FIT TECH DAY 2026 | Lumentum, Avalanche Thinking, NTT, ams OSRAM | Light at Scale: The Next Frontier in AI Data Center Interconnects

2 days ago

41 min read

The most important takeaway from this event isn't “which packaging approach won.” It's that eight speakers, in five different ways, said the same thing: interconnect has graduated from component to architecture. Links are no longer just the wiring that connects compute; they are a fifth subsystem that must be designed alongside compute, memory, power, and cooling.
The four keynotes reached a remarkably consistent consensus: pluggables, NPO, CPO, and OIO are coexisting, stacked generations, not replacements. Avalanche Thinking's Chuck Mattera put it most plainly — “these generations will layer on top of each other; they won't simply replace one another” — while ams OSRAM's Ashkan Seyedi added the practical angle: NPO is the easiest tier to deploy.
The timeline was spelled out: Lumentum's Wupen Yuen said outright that 2028 will be the first year of optical scale-up, and 2030 will bring full optical scale-up, with scale-up consuming 100x as much optics as scale-across. It was the only year stated all day that can actually be checked against reality.
The two hardest contrarian lines came from the panel. Semtech CEO Hong Hou: “Don't ignore copper — equalizers can still recover 15–20 dB, pushing copper's effective reach to about 3 meters.” Wei-Ping Huang: “At meter-scale distances, copper still leads optics by one to two orders of magnitude in bandwidth density and energy efficiency; CPO can close some of the gap, but the gap is still there.”
The real ceiling isn't the choice of technology path; it's materials and alignment. When moderator Dr. Thomas Liu wrapped up, he didn't talk technology at all. He told the room: “Please help us solve the material shortage.” That is also why Hon Hai Research Institute is betting on both InP at 300 Gbps+ and quantum-dot comb lasers.
For Taiwan's supply chain, the signal was concrete: FIT has gone from connector supplier to co-designer of NTT's photonic-electronic convergence interface (500 pins, 200G per pin, detachable). The real opportunity isn't “can it be built” but “can it ship in volume every quarter.”

1. Introduction: A Conference That Put “Connectivity” Center Stage

On September 16, 2026, FIT (Foxconn Interconnect Technology) hosted FIT TECH DAY 2026 in Taipei. The subtitle put its position right on the wall: Light at Scale — The Next Frontier in Data Center Interconnects.

That subtitle is worth unpacking. It isn't “Light” (optics is great), nor “at Scale” (things are big). It binds the two together: light must exist at scale. For thirty years, optics' problem was never “we can't build it”; it was “we can't build enough, fast enough, consistently enough.” Putting scale in the theme declared that the event was about industrialization, not technical feasibility.

The agenda design was telling too: four keynotes plus one panel, deliberately layered from the bottom up — light sources and optical engines → industrialization and volume production of photonics → electronic-photonic convergence in network architecture → new materials — closing with a panel tying together optical modules, CPO, high-speed interconnect, and the broader ecosystem.

The speaker lineup was far heavier than a typical single-company tech day:

Session	Speaker	Title
Keynote 1	Dr. Wupen Yuen	President, Global Business Units, Lumentum
Keynote 2	Dr. Chuck Mattera	Founder & CEO, Avalanche Thinking; former CEO of Coherent
Keynote 3	Dr. Yoshiaki Sato	Executive Adviser, NTT Innovative Devices Corporation
Keynote 4	Dr. Ashkan Seyedi	VP & GM, Optical Interconnect BU, ams OSRAM
Panel moderator	Dr. Thomas Liu	Optical Communications Consultant, FIT
Panel	Dr. Hong Hou	CEO, Semtech
Panel	Dr. Wei-Ping Huang	Founder & Non-Executive Director, Ligent Group
Panel	Mr. Ernest Muhigana	Senior Director, Lightwave & MLH BU, MACOM
Panel	Dr. Hao-Chung Kuo	Director, Semiconductor Research Center, Hon Hai Research Institute (HHRI)

Compress the whole day into one line, and it looks like this:

Every link in this chain had someone presenting it, and the speakers barely contradicted one another. That in itself is a signal — when upstream, midstream, and downstream players lay out the same causal chain on the same day, it usually means the industry has moved past “should we do this” and into “who can deliver volume.”

2. The Host Speaks First: The Heart of Compute and Its High-Speed Arteries

The opening address came from Brand Cheng, Chairman of Foxconn Industrial Internet (FII). His framing was very much the group line: Hon Hai is pushing “four-capability integration” — technology, manufacturing, management, and vertical integration. But two sentences are worth writing down:

Simply stacking chips is no longer the answer. The bottleneck that truly sets the upper limit of overall compute lies in interconnect bandwidth, power consumption, and heat dissipation.

If FII's end-to-end system integration in high-end AI server racks and power architecture is the powerful heart of compute, then FIT's world-leading high-speed connectivity, copper cable, and optical communications technologies are the high-speed arteries that let compute flow.

This matters more for the industry than it sounds. It amounts to the Hon Hai group publicly rewriting FIT's positioning from “connector supplier” to “a necessary subsystem of compute infrastructure”. His next line was more direct: AI infrastructure has moved from point-technology competition to whole-system efficiency competition — whether compute, connectivity, power, and cooling can be co-optimized on one platform.

He closed by quoting Hon Hai Chairman Young Liu: this era is not just about make in Taiwan, but make with Taiwan.

When the afternoon panel opened, moderator Dr. Thomas Liu recapped the morning with a number: the four keynotes were all saying the same thing — AI compute demand keeps accelerating, and compute capacity is doubling roughly every seven months (citing the opening talk). That figure is the starting point for every argument that followed.

FIT Chairman Sidney Lu then introduced the agenda, framing the day in one sentence: from light sources, materials, components, and network architecture to systems and manufacturing, every talk would help piece together the connective landscape of AI's next era.

STT Take: These two addresses are classic “home-field narrative,” but they make the event's motive clear — FIT isn't here to sell connectors; it's here to secure a seat at the table while optoelectronic architectures are being defined. The NTT keynote later showed it already has that seat.

3. Keynote 1 | Lumentum's Wupen Yuen: This Time Really Is Different — and 2028 Has a Date

Wupen Yuen opened with a self-deprecating line that hits a nerve in optical communications:

I've been in this industry for thirty years. For thirty years, optics has always been “the next big thing.” This time is different.

He then broke down “why it's different” into a clean argument.

3.1 Demand Side: Data Producers Have Shifted from Humans to Machines

All past traffic, whether YouTube or cloud computing, was fundamentally produced and consumed by humans. AI changes that: we are now creating and delivering human-level, even superhuman-level, intelligence at massive scale. Producers and consumers are no longer bounded by human population or hours in the day, so the required compute and communication jump to an entirely different order of magnitude.

The result: data centers are now measured in GW. And because semiconductors themselves have hit scaling limits, the only way to expand is to connect a very large number of processors and use them as one computer — making the network the layer that sets the compute ceiling.

3.2 Four Irreplaceable Properties of Light

Yuen distilled the value of optics into four things, each mapping directly to an AI pain point:

Ultra-low latency: light communicates at the speed of light, with latency in the nanosecond range.
Enormous bandwidth: a single fiber can carry on the order of 30 Tbps across roughly three wavelength bands — hundreds of times today's copper cables.
Scalable connectivity dimensions: optics can add connections not only by “adding fibers,” but also open another dimension through wavelength.
Power and reliability: with CPO, for example, interconnect power can drop by about 70%; and the data he cited says CPO reliability is already at least 20x that of pluggable modules (a speaker citation, not public standardized data — rely on each vendor's own measurements).

He added a line that lands hard in AI data centers: every watt not spent on compute is a lost opportunity. That is the real business case for technologies like CPO — what you save isn't the electricity bill, it's power budget you can convert into compute.

3.3 What Happened to Copper

Yuen gave minimal but devastating arithmetic: a data center rack is about 2.2 meters tall, and connecting a server at the top to a switch in the middle takes roughly one meter of copper cable. At the 1.6T generation that's barely achievable; go higher, and even if the distance is just within reach, you burn so much power in DSPs that it defeats the purpose of using copper.

3.4 Four Layers of Scale — and the Year 2028

This was the most information-dense section of the day. Yuen divided the optical battlefield into four layers and gave the relative magnitude of each:

Scale across (between data centers): optics' home turf, an extension of telecom technology. Baseline 1x.
Scale out (within the data center): took shape around 2023, 10x. This is why the entire optical module industry is sprinting today.
Scale up (in-rack and across rows): runs at NVLink-class rates rather than Ethernet, 100x. This is the leap that hasn't happened yet.
Scale in (in-package die-to-die): using optics to handle I/O expansion inside the XPU, another 10x in the future.

Then he gave the timing: There is broad industry consensus that 2028 will be the first year of optical scale-up; 2030 will be full optical scale-up.

He also mentioned XPU bandwidth growth — roughly doubling every two years, or about 40% per year — and that optical circuit switching (OCS) is evolving from large-scale scale-out applications toward scale-up scenarios, which are smaller in scale but larger in connectivity dimensions.

We broke down the divergent paths for scale-up optical interconnect in After Copper Runs Out for AI: Seven Paths for Scale-Up Optical Interconnect, and Two Ways to Survive Each Bottleneck; the corresponding map from the official-spec side is in OIF Draws an Official Map for AI Interconnect: Three Networks, One pJ/bit Battlefield, and CPO as the Written Endgame.

3.5 The Real Problem: How to Build Volume

Yuen ended by turning to what he sees as the industry's biggest challenge — not technology, but capacity ramp:

Optics spent twenty to twenty-five years in the wilderness. Nobody knew how to make it go fast. Now we suddenly have to ramp capacity tenfold in a year.

He used his own company as an example: Lumentum grew laser capacity about 12x in two years, and its module products have been ramping every quarter. He said the answer is to learn from the silicon industry — which is also why he believes CPO is so powerful: the essence of CPO is moving optics from “the optics way” to “the silicon way” — from components built one by one in a lab to a high-yield, high-throughput manufacturing flow that can ramp on demand.

He explicitly called out Taiwan: Taiwan understands what scale means, how to handle volume, and how to get quality right quickly.

STT Take: The value of this keynote is that it replaced the empty phrase “optics matters” with two checkable items — a 100x jump in magnitude and the year 2028. Over the next two years, if optical scale-up orders don't start showing up in supply-chain book-to-bill by the second half of 2027, that year will have to be pushed out. Lumentum's own financials have already started reflecting this line; see Earnings Highlights: Lumentum (LITE) | FY2026 Q4.

4. Keynote 2 | Avalanche Thinking's Chuck Mattera: The Century of Photonics, and “Interconnect Becomes Architecture”

Chuck Mattera is the former CEO of Coherent and now founder of Avalanche Thinking. His style was the opposite of Yuen's — he didn't talk about the next product cycle; he talked about the direction of a century.

4.1 A Seventy-Five-Year Chain of Accumulation

He started in 1947: that year the transistor gave us semiconductor amplification and switching, and a few months later Claude Shannon gave us information theory. Electronics and information were born together. Then light joined: the laser in 1960, the semiconductor laser two years later (light that could be made from III-V materials), and low-loss fiber in 1970 (light that could cross a continent). By 1980, optical communications had only just become a business.

He wanted the audience to notice the pattern: discovery becomes capability, and capability becomes infrastructure. The internet in the '90s and the cloud in the 2000s — every layer was built on a photonic foundation.

Then came ChatGPT at the end of 2022, which he said did something different:

It made the movement of information the constraint. Not storage, not processing — movement.

4.2 How a Discovery Becomes Infrastructure

This was the methodological core of his talk: it happens through the accumulation of capabilities — materials, components, subsystems, systems, growing layer by layer. Each layer must deliver differentiated functionality and must integrate systematically with the layers above and below.

A discovery by itself only creates possibility. But possibility doesn't change the world until engineers can design with it, manufacturers can scale it, customers can depend on it, and markets can organize efficiently around it.

He offered two halves of his own career as evidence: the first twenty years at Bell Labs, watching this accumulation happen around him; the next twenty at II-VI and Coherent, living it again firsthand. II-VI started in 1971 as a materials company and spent fifty-plus years on hard problems — materials became components, components became subsystems, subsystems became systems. His conclusion: when you've been accumulating for fifty years, the accumulation itself is the competitive advantage.

4.3 AI Is an Infrastructure Revolution, and Connectivity Is the Fifth Peer Subsystem

Mattera said AI isn't just a software transformation; it's an infrastructure revolution — easy to forget, because most people experience AI as a prompt on a screen. But behind it is a giant physical machine, and the unit of compute is no longer just a chip, server, or rack; increasingly it's a building, sometimes an entire campus.

And inside that machine, five systems are now peers: compute, memory, power, cooling, and connectivity.

When compute reaches the scale of a building, interconnect becomes the architecture itself.

He then traced how the constraining resource in computing has moved over time: in the '90s we optimized clock speed until heat stopped us; in the 2000s we pushed transistor density; in the 2010s memory bandwidth became the pressure point; then packaging and data movement tightened. Every time the industry hit a bottleneck, the bottleneck moved forward. Now the bottleneck has moved to interconnect — power per bit, high bandwidth, low latency, and distance all have to be solved at once.

Then came the key correction of his talk:

This is not a war between electrons and photons. Electrons compute; photons move. The future is the two engineered together.

4.4 Four Architectures, and Four Signals That “the Transition Is Already Underway”

He made Light at Scale concrete with four architectures, stating explicitly that they are stacked generations:

Pluggables (front-panel pluggable): power today's AI deployments and will keep scaling.
Near-package optics (NPO): serves as the bridge.
Co-packaged optics (CPO): delivers integration.
Optical I/O / photonic fabrics: pushes the boundary further into compute, ultimately making light the connective tissue of the whole machine.

These generations will layer on top of each other; they won't simply replace one another.

Four signals (all market data cited by the speaker, not independently verified by STT):

Bandwidth: 1.6T optics are ramping in volume; 3.2T transceivers with 400G optical lanes have already been demonstrated.
Pluggables are still growing: Ethernet optics grew about 80% last year, and analysts expect roughly another 65% this year.
CPO has left the lab: platforms are shipping, and 102.4T systems have entered the market.
Power is the reason: moving optics next to the ASIC cuts interconnect power by about 70% versus conventional pluggable approaches.

He also gave a monetary scale: the backdrop for understanding how this transition will unfold is an estimated ~US$7 trillion of AI infrastructure investment between now and 2030.

4.5 The Collapse of Distance

This was his most beautifully described section. For decades, fiber solved long-distance problems — across oceans, continents, cities, and data centers. Then the same technology was asked to move data one kilometer, ten meters, one meter, one centimeter, and ultimately one millimeter.

Distance is collapsing, bandwidth is multiplying, and device physics is being pushed to its limits — it starts taxing the materials. Copper works extremely well over short distances, but as bandwidth rises, distance gets expensive, power gets expensive, heat becomes intolerable, and latency becomes unmanageable. That's why light keeps moving closer to compute.

For fifty years we brought compute into the network. Now the network is moving into compute, and compute is moving into optics.

Finally, he turned to FIT: interconnect is no longer just a passive component at the edge of the system; it is becoming part of the machine's architecture. Very few companies combine electrical, optical, power, and thermal design capabilities with long-term manufacturing scale — and he believes FIT now does. He said he sees in FIT what he saw when he walked into Bell Labs in 1984: engineering and manufacturing capability that is accumulating, enough to drive an entire technology transition.

STT Take: Mattera's talk had the highest density of ideas all day, but be clear about its position — a senior industry leader, on the host's home turf, publicly endorsing the host. The real takeaway is his methodology: discovery creates possibility; manufacturing creates scale. That thread connects directly to the storyline we've been following in [CPO Teardown 5/6] The CPO Ecosystem Is Being Reshuffled: Foundries and OSATs Move from Supporting Roles to Lead: the winners of this redistribution won't be those with the prettiest roadmaps, but those with the deepest accumulated manufacturing capability.

5. Keynote 3 | NTT's Yoshiaki Sato: Three Generations of Photonic-Electronic Convergence, and FIT's 500 Pins

The third keynote was delivered in Japanese by Yoshiaki Sato, Executive Adviser at NTT Innovative Devices. It was the only talk of the day about products already in volume production, with real hardware brought on stage — and the most substantive.

5.1 Why a Telecom Operator Builds Its Own Components

Sato first explained what makes NTT unusual: NTT is a telecom operator in the same category as Chunghwa Telecom and AT&T, not originally a manufacturer; but NTT Innovative Devices is a manufacturer, making both digital ICs (fabless — designed in-house, fabricated by a foundry, ultimately TSMC) and optical components (designed, developed, and manufactured in its own fabs).

A company that provides telecom services also owning a manufacturing business is a rather rare model.

The company is three years old, but counting its predecessors, it carries 40 years of optical communications innovation. He walked through three milestones: in the 1980s, to push a single fiber past 1 Gbps, they built GaAs compound-semiconductor high-speed ICs; in the 1990s, as the internet drove explosive demand and telecom had to break tens of Gbps, they built optical multiplexers for wavelength-division multiplexing (WDM); later, to break 100 Gbps, a new transmission method was needed, so they built digital ICs for digital coherent.

Having both digital ICs and optical components — these two faces are what give them the confidence to bet on photonic-electronic convergence.

5.2 A Chart Spanning 30 Centimeters That Set the Direction of the Whole Company

Sato showed a chart he described as “only 30 centimeters on the horizontal axis”: the power-consumption curves of optical wiring versus electrical wiring over a few tens of centimeters.

The blue line is electrical wiring: as distance grows, power rises sharply; push the signal rate higher, and it jumps up another step.
The yellow line is optics: we use fiber over hundreds of kilometers in telecom, and using light over these 30 centimeters, power barely rises with bit rate.

He tied AI's power problem to this chart: cutting-edge GPUs and CPUs are extremely dense, with large numbers of I/O interfaces, and interface speeds keep rising — so power inevitably rises. He cited NVIDIA's just-announced Rubin: “excellent performance, but power as high as 2,300 W — 2,300 W for a single IC.” (This matches the widely reported 2,300 W TDP for Rubin Ultra; it is the speaker's on-site citation.) He also cited research-firm figures: by 2030, data center power consumption will be double that of 2025.

He then drew GPU history as two lines: from Volta to Blackwell, performance rises and power rises with it; the yellow line isolates just the GPU's interface portion — a large chunk of power growth comes from “more interfaces, and faster interfaces.”

Replace this portion with optics, and GPU power and XPU power will come down.

5.3 Three Generations — Deployed in Reverse Order

This is NTT's most distinctive strategy. In the GPU world, the usual progression is scale up → scale out → scale across; but because optics was originally a long-distance technology, NTT deploys in reverse, starting from the farthest layer.

Generation 1 (PEC-1): Between data centers

The problem is clearly defined: the per-fiber capacity needed between data centers is moving from 400G (already in use) to 800G and 1.6T, and building 800G the conventional way is extremely difficult. The reason is in the cross-section: a typical transceiver interconnects multiple components on a PCB, and the moment a signal goes onto the board it incurs loss — you can't reach 800G the ordinary way.

NTT's approach is to integrate silicon semiconductor devices and compound-semiconductor optical devices into a single chip: it looks like an ordinary IC with many contacts, but also carries a fiber outlet. He described it as “a new type of chip with both optical and electrical inputs and outputs.”

He compared two photos on stage: on the left, a typical 100G transmission unit from a partner, with a transceiver measuring about 4×5 inches; on the right, a pluggable module the same company built using NTT's first-generation photonic-electronic convergence device — reaching 800G while lowering power.

Generation 2 (PEC-2): Inside the data center

A typical electrical switch places a switch chip at the center of a large board, connected to the front panel via roughly 30 cm of electrical interface. Power consumption is very high. NTT's approach moves high-capacity optical interfaces right next to the switch chip. He brought the actual hardware on stage:

At the center is a Broadcom Tomahawk-series switch IC. Around it are 16 optical interfaces, each delivering 6.4 Tbps of high-capacity communication.

16 × 6.4T = 102.4T — the same thing Mattera meant by “102.4T systems have entered the market.” The effect: overall power about 50% lower than simply keeping electrical interfaces. He said the product is being commercialized this year.

Then came the most important line of the talk: “mou hitotsu houkoku ga arimasu” (“I have one more thing to report”).

This optical interface can now be made detachable. FIT helped us develop this technology together. The interface has 500 pins, and each pin can carry 200G.

That sentence carries a lot of information. CPO's biggest commercial obstacle has always been serviceability — optical engines soldered next to the switch chip mean one failure requires replacing the whole module. “Detachable” attacks this pain point directly. And the company that made it happen is a connector maker, not an optical module maker. He also noted that beyond connectors, the two sides have multiple joint developments underway, including external laser sources.

Generation 3 (PEC-3): Into the package

Generations 1 and 2 both combine “digital IC/analog IC + optical components,” and the optical components contain lasers. To fit this whole package inside the IC package, existing laser structures basically can't do it, so the architecture has to change fundamentally. NTT's answer is the membrane laser — extremely small, with extremely low power — fused with electronic devices into an optical chiplet connecting digital ICs. Still in R&D.

5.4 An Easily Skipped Ending: The Osaka Expo Performance

Sato spent his last minute on a case with no specs: at last year's Osaka Expo, the NTT Pavilion and Taiwan's Chunghwa Telecom were connected over an optical network (with first- and second-generation photonic-electronic convergence devices already in the network), and live performers at both sites performed simultaneously, exchanging video.

Most of the audience watching in Taiwan probably didn't realize the images in front of them were coming from faraway Japan. Latency was so low that the performers could talk to each other directly, with nothing looking unnatural.

STT Take: This was the most valuable talk of the day — not because the technology was novel, but because it nailed the timeline to the ground: Gen 1 is commercial, Gen 2 commercializes this year, Gen 3 is still in R&D. Compared with Yuen's “2028 as year one of optical scale-up,” NTT's sequence is arguably more honest: optics first establishes itself over long distances, then moves inward step by step. We covered NTT's full scale-across strategy in NTT Unveils AICC at OCP: AI's Next Bottleneck Isn't Compute, It's Connecting Distributed Compute with Light.

6. Keynote 4 | ams OSRAM's Ashkan Seyedi: The Data Center You're Talking About Is Yesterday's World

Ashkan Seyedi opened in a posture both humble and provocative:

Going last comes with a special honor and a special challenge. The previous speakers made you all smarter; I'll do my best to mess that up.

Then he delivered the biggest frame shift of the day:

So far everything we've discussed has been the data center. But let me tell you, in my world that is yesterday's world. Cloud training and cloud inference have been proven to work; the world will now spend five, ten, fifteen years supplying that ecosystem.

6.1 Physical AI and Personal AI

He believes the real impact lies in two places.

Physical AI: hands, arms, eyes, wheels. He cited an investor friend — of the three to four hundred companies that listed in China this year, roughly 30–40% are actuator companies (on-site citation, not independently verified by STT). Once these things take shape, intelligence will be unlocked from the data center, and tokens can be generated in the robot next to you, in your car, in the drone above your house. He also noted that Hon Hai itself is bringing this kind of automation into its factories.

The conclusion: you'll need a large number of micro data centers — small server rooms of ten or twenty racks, sited next to factories packed with robots, doing “local thinking, remote orchestration.”

Personal AI: he ran a very handy calculation. There are four to five billion phones in the world, each in roughly the 25 W class — the global phone population is itself a GW-class data center, and it's already installed. The only problems are memory bandwidth and compute bandwidth, and they aren't interconnected. But in this world, they don't all need to be.

The price is energy efficiency. He gave a very concrete challenge figure:

Today it's roughly 100 to 200 millijoules per token. In a data center server drawing thousands of watts, that's fine. But on your watch or glasses, at that rate the battery is gone in twenty minutes. We need another 100x reduction in joules per token.

He noted this is a memory wall problem, then half-jokingly said his talk strategy was to complain about his own problems and hope someone in the audience would come up and say “I have a solution.”

6.2 The Most Counterintuitive Line of the Day: Interconnect Isn't the Data Center's Big Power Hog

This line deserves to be pulled out on its own, because it directly challenges the main selling point of the previous two keynotes:

Everyone talks about optics and CPO cutting power. I've worked in CPO for over ten years, and people always joke that it's forever “two years away.” But — interconnect is not suffering. Thanks to forty years of work in telecom, interconnect is not the biggest power consumer in the data center; compute and memory are. Optics accounts for only a single-digit percentage. What really kills you are reliability problems, the latency added by forward error correction (FEC), and things breaking.

And when physical AI arrives, he said, these pressures will only grow.

This was the most important correction of the day. If optics accounts for only a single-digit percentage of data center power, then the system-level benefit of “CPO saves 70% of interconnect power” gets compressed — what can really be sold is reliability, latency, and density, not the electricity bill.

6.3 The Scale-Up Domain Is Getting Crowded

He showed a chart of a scale-up rack row: two power racks, six compute racks, one LPU rack, one context-memory storage rack, and two network switch racks. The key was the text beside it: larger scale-up domains will span multiple rows of racks and need ultra-low-latency, power-efficient interconnect; AI inference has entered the era of heterogeneous compute (GPU/LPU/TPU/xPU + context memory + scale-up & scale-out networking); and today's single-mode optics don't meet scale-up's requirements for power, cost, and latency.

He also gave a concrete number: in NVIDIA's reference architecture for 120,000 GPUs, about 60% of interconnects occur within the same row. The more accelerators you pack in, the tighter this demand for connections within about 100 meters becomes.

6.4 Wide and Slow: Pushing the Problem onto Mechanics

Seyedi's technical position is wide-and-slow — many lanes at low rates. He explained how we got here:

The reason serializers exist, and the reason rates keep going up, is metallurgy. Microbumps haven't gotten smaller. Bump pitches of 30 microns, 40 microns — that's where we've been stuck for the past decade. So how did chips get faster? Serializers. And once you're in the PAM4 world, you have to make decisions, you need hard specs, and that only adds latency.

So when you go wide and slow, you're pushing this problem onto mechanics. You have to go multi-fiber, multicore fiber, and it becomes an alignment problem, a precision mechanical engineering problem — which is exactly what Hon Hai is very, very good at.

What you get in return is cost savings, latency savings, and of course power savings.

He also spoke to his company's capacity: ams OSRAM does InP, GaAs, and GaN, and interconnect based on GaAs VCSELs and GaN LEDs is consumer-grade.

When you're talking ten million, fifteen million GPUs a year — and by GPU I mean any accelerator that needs optics for scale-up — we have that capacity. Today we're on 8-inch, and if needed, compound semiconductors can move to 12-inch.

6.5 Why NPO Deploys First

This was the most practical slide of the day. He labeled the electrical channel length of three form factors directly: OSFP about 20 cm or more, NPO about 5 to 10 cm, CPO/OIO under 1 cm.

His argument was clean enough to memorize:

OSFP: a typical AI compute tray is only about 0.9RU yet must accommodate 300+ scale-up data lanes, and scale-up IO is often at the back of the rack. The OSFP form factor will be tight on serviceability and density for scale-up, but can support NIC-L1 and in-rack links.
NPO: shortening the electrical channel yields a significant power reduction; no new connector is needed as SerDes speeds rise from 200G to 400G and 800G; and optics stay outside the XPU/HBM failure domain — an optical chip failure doesn't drag memory down with it; the standard footprint also preserves the option of mixed-media IO later.
CPO/OIO: the ideal form factor for power and signal integrity, but it brings significant packaging challenges, and ever-growing interposer size is itself a threat.

His own conclusion was politically correct, but also honest:

CPO clearly has its benefits and challenges, and NPO is getting a lot of attention, deservedly. Someone has to go out and build the ecosystem, and then stand up the entire supply chain.

He also pointed out the least sexy detail that's holding everyone back: vertical fiber attach — actively aligning 10, 20, 30 optical contacts is a very complex problem; but once you learn it on one form factor, you can apply it across all form factors. This is exactly what we covered in [CPO Teardown 4/6] Fiber-to-Chip: The Least Sexy Step in CPO That's Holding Everyone Back.

Finally, he described how the conversation has changed, in a vivid line:

Two years ago, if you said CPO, people would say “don't talk to me about that.” Now if you say integrated optics, they ask: which kind? CPO? NPO? InP? Silicon photonics? TFLN?

STT Take: Seyedi's talk was the most valuable “dissent” of the day. He did two things useful for readers: first, he downgraded CPO's power-saving pitch (optics is only a single-digit percentage of data center power), shifting focus to reliability and latency; second, he framed NPO as today's best compromise, for reasons that are entirely engineering and supply-chain (no connector change, failure-domain isolation, preserving mixed-IO options), not performance. We quantified the thermal and system costs of the NPO path in 51.2T NPO Switch Thermals: 835W ASIC Plus 16 Optical Engines — Air or Liquid Cooling?.

7. Panel | How Optical Modules and CPO Are Moving into AI Data Centers

The afternoon panel was moderated by Dr. Thomas Liu, FIT's optical communications consultant. Its value was entirely different from the keynotes — the keynotes were each company's position; the panel was where positions collided.

7.1 Question One: How Is This Different from the 2000 Bubble?

The moderator put this to Wei-Ping Huang — someone with forty years in optoelectronics research, teaching, entrepreneurship, and management, who lived through the 1999–2000 telecom bubble firsthand.

He first laid out the history: that bubble was fueled by a belief — internet traffic would grow exponentially. Carriers laid fiber extremely aggressively, and investors believed it too. Then after 2001 the industry collapsed under fiber oversupply and failed forecasts; a dozen or so companies went bankrupt, Nasdaq fell back from 5,000, and he said the industry took about fifteen years to truly recover.

A scary story, right? Well, yes and no.

He founded Ligent (Light for Intelligence) around 2002, the darkest period for optical communications. He said he believed two things then: light is the best medium for information, and we need optical-electrical conversion. He formed a joint venture with Hisense (a TV brand with almost zero knowledge of the field) because its chairman was his college roommate — “in that dark moment, that was the only combination that worked for a startup.”

So why is this time different? His answer:

This time is different because AI really is a revolution in productivity and economic growth, and an upgrade in consumer experience will be layered on top. Leading AI service providers have already proven the business model with astonishing revenue and profit growth. We haven't yet seen AI's impact on the consumer side, but when AI applications reach end users, it will happen — a consumer revolution possibly bigger than the internet era.

Hong Hou followed with a quantified version of agreement, and the contrast was razor-sharp:

Back then, the scale tier of that boom and bust (we didn't have this term then) was roughly 1x. At that tier, fluctuations make a very significant difference. But now, the traffic through a single switch already exceeds total north-south traffic across the Atlantic and Pacific. The scale is enormous. Scale out is already 10x, scale up will be 100x. So this time will be different — there will still be fluctuations year to year, of course, but the base is much higher.

STT Take: The value of this question is that it turned “this time is different” from faith into a structural argument. The two gave completely different reasons — Wei-Ping Huang on the reality of demand (the business model is proven), Hong Hou on the absolute scale (the base is too high for fluctuations to punch through). Both arguments hold, but neither answers one question: a high base doesn't mean healthy cash flow. Wei-Ping Huang addressed this himself in the next question.

7.2 Joys and Worries: Wei-Ping Huang's Three Concerns

The moderator noted that Wei-Ping Huang had spoken about “joys and worries” at CIOE in Shenzhen the previous week and asked him to reprise it. The slide itself is worth saving.

The joys he covered quickly, because “everyone here feels the same.” But wearing his professor's hat, he added a beautiful line:

To make a networked computer work like one computer, you need full connectivity, nearly unlimited high bandwidth, and extremely low latency. We're still far from that goal. Our ceiling is far above us.

Then he said: so I'm happy — I have at least another twenty years to work in this field, purely for fun.

The worries come in three parts:

Capital market patience: he worries about the financial structure behind the AI narratives and massive investments. The slide was more direct: leading AI service providers are in a fragile position of negative cash flow and huge debt. In his own words: when you hand control to market sentiment and the patience of capital, that may or may not be a problem.
Geopolitics: he gave a set of numbers (speaker citation) — US vendors supply about 95% of high-end electronic chips and about 90% of photonic chips, while Chinese vendors make about 65% to 75% of optical transmitters. This was a very elegant, healthy, efficient global supply chain; but under political pressure, things changed: the US is trying to bring manufacturing back onshore, and China is investing heavily in semiconductors to meet domestic demand. His conclusion was cold: we may end up with two separate ecosystems supporting two competing AI ecosystems. I don't know if that's ideal, but we'll have to learn to live with it.
The shift from pluggables to NPO/CPO is not a simple substitution: he said it will bring significant technical challenges as well as industry-ecosystem challenges. His analysis is worth reading verbatim:

Looking at CPO in the extreme: high resolution, high power, high bandwidth density, and these are usually handled by CMOS. Going to higher speeds, on the electronics side you may need different materials. We just talked about all those fundamental limits.

And NPO looks like a good balance: it can be handled by ordinary transceiver manufacturers, it can accommodate future upgrades to more advanced optoelectronic chips, and it saves power and captures the bandwidth-density benefits.

These are two technical solutions, but also two very different value propositions. The first is CPO — a bundled solution: all the transceiver players get moved out of the package, and customers must accept a bundled offering. NPO may be different — it can be independent and unbundled, but maybe gives up some of the nice properties.

Asked by the moderator to sum up in two sentences, he said:

First, my joys are well-founded; my worries may be biased or exaggerated, I have to admit that. Whether it's a bubble depends on the patience of capital markets; AI will be real, it just takes time, and some turbulence. The economic scale is huge, so we have to be on the winning side, not the losing side. People often say optimists are usually right, but truly successful people must be optimists. I am an optimist.

STT Take: The read that “CPO is a bundled solution” was the most underrated line of the event. It isn't a technical judgment; it's a judgment about industry structure — once CPO becomes mainstream, optical module makers get squeezed out of the package, and customers buy bundled solutions from switch or chip vendors; NPO preserves room for “independent third-party supply.” That explains why players in different positions show such different levels of enthusiasm for NPO versus CPO. We've followed this structural thread in [CPO Teardown 6/6] Every CPO Roadmap Reviewed: Beyond the Big Four, a Whole Row of Challengers and CPO Is Finally No Longer “Crying Wolf”: Six Real Signals from LightCounting's CPO/NPO Event.

7.3 Connectivity Options: Semtech Wants to Make Analog Great Again

The moderator asked Hong Hou what connectivity options can simultaneously deliver high bandwidth, high bandwidth density, low power, and low latency.

Hong Hou first did something important — he laid out all the technology paths and refused to pick a side:

From EML, to CW lasers with silicon photonics modulators, to VCSELs, to micro VCSELs, to micro LEDs — each has its pros and cons. Take the micro LED that ams OSRAM talked about: it borrows the display industry's ecosystem, so you never have to worry about capacity constraints, which is good news. Of course it has its own challenges — coupling, reach, and it needs ecosystem support such as drivers, TIAs, and gearboxes. But that's the beauty of an ecosystem. Every technology will have its place.

Then he laid out Semtech's position, with the most memorable line of the event:

We specialize in analog and mixed-signal. If you don't take the DSP route — analog-to-digital, then DSP, then digital-to-analog — you save a lot of power and reduce latency. So we made a joke for ourselves: Make Analog Great Again.

He added that O-band and C-band DFBs, VCSELs, and micro LEDs all have their own places — different reaches, different data rates, different requirements.

MACOM's Ernest Muhigana then put this in the context of AI infrastructure scale: the volume of data exchanged inside data centers is enormous — east-west, across halls, across campuses, then scale across to other data center clusters. He cited a very concrete new pattern: large hyperscalers now connect out to customer-side operators to provide services — a customer trains a model in one hyperscaler's data center, then runs inference in another data center. This creates massive interconnect demand.

7.4 Fast-and-Narrow vs. Wide-and-Slow: No Winner Yet

The moderator put Seyedi's position on the table: these are two different design philosophies, so what's common in IC development strategy?

Hong Hou's answer was a cost list: with wide and slow, every lane still needs a driver and a TIA for optical-electrical and electrical-optical conversion; on top of that you need a gearbox to convert high-speed signals into multiple low-speed lanes. You pay some power overhead in the electronics, but energy efficiency on the optical side is much higher. So it depends on the overall system, and system architects must do extensive trade-off studies to decide which option fits which application.

Ernest Muhigana's answer was more direct:

I don't think there's a clear winner right now. Because of the rush to deploy infrastructure, everyone is using existing technology — 100G, 200G, 400G — so it started with fast and narrow. But slow and wide does have its value, especially in enabling very low bit rates and robust NRZ communication.

But slow and wide still has to be built and proven. It's many fibers in parallel, so the fiber part isn't trivial. Also from an architecture standpoint: fast and narrow is switched in some cases, or is typically a point-to-point interface between XPUs. What about slow and wide? You'd have to introduce more boxes. But what would the switching architecture become? And if it's not switched, would it be the best way to move data between GPUs and TPUs?

His conclusion was honest: fast and narrow is there because the technology is there and proven; slow and wide still has to be built, but it will arrive with its benefits.

Wei-Ping Huang's addition had the most historical depth of the panel:

I may have a bit of a bias from the optics community. In the past, so-called slow and wide was the inferior solution. The best example is DWDM — stacking 10G, 10G like that. For optics, using this kind of parallelism to create enormous bandwidth was preferable.

I didn't expect us to go from 1G, 5G all the way to 200G or even 400G per link. But now I understand it was purely the computing industry dreaming — because service demand kept going up. I used to think 400G would be the top, and after that we'd get wider and wider.

But now I'm a bit surprised to see everyone suddenly going back to very slow and very wide. That's genuinely a bit unexpected for me. I think there will be a balance somewhere in between — not necessarily micro OE; maybe larger OE will be a good balance.

He closed with an optics person's stance: he really wants to see photonic integration move toward higher density and more functions, designed with a bit more freedom, rather than being led purely by existing computing architectures.

7.5 Don't Bury Copper Too Soon

The moderator raised scale across — from 10 km to 1,000 km, connecting campus AI and metro AI to backbone sites. It's reminiscent of traditional long-haul transmission, but with much higher data rates and bandwidth, so supply and security of coherent optics and electronics will be challenging.

Muhigana's answer pinned down the definition of scale across: optical transport networks were traditionally built for telecom, serving the internet and cloud services; but the term scale across is mainly about interconnecting different clusters — TPU, GPU, and XPU clusters within large campuses, and across campuses. You're connecting accelerators over distances of roughly 20 to 30 km.

You can think of it as one big computer with huge amounts of data moving inside it. One interface option is coherent; another is ZR lite. So it's no longer ZR, but a lower-cost, lower-complexity solution closer to the PAM4 interfaces we know well.

He then offered a magnitude estimate well worth tracking (speaker's estimate):

If you look at how many coherent lite interfaces scale-across networks will use, plus the coherent lite interfaces that will also be used inside data centers, the volume will be on the order of ten times the ZR we know. So we have to produce very small, very energy-efficient interfaces in enormous quantities.

Then Hong Hou delivered the most contrarian passage of the event:

I don't want to be a party pooper. The theme is light at scale, but when you scale up, you can also consider scaling up with copper inside the rack. Rack height is about 2.2 meters, but if you can use redrivers or retimers to extend copper cable reach to three meters, you can effectively complete scale up within the rack. A combination of copper plus fiber may well be the best solution inside the rack.

For example, we offer linear equalizers that can amplify different levels in the frequency domain. The result is you can improve signal integrity and raise gain by 15 to 20 dB, which may be enough to make up for attenuation and signal distortion. Imagine — even when a high-speed signal passes through a connector from a big connector maker like Foxconn, not only do you not lose power, if the equalizer is embedded in the connector, you might even gain some.

So it's a combination. Copper hasn't reached the end. We tried to kill copper about twenty years ago, but copper has kept going strong. Don't ignore copper. Treat copper as an extension of fiber.

He also gave his company's product status: at current transmission rates such as 224 Gbps, there are already products supporting 2.4T cables, and they are co-developing 448 Gbps with major customers; meanwhile PCB materials are also improving (M8, M9 grades lowering loss). His judgment: even at 448 Gbps, equalizers can still play a key role in improving signal integrity, perhaps enough at the board level.

Wei-Ping Huang's follow-up cut even deeper:

In the fiber-to-the-home era, we said fiber advances and copper retreats. That may have been right. But in the computing era, my view is that copper has to retreat first before optics can advance. Maybe it succeeds, maybe not. If you look at the figures of merit — bandwidth density, energy efficiency — at meter-scale distances, copper still leads optics by one order of magnitude, maybe two. CPO can indeed close some of the gap, but the gap is still there. The competition here really depends on how smart and advanced we can make optical technology.

Muhigana closed with an old industry saying: use copper where you can, use optics where you must. But he added that the demand and build-out scale of AI infrastructure is very large; accelerators within a rack can still be connected with copper, with equalizers and other component solutions; but once you leave the rack, you definitely need another solution that gives you extra reach, and that has to be optics.

STT Take: This was the most valuable part of the panel because it produced genuine disagreement. The three positions line up on a spectrum: Muhigana says “use optics where optics is needed,” Hong Hou says “copper can stretch to three more meters, don't forget equalizers,” and Wei-Ping Huang says “copper still leads by one to two orders of magnitude at meter scale; optics has to get smarter before it can come in.” For the supply chain, this means the lifecycle of ACC/AEC and redrivers/retimers may be longer than the optics-camp narrative implies. This matches the judgment SemiAnalysis made at OCP, which we broke down in Scale Up Sophistry: Copper vs. Optics Is a False Dichotomy.

7.6 Hon Hai Research Institute: Betting on InP and Quantum Dots at the Same Time

The moderator handed the floor to Hao-Chung Kuo, first announcing his team's results: Hon Hai Research Institute's team co-developed modules with FIT, with a live demo on site.

Hao-Chung Kuo presented two tracