrite.

That’s the clean re-rate case, and it’s still the easiest way to explain why the stock has worked and why it can keep working.

Read more on all of that here:

But there’s a deeper layer here that I haven’t discussed, and that’s what this report is about.

Nokia may have one of the few research engines capable of shaping the next decade of AI networking.

The optical and IP thesis is about what drives the stock today.

But Bell Labs is about why Nokia will still be relevant as AI infrastructure spreads beyond data centers and into factories, robots, drones, defense systems, private networks, and eventually 6G.

Over time AI becomes a physical-world infrastructure story. If you haven’t been following, I am a strong believer in the Physical AI supercycle that is coming and am researching it extensively.

When AI moves into the physical world, the network layer becomes a lot more important.

That’s where Nokia gets really interesting. And this angle isn’t discussed often. But that changes today!

The Network

Let’s start by thinking about what an AI system actually needs. A giant AI data center is basically one large machine made up of GPUs, memory, storage, power, cooling, software, and networking. The network links the machine together. If the network is slow, congested, unstable, or inefficient, expensive compute sits idle or underperforms.

That’s why optical and IP networking are becoming more valuable. AI clusters need to move enormous amounts of data inside data centers, between data centers, across metro areas, and eventually out to edge locations. Nokia’s current strength in optical transport, IP routing, coherent pluggables, and data-center networking gives it a real seat in that buildout.

The Q1 2026 numbers show the shift is already happening in a very meaningful way. Network Infrastructure grew 6% on a constant currency basis. Optical Networks grew 20%. AI and Cloud customer sales grew 49%. AI and Cloud customers reached 8% of group sales, and Nokia booked €1 billion of AI and Cloud orders in the quarter. Management raised its view of the AI and Cloud addressable market to a 27% CAGR from 2025 to 2028, up from a prior estimate of 16%.

But the bigger point is that AI infrastructure will keep spreading. Today the network helps connect GPU clusters. Tomorrow the same idea applies to factories, robots, vehicles, defense systems, and physical environments of all kinds. AI systems in the real world need fast, reliable, secure connectivity. That’s why Nokia’s network assets could become more strategic over time, and that’s where Bell Labs connects the near-term and long-term thesis.

Bell Labs

This is a blast from the past! I had to spend some time really digging into the history here.

The transistor, information theory, UNIX, communications satellites, and more. I won’t get too in the weeds here.

Nokia acquired Bell Labs through its 2016 takeover of Alcatel-Lucent, a roughly €15.6 billion all-share transaction announced in 2015. That deal brought Nokia optical scale, IP/networking assets, and one of the most important industrial research labs in technology history.

The main question is how does Bell Labs help Nokia make money, win customers, and create value today?

The answer is a flywheel that a normal equipment vendor doesn’t have.

Research creates inventions. Some inventions feed into standards. Standards create patents. Patents support licensing revenue. Licensing revenue helps fund future research. At the same time, that research feeds product roadmaps in optical networking, wireless, automation, security, and network software.

Nokia has invested more than €150 billion in R&D since 2000, owns around 26,000 patent families, has more than 8,000 patent families declared essential to 5G, and has more than 250 licensees across its patent licensing programs. Nokia also says licensing revenue is reinvested into future inventions including 6G. Nokia sells hardware and software, but it also owns a standards and IP engine that Bell Labs helps power.

Nokia reorganized around this direction in September 2025, creating a Technology and AI Organization that integrates Nokia Bell Labs, Technology and AI Leadership, and Group Security under Pallavi Mahajan. Bell Labs now sits inside the current strategic architecture of the company.

That flywheel is hard to value in a quarterly model, but it’s real, and it explains why Nokia keeps showing up across so many future network categories.

The rest of this post is for paid subscribers. Below the paywall I get into the specific product areas where Bell Labs is doing work that matters for the Physical AI thesis: network-as-a-sensor, AI-RAN, private wireless, defense, quantum-safe networking, and how the Pennsylvania expansion fits the supply chain story. I also cover what I’m watching and why I remain bullish on the full combination of near-term re-rate and long-term relevance.

Read more

I have written multiple articles on it. This linked one here touches upon several aspects of it.

But this article I am writing today takes a step back and looks at what the demand is that is driving the need for this layer.

As AI shifts from training ever-larger models toward deploying them at scale, and as that deployment increasingly takes the form of autonomous agents rather than single question-and-answer exchanges, the demands placed on networks are changing in both volume and character.

Two forces sit at the center of that change:

1 - The rise of agentic AI

2- The emergence of the distributed, multi-campus data center.

Understanding how they reshape traffic, and which parts of the network capture the result, is the key to understanding one of the most durable parts of the AI infrastructure investment case.

This article lays out the mechanism, the evidence, and the map of who is positioned where.

It leans on reported company results and established engineering realities.

The picture that emerges is one of genuine, broad-based growth across the networking stack, with the layers benefiting in different ways and on different timelines.

The rest of this article is for paid subscribers. In it I unpack what is already happening and what signals we are seeing, how agentic AI changes traffic, what a distributed campus looks like, and where the demand lives.

Read more

The focus is going to be on the InP substrates / wafers.

There was a report put out by Reuters and some other news around InP lately so this is a good time to revisit.

China's control over indium phosphide exports threatens AI data centre rollout

These are the foundational building block of the entire optical supply chain.

There are many steps prior to the creation of lasers.

1. First, you have the raw materials. That means indium, phosphorus, dopants, crucibles, and high-purity inputs.

2. Then those materials get turned into a crystal. That crystal gets sliced and polished into a blank InP substrate.

3. That blank substrate then goes to an epi house or an internal epi line. The epi process grows extremely precise semiconductor layers on top of the blank wafer.

4. Then that epi wafer goes into a device fab. The device fab turns it into lasers, photodiodes, modulators, or photonic chips.

5. Then those devices get packaged into transceivers, optical engines, coherent modules, CPO systems, NPO architectures etc.

So the chain looks like this:

Raw materials
→ InP crystal growth
→ blank InP substrate
→ epitaxy
→ laser / photodiode / modulator / PIC device fab
→ optical component
→ module or system
→ AI data center network

So you can see how if InP substrate supply gets tight, the impact can move through the entire optical supply chain.

But there is one thing that I want to emphasize because it could get really confusing and is often mixed up in conversation.

‘InP capacity’ can mean several different things.

• It can mean blank substrate capacity.

• It can mean epi wafer capacity.

• It can mean device fab capacity.

• It can mean final laser output.

Those are different layers.

Subscribe now

The most important distinction

The blank InP substrate is the starting wafer. It is the foundation.

Before Coherent, Lumentum, AAOI, Nokia/Infinera, IQE or any other device maker can build an InP laser or photodiode, somebody has to make the blank wafer.

The main substrate / wafer names are:

• AXT / Tongmei

• Sumitomo Electric

• JX Advanced Metals

• Yunnan Germanium

• InPACT (I believe)

• Guangdong Xiandao

• Zhuhai Dingtai Xinyuan

Coherent and AOI have referenced using more than five raw substrate suppliers. That tells us they are all trying to diversify the input layer, but the full supplier roster is private.

AXT, Sumitomo, (and JX less-so) are the major non-module names in the blank InP substrate discussion.

Coherent, Lumentum, AAOI, and Nokia/Infinera are mainly downstream device, laser, PIC, module, or system companies.

They can be very vertically integrated.

They can own InP fabs. They can run internal epi. They can produce lasers and photodiodes.

But they still need blank substrates feeding the front of the process.

This report is for informational and educational purposes only. It reflects my personal opinions and research process as of the publication date.

This report should be treated as general market commentary, rather than personalized financial, investment, legal, or tax advice. Every reader has a different risk tolerance, time horizon, financial situation, and portfolio construction.

I may own shares, options, or other securities in companies discussed in this report. I may buy, sell, add to, reduce, or exit any position at any time without prior notice.

AXT

AXT makes compound semiconductor substrates. In this case, the important one is InP.

AXT is trying to add more substrate capacity. As much as possible. They are doubling this year and doubling again next year.

Making InP substrates is a manufacturing skill.

You cannot simply decide to make InP wafers and then ship qualified material into the optical supply chain a few months later.

Crystal growth is hard. Yield is hard. Qualification is hard. Scaling is hard.

That is why substrate capacity can become such a powerful bottleneck.

It is small in dollars compared to the full AI data center buildout, but it sits at the front of the chain.

If this layer gets tight, everything downstream can feel it.

AXT is no longer only hearing from direct customers.

They are hearing from the customer’s customer.

The hyperscalers care because they need the optical supply chain to scale.

The transceiver makers care because they need lasers and detectors.

The laser makers care because they need epi wafers and substrates.

The substrate suppliers care because they need enough capacity, permits, furnaces, raw materials, and qualified customers.

“There’s been a lot of encouragement from hyperscalers and hardware companies for their suppliers to enter into long-term supply agreements with AXT.”

The permit issue is more nuanced than it looks

The Chinese permits get brought up often. We just heard that the CEO of Coherent went to China to discuss the export permits. China is in a really powerful position here where they can threaten the U.S. domestic buildout if they can choke out these materials.

But the issue is not simply “AXT cannot ship to U.S. customers.”

The real situation is more nuanced.

AXT said:

“We’re still being encouraged to apply for export permits for U.S. customers both in the U.S. and in other global regions.”

Then they said:

“At the moment, obviously we are getting permits pretty readily for U.S. customers based in other global regions.”

That means location and destination matter.

A U.S. customer receiving material outside the U.S. can have a different permit outcome than a shipment directly into a U.S.-based facility.

AXT also said:

“We’ve been contacted by the Ministry of Commerce in China on a number of U.S. applications right now to submit more data that gives us an encouraging sign that they’re still looking at U.S.-based permits, and there’s still a possibility to get a permit for the U.S. in the near future.”

So the right framing is that AXT is getting permits for some global shipments.

Permits for U.S. customers outside the U.S. appear to be moving more easily.

Direct U.S.-based permits remain the harder category.

And the timing is outside AXT’s control.

AXT said:

“The most significant single factor to our growth in Q2 and beyond is the success and timing of getting export permits.”

So, demand can be there, inventory can be there, customer orders can be there.

But if permits lag, revenue and supply flow can lag.

Many of the companies we care about have manufacturing facilities outside of the U.S. So hopefully they are having an easier time sourcing substrates. More on that later.

The epi layer

The blank substrate is only the starting wafer.

Before that wafer becomes a laser, photodiode, modulator, or photonic chip, someone has to grow the actual semiconductor layers on top of it.

That process is called epitaxy.

An epi house takes the blank wafer and grows atomic layers on it.

That is the middle of the chain. They are not necessary, as we will see in the next section how vertically integrated many of these companies are. But they can be leveraged during constraint.

IQE is one of the public examples of this epi layer.

“In the last year, we have seen market demand from these indium phosphide solutions significantly increase due to the exponential growth in AI.”

Then they made the timing clear:

“The nature of the supply they’re demanding is very much in the here and now.”

This matters because epi houses can also become constrained.

If they have demand but lack blank substrates, they can run into input shortages.

If they have substrates but lack enough MOCVD or MBE capacity, they can run into epi capacity limits.

And if customers want more capacity, IQE is saying those expansions need customer commitment.

On substrate availability, IQE said:

“Indium phosphide is a very sought-after commodity these days.”

They also said:

“We are working closely with our suppliers to mitigate any kind of indium phosphide shortages that we are seeing.”

And then they explained the strategy:

“Ultimately, the main component here will be to dual source or to increase really our sourcing capabilities and ensuring that we are just not reliant on one supplier alone, and which will expose us even more.”

Lumentum

Lumentum operates fundamentally as a downstream InP company. While they possess robust internal InP capabilities, they still require blank InP substrates before they can turn those wafers into finished optical devices.

Luckily (or tactfully) they had foresight in securing meaningful substrate supply early in the cycle. They tied this supply to Japanese suppliers like JX and Sumitomo, rather than relying exclusively on China-based production. This proactive approach provided them with a distinct initial advantage and much-needed visibility during a growing global InP shortage.

Initially, this strategy appeared to provide total coverage. During their Q3 FY2026 earnings call, CEO Michael Hurlston:

“we’ve executed some long-term agreements that we feel leave us in pretty good shape on substrates”.

However, the AI-driven optical ramp expanded far faster than their original supply plan could accommodate. Just weeks after that earnings call, Hurlston spoke at the Mizuho Technology Conference and captured the tension of their setup:

“We’re uncomfortably comfortable right now with our substrate position”.

Addressing the aggressive shift in the market’s demand curve, he explicitly admitted that their early agreements are no longer enough:

“Numbers have gone up considerably since then. So I’d say, no, we’re not well covered anymore. We’re going to have to find alternate sources of supply”.

This introduces a vital geographic workaround to the story. While Lumentum’s Japanese suppliers offer a cleaner supply base, the reality is that Chinese suppliers are bringing capacity online at a much faster rate. To access this essential capacity while navigating complex geopolitical trade restrictions, Lumentum relies on a highly creative supply chain route.

“...we can have substrates go from China to the United Kingdom. And in so doing work around some of the licensing and restrictions, restrictive environment, so we’re uniquely positioned in that way”.

Downstream capacity is equally critical to this equation. Lumentum is actively expanding its InP device footprint by bringing a newly acquired fifth fab in Greensboro, North Carolina, online. As management confirmed:

“plans to convert the facility from gallium arsenide to indium phosphide are well underway”

which will significantly boost their domestic laser output. Yet, even a massive new InP fab cannot scale unless blank substrates consistently arrive at the front of the line.

Coherent

Coherent also operates exclusively as a highly advanced downstream InP device manufacturer, partnering with upstream providers for blank substrates. When management highlights InP wafer capacity, they are detailing their internal device fabrication capacity. Coherent transforms raw InP substrates into highly complex optical devices like Lasers, Photodiodes, Continuous Wave (CW) laser sources, Electro-absorption Modulated Lasers (EMLs), and other critical optoelectronic components.

Coherent has built one of the strongest internal InP manufacturing footprint in the global optical industry, with their 6-inch InP platform serving as the absolute center of the story.

“We are ramping production in parallel at two sites, Sherman, Texas, and Järfälla, Sweden. We are in production with three different types of key transceiver components on six-inch indium phosphide, EMLs, CW lasers, and photodiodes.”

The structural economic advantage makes this completely transformative. A larger wafer fundamentally changes the output curve, giving Coherent massive scale from the exact same manufacturing footprint.

“A six-inch wafer, compared to a three-inch wafer, will produce more than 4x as many chips at less than half the cost.”

“Every new 6-inch line we put in place is like us putting 4 3-inch lines, right? So we’re scaling at 4 times the rate of if we’re doing it on 3-inch lines.”

The manufacturing yields make this even more compelling. The team is executing well, seemingly managing standard early-ramp friction.

“We’re in production on three different types of devices, EML, CWs, and PDs. All three of those devices have yields on 6-inch that are higher than our 3-inch production yields.”

Coherent dedicates this platform to the most critical, high-margin AI optical components. EMLs power high-speed pluggable transceivers. CW lasers provide the necessary light sources for silicon photonics and future CPO architectures.

They are also hyperscaling this capacity aggressively. The execution metrics highlight a massive acceleration in wafer starts and a rapid timeline for doubling internal InP output.

“And if you look at the number of wafer starts that we’re starting this quarter, we’re basically at 80% of that target capacity already. So we’re starting wafers at 80% of the goal of doubling capacity, which is really strong and frankly ahead of schedule.”

“We remain on track to achieve our goal of doubling internal indium phosphide output capacity by the end of this calendar year. Based on current execution, we now expect to reach that milestone one quarter earlier than originally planned.”

“We also expect to more than double our internal indium phosphide capacity again by the end of calendar 2027.”

Coherent expands the device layer precisely while the broader industry remains in a state of high demand.

Expanding device capacity requires securing the underlying raw materials though. Six-inch device production completely relies on qualified six-inch blank substrates. Coherent addresses this proactively by locking down their upstream supply well in advance and securing a highly diversified supply base.

“...we have over 5 different substrate suppliers. Today... we work with those suppliers... on the next 1, you know, like 3 to 5 years of capacity that we’re going to need. So we have, you know, in some cases very long-term agreements place.”

“...we have commitments from our suppliers to supply the necessary indium phosphide substrates to support that.”

By structuring their supply chain this way, management achieves complete supplier diversification, locking up multi-year visibility, and securing the raw substrate supply required to support their massive internal capacity ramp. As raw substrates become increasingly scarce, Coherent’s ability to extract four times the chips per wafer at significantly higher yields makes their 6-inch economics profoundly more valuable.

That being said, in the Reuter’s report, it was noted that the cost of a 6-inch substrate has surged 250%. I don’t know over what timeframe.

Also, a good question is ask is if they are so confident in their substrate supply, why would the CEO go to China and discuss InP export permits? My thought process is two fold: One is that they are ramping so hard and like Lumentum, the capacity that they want to bring online needs more substrate than forecasted. And two is that more is better. The more visibility they have in supply, the more confident tehy can be on their ability to take on orders. it also helps their pricing power. if substrates are so tight, and Coherent doesn’t have a meaningful ability to multi source enough, they are subject to significant price increases without leverage. So the question is whether the economics of the 6-inch could offset it, in addition to obvious price increases to the end customer.

Applied Optoelectronics

Just like Lite and Cohr, AAOI is heavily integrated downstream from the blank substrate layer, operating as a sophisticated laser and transceiver manufacturer anchored by its own InP fabrication plant in Sugar Land, Texas.

“Again, when you look at the universe of companies that have optical technologies that can make transceivers, there’s a certain subset of companies that can do that. When you narrow that down to companies that actually make their own optical devices in-house and so have the capacity to expand and grow their ability to make those indium phosphide devices themselves, that’s an even smaller subset of those companies, and I think that’s really where the bottleneck right now is in the industry is on that laser capacity...”

Murry frames the substrate constraint it as a highly solvable, geopolitical issue rather than a terminal resource constraint. He notes that because Indium Phosphide is a fundamentally common material, the supply chain constraint is a “more tractable problem to solve than something where you’re dealing with a material that is fundamentally rare”.

To help insulate their Texas fab from these geopolitical shocks, CEO Thompson Lin confirmed the company has aggressively diversified its sourcing and built a massive inventory buffer:

“We already got 4 to 5 suppliers... Four of them are outside of China. I would say right now we should have enough inventory minimum for almost 1 year. Since the volume will increase so fast, we are making calls with all the suppliers.”

AAOI is not just making phone calls though. They are actively managing this supply chain on the ground. Leveraging their “very good long term relationships” with major InP suppliers, Murry noted that AOI actually has “people on the site to address any capacity issues”. While management has kept the exact details of their future procurement contracts under strict wraps, with Murry saying “we can’t say too much about it specifically at this point because a lot of it’s under discussion still”, their ultimate baseline is unwavering. As Murry explicitly summarized their substrate position heading into the 2027 ramp: “we feel incredibly comfortable”.

With the raw materials seemingly secured, the strategic focus shifts to expanding internal output.

AAOI is aggressively pushing its engineering roadmap forward, upgrading its fabrication equipment to process significantly larger wafers.

“Right now, we are working on 3-year long-term agreement with several customers... That’s why, not only transceiver, we are expanding very fast about our laser capacity. Right now, we have been doing the 4-inch growth process. Our target is go to 6-inch by end of next year. Yes, I think we need to do more investment to meet the demand for CPO market. As you know, the CPO laser is about 300-400 milliwatt compared to 70 milliwatt for 800G transceiver and 100 milliwatt for 1.6 terabit transceiver. The size is much bigger, okay? Minimum maybe 5 times or 6 times bigger. That’s why we already go from like 2 inch to 3 inch to 4 inch in the past 18 months. We still plan to go to 6 inch by end of next year. That will increase our capacity a lot.”

AAOI has no intention of functioning as a merchant laser supplier to bail out its competitors.

Instead, they intend to capture the full margin of this bottleneck by hoarding these high-power InP lasers exclusively for their own finished module production. Murry made this strategy perfectly clear regarding their future 400,000-unit-per-month ELSFP capacity targets:

“...our anticipation is that we’re going to be shipping mostly those ELSFP models. We’re going to be making the high power lasers for those modules in house. But we don’t anticipate selling that. We’re going to use it pretty much for the in house production...”

If AAOI can execute the integration of this new equipment and successfully upgrade their wafer sizes, their ultimate ambition is quite clear. By utilizing their secure substrate supply to feed their internal fabrication plant, Lin projects that AAOI should become a “top 3 laser supplier worldwide” by the end of 2027.

Let’s see!

So this report was really intended to just explain the supply chain in (hopefully) a digestible way. And to introduce the companies and how they are positioned.

I do believe this problem is workable. But if in the coming quarters these companies capacity ramps are increasingly being gated by substrate supply, and it isn’t easing, that could be a concern. Because lots of these valuations are dependent on continued growth.

I feel confident in the coming quarters based on everything management is saying, but of course I will keep you all updated as we get more and more updates.

Disclaimer

This report is for informational and educational purposes only. It reflects my personal opinions and research process as of the publication date.

This report should be treated as general market commentary, rather than personalized financial, investment, legal, or tax advice. Every reader has a different risk tolerance, time horizon, financial situation, and portfolio construction.

I may own shares, options, or other securities in companies discussed in this report. I may buy, sell, add to, reduce, or exit any position at any time without prior notice.

The companies discussed may be highly volatile, speculative, small-cap, or exposed to significant business, execution, customer concentration, supply-chain, regulatory, geopolitical, and valuation risks. Investing in public equities involves risk, including possible loss of principal.

Any forward-looking statements, estimates, price targets, frameworks, scenarios, or projections are based on assumptions that may prove inaccurate. Actual results may differ materially from expectations.

Information in this report is gathered from sources believed to be reliable, including company filings, earnings calls, investor presentations, press releases, industry reports, and other public materials. Accuracy, completeness, and timeliness cannot be guaranteed.

Readers should conduct their own research, review primary sources, and consult a qualified financial, tax, or legal adviser before making investment decisions.

Nothing in this report should be interpreted as a solicitation, offer, or recommendation to buy or sell any security.

If you invest in Photonics you gotta read this.

Yesterday I mapped out NPO and the companies that could benefit as optics moves closer to the package.

This is the follow-up. And some of my opinions have actually changed in the last 24 hours.

The SemiAnalysis CPO report that was released today is forcing the market to revisit the other side of that debate.

If full scale-up CPO takes longer than investors expected, who actually wins and loses? And which companies are positioned across the transition instead of relying on one perfect architecture outcome?

This idea has precedent.

If you were with us back in March, Broadcom’s earnings call created the first major version of this trade. Hock Tan said Broadcom’s custom XPU customers could likely keep using direct-attached copper through 2028. The stock reaction was immediate. Credo rallied. Lumentum and Coherent sold off. The market read was simple… copper gets more time, CPO moves right, optical names face pressure.

Now SemiAnalysis has brought the same question back into focus.

I think the first market reaction was too blunt. We have always known that Optical Scaleup was a 2028-2029 volume story.

But the CPO timing debate changes how you could look at the optics stock map. It gives copper more duration. It gives pluggables more runway. It creates timing risk for companies that were being valued on a fast optical scale-up TAM expansion. It also supports NPO, lasers, InP capacity, fiber attach, analog components, silicon photonics, and test.

The optical thesis is still very strong. But maybe the form-factor sequence changed.

The industry is still moving toward more optical content closer to compute. The path there looks more fragmented.

• Copper stays where reach allows.

• Pluggables stay where serviceability and multi-sourcing win.

• LPO, LRO, and XPO bridge the gap for customers that want lower power while preserving the pluggable ecosystem.

• NPO gets pulled forward where 200G per lane copper becomes harder.

• CPO stays strategic where density and system control justify deeper integration.

• OCS becomes a separate optical branch for customers that can redesign the fabric around optical switching.

The question is where the dollars accrue first.

I will lay out my thoughts in this report on who the main beneficaries of this report are.

Who is largely un-impacted.

And who could really feel the headwind.

Because all of the companies sold off.

But many of them shouldn’t have.

And in my opinion, that creates a great opportunity.

Read more

We have fairly consistently seen a 20-30% drawdown after earnings.

Only for that gap to be closed (in most instances) within a couple weeks.

Lot’s of these names are consolidating now off of highs.

This one company in particular I am going to present today has retreated to a level that I find attractive.

It is still “expensive” to many people (as are most of these companies).

But it is my opinion that the market will reward it over the coming days/weeks.

So in this post I am going to layout the company, their products, how they are positioned over the coming years, what my strategy has been, what levels I am interested in, and my 12 & 24 month financial frameworks.

This report is for educational and informational purposes only. It reflects personal research, opinions, estimates, assumptions, and interpretations as of the publication date. Nothing in this report should be interpreted as investment advice, financial advice, legal advice, tax advice, or a recommendation to buy, sell, hold, short, or otherwise transact in any security.

Read more

It is gaining a lot of traction this weekend and I want to revisit it in depth.

And present to you the companies that I think are best positioned to capture dollars.

As a quick foundation, give this a read. It’ll only take a few minutes and it’s very watered down.

There was a FundaAI report that dropped recently.

I haven’t read the article, but I wanted to take this opportunity to discuss all things NPO.

• Why it is being used

• Who supplies into it

• What it means for CPO adoption

• Where do dollars accrue?

I put out a tweet asking people who they think the main beneficiaries are from this information.

And it was exciting to me that most of the relevant companies weren’t being highlighted.

That means there is alpha here, even in a follower base that is very intune with this industry.

Some of the companies with exposure here will surprise you…

So let’s unpack it all.

First, what is NPO?

The simplest way to think about it is that the optical engine moves much closer to the GPU, XPU, switch ASIC, or accelerator.

But it still stays outside the most aggressive part of the advanced package.

A traditional pluggable optical module sits at the front of the box. The ASIC has to send very fast electrical signals across the board to reach that module. At lower speeds, that worked well enough. But at 200G per lane, 224G signaling, 1.6T links, and eventually 3.2T links, the electrical path starts to become a problem.

The signal degrades, power goes up, the board gets harder to design, retimers and equalizers become more important, thermals get more complicated, and the whole system becomes harder to scale.

NPO tries to solve that by moving the optical engine much closer to the chip. That shortens the electrical path. It lets the signal turn into light earlier. And that can help the system become more power-efficient, higher-density, and easier to scale.

NPO is so important is because it gives the industry a practical middle step.

It is more advanced than a front-panel pluggable and it is less aggressive than putting optics directly inside the full package.

That middle ground is why hyperscalers and semiconductor companies are interested.

Why NPO is being used now

The AI network is changing over time.

The first major AI networking wave was mostly about scale-out. That meant connecting many servers and many racks together with Ethernet fabrics, switches, pluggable optics, DSPs, retimers, and a lot of fiber.

The next wave is increasingly about scale-up. Scale-up is the network that helps many GPUs or XPUs behave more like one large compute system. This layer needs extremely high bandwidth, low latency, and strong reliability. It also sits much closer to the compute layer than traditional data center interconnect.

That is why copper starts to get pressured. Copper can still work well over short distances. Inside the rack, it can be lower latency, cheaper, and easier to deploy than optics. But as the scale-up domain gets larger and the lane rates move higher, copper becomes harder to push. The reach gets shorter, the power rises, the signal integrity challenge gets worse, and the system needs more high-speed electrical support around the link.

That is where NPO enters the conversation.

Why the Rubin Ultra / NVL576 discussion

The reason NPO gained so much attention this weekend is because of the claim that Rubin Ultra / NVL576 could use meaningfully more optical engine content per GPU than many people had modeled.

If optical engine density rises per GPU, the supply-chain impact can be large.

For example, if a 576-GPU system uses 2.25 optical engines per GPU, that implies roughly 1,296 optical engines.

If that moves closer to 4.0 optical engines per GPU, that implies roughly 2,304 optical engines.

So rather than quesitoning if CPO will ever be fully adopted, I want to focus on how much optical content gets pulled closer to the GPU before full CPO becomes dominant.

This may change the beneficiary list.

If NPO scales, dollars flow into optical engines, silicon photonics, drivers, TIAs, photodetectors, lasers, VCSELs, external light sources, fiber attach, connectors, and passive photonics.

What NPO means for CPO adoption

I view NPO as a bridge.

Both NPO and CPO are pointing in the same direction. Optics are moving closer to compute. The difference is how aggressive the integration is.

CPO pushes optics deeper into the package or directly next to the switch ASIC. That can offer major power and density advantages over time. But it also creates harder packaging, yield, thermal, and serviceability questions.

NPO gives the industry a more modular way to move first.

It gives customers more flexibility around qualification, repair, sourcing, and manufacturing.

That is why NPO should become the first major commercial ramp before full CPO becomes more mature.

Where the NPO dollars accrue

Read more

Read this as a primer:

TLDR; AI clusters are getting too large to stay inside one building. Power, land, cooling, fiber access, permitting, and physical infrastructure are becoming real constraints, so hyperscalers are spreading compute across multiple buildings, campuses, metros, and regions.

That’s scale-across. And Ciena just showed us it’s moving forward with full force.

Today we witnesse a company who prints 40% revenue growth, 290% EPS growth, raises full-year guidance, grows backlog by $600 million in a single quarter while already shipping at record levels, and announces the first hyperscaler multi-rail order with language around hundreds of millions of dollars over multiple years.

So it’s clear that distributed AI compute is here, and the optical network is a major bottleneck.

The read-through spreads well beyond Ciena. Pump laser constraints point to Lumentum. Multi-rail architecture points to Nokia. InP capacity expansion points to Coherent. More fiber routes and physical density point to Corning. I want to walk through all of it, starting with the Ciena numbers because they set the table for everything else.

The Quarter

Q2 FY26 revenue was $1.5707 billion, up 39.5% year over year.

Adjusted EPS was $1.64, up 290% versus last year.

Adjusted gross margin was 44.9%, up from 41.0% last year.

Adjusted operating margin was 19.5%, up from 8.2% last year.

Adjusted EBITDA was $341.8 million, up 192.9% year over year.

Revenue growing nearly 40% while operating expense grew only 7.7% is what real operating leverage looks like. We’re talking about a company in a supply-constrained demand environment where pricing discipline, product mix, and scale are all working simultaneously. That is a really big deal.

The guide moved higher too.

Q3 revenue guided to $1.625 billion plus or minus $50 million, adjusted gross margin to 45% plus or minus 50 basis points, and adjusted operating margin between 19% and 20%.

Full-year FY26 revenue guidance moved to $6.3 billion plus or minus $100 million, implying 32% growth at the midpoint.

That is a mid-40s gross margin profile while revenue grows above 30%. Very different from what most people still associate with telecom equipment.

Hard to believe the stock went down on all of this! But many of these companies are priced high and the market tends to sell off quickly. It happened last earnings as well.

And we don’t need to guess what happened after last earnings selloff…

Backlog

Backlog increased by more than $600 million sequentially to $7.7 billion, and management expects to exit the year with backlog even higher. Think about what that means. They just printed record revenue, raised guidance, and still grew backlog by $600 million in one quarter. Customers would take more product in 2026 if Ciena could ship it. This is a supply-constrained cycle, and the demand environment is deeper than current revenue suggests. We are seeing this across all of our names and it’s not slowing down.

Management was clear about what this backlog actually represents. These are real deployments tied to services, customer programs, and urgent build-outs, not inventory channel fill. That distinction is impotant because it tells us the demand is structural.

They called out two constrained inputs on the call: modems and DSPs, and pump lasers. The pump laser comment is one of the biggest sector signals in the entire call, and I’ll come back to it when we get to Lumentum. We’ve talked about it already and it’s nice to hear the validation on the other side.

The Segment Mix

Optical Networking came in at $1.1 billion, or 70% of total revenue.

Routing and Switching was $174 million, or 11.1%.

Total Networking Platforms hit $1.274 billion, or 81.1% of the company.

Platform Software and Services added $93.9 million

Blue Planet $23.4 million

Global Services $179.4 million.

Optical is still the engine, which makes sense given Ciena’s core advantage in coherent optical systems and line systems. But Routing and Switching growing 88% year over year tells us the AI exposure extends beyond traditional WAN and DCI. Ciena is getting pulled deeper into data-center operations, and the direction of travel is clear.

WAN and DCI today, around the data center today, inside the data center over time.

HyperRail

HyperRail was the star of the call.

Ciena announced the first multi-rail order from a leading hyperscaler. This begins rolling out in 2027, and management described these opportunities as potentially hundreds of millions of dollars over multiple years. HyperRail moved from a product concept to a customer-backed deployment path real fast.

Quick bit on Multirail. A normal optical upgrade is about faster wavelengths or higher-speed modules. Multi-rail is a density architecture. The question is all about how many fiber pairs can I support, how much power does the amplifier site consume, how many racks fit in the hut, and how fast can I turn up hundreds of fiber pairs. This pulls on a much broader set of the supply chain.

More rails mean more optical paths.

More optical paths mean more amplification.

More amplification means more pump lasers.

More coherent links mean more narrow-linewidth lasers.

More physical links mean more fiber.

Every layer of the stack gets pulled when multi-rail scales. Ciena’s own materials describe multi-rail photonic network designs requiring tens of petabits of capacity across hundreds of fiber pairs. That’s not incremental optical demand. That’s a new category of optical demand sitting on top of everything that already exists.

DCOM

Routing and Switching grew 88% year over year, driven primarily by DCOM, and this deserves some time.

DCOM, Data Center Out-of-Band Management, is the network layer used to provision, monitor, recover, and manage data-center infrastructure at scale. Hyperscale AI data centers need more than a front-end training network. They need a backend operational layer that can handle provisioning, configuration, recovery, and management across massive numbers of servers, racks, and clusters.

Ciena co-developed DCOM with Meta. The critical update this quarter is that it’s no longer a Meta-specific architecture. Ciena received initial orders from a second hyperscaler, and lab qualification is progressing with a third. Management framed the total DCOM opportunity at $1 billion to $3 billion by 2029.

The way I think about it is that HyperRail connects AI factories across distance and DCOM helps operate the AI factory. Both are part of the same larger shift, which is Ciena moving from telecom equipment into AI infrastructure architecture.

Coherent Modules, Pluggables, and the Nubis Portfolio

Ciena secured a new hyperscaler win for high-performance coherent modules to be deployed across metro and long-haul DCI networks. That shows Ciena monetizing its coherent systems capability across multiple consumption models like full systems, modules, pluggables, components, and software and services wrapping all of it.

400G and 800G pluggables remain on track to more than double from 2025. Ciena also won with a major switch OEM using WaveLogic 5 and 600G plugs.

The Nubis product family maps cleanly across the full architecture. Nitro, the linear redriver, is on track for general availability this summer. Vesta 200, the 6.4T CPO optical engine, has seen increased demand for open ecosystems over the past 90 days. Scale-up gets Nitro and eventually optical inside or near the rack. Scale-out gets Vesta, CPO, optical engines, and pluggables. Scale-across gets RLS, HyperRail, coherent modules, WaveServer, and 800ZR. DCOM handles the operational network layer. That’s a real portfolio covering multiple layers of the AI infrastructure buildout.

I am not going to unpack all of this in this earnings debrief, but I have unpacked it all in previous articles!

Click here, scroll down to CIEN, and read through them if interested.

Just know the Ciena is expanding its offerings outside and inside the data center.

Service Providers and MOFN

Service provider revenue grew 28% year over year, and India service-provider revenue more than doubled, driven by MOFN deployments.

MOFN, Managed Optical Fiber Network, is where service providers build or manage dedicated optical connectivity for cloud and AI customers. This is worth understanding because hyperscaler AI demand can show up in two ways. One is direct cloud purchases from Ciena, and the other is service-provider purchases that are really cloud demand in disguise. A hyperscaler expanding in India, Southeast Asia, Europe, or the Middle East may rely on local carriers for fiber, regulatory relationships, route ownership, and operating capabilities. That turns cloud AI demand into service-provider optical spend, and Ciena captures both sides.

So the demand base here is broad. Direct cloud demand, service-provider optical refresh, MOFN pulled by cloud and AI infrastructure, neoscaler demand, subsea and long-haul demand. That’s quite a durable setup.

Margins

Adjusted gross margin was 44.9%.

Adjusted operating margin was 19.5%.

Revenue grew 39.5% while adjusted operating expense grew 7.7%.

The operating leverage is flowing through the model exactly the way we like to see it.

Management talked about engineering cost reductions, product mix, price optimizations, and value exchange with customers. Value exchange is the phrase worth focusing on. It means Ciena is using supply scarcity and customer urgency to improve economics across pricing, terms, product mix, working capital, and supply-risk sharing. When demand exceeds supply and customers need your product to run their AI infrastructure, that’s where pricing power lives. Rosenblatt raised concerns about laser pricing impacting Ciena’s margins. Was nice to see the proof against this and their ability in insulate and pass through.

HyperRail is margin-accretive versus single-rail RLS. If multi-rail becomes a larger part of revenue in 2027 and carries better economics, this cycle can support both revenue growth and margin expansion at the same time. That’s the operating leverage setup.

The rest of this post is for paid subscribers. Below the paywall I get into the names. Lumentum, Coherent, Corning, Nokia. What Ciena's call means for each one and how I'm thinking about positioning across the optical stack as this cycle broadens. I will also discuss the levels I am interested in adding more shares, my current weighting, and my 12 & 24 month financial modeling.

None of this is financial advice or a recommendation. Do your own research.

Read more

This article lays out three things:

1 - My Optics MOCK ETF. This is a reflection of how I would weight these companies and how I would build out a diversified portfolio for exposure across the stack.

2 - All my 12 & 24 month price targets for my optics/interconnect adjacent stocks

3 - A link to my Tradingview watchlist for all the companies I am tracking in this industry

Along with that, every company that I cover will have every article I’ve written attached to it. And this will get updated each day. So if you are interested in any of these companies, you can easily find every article I have written about them.

This document is going to be extremely valuable and provide you with so much insight into how I see the industry.

Read more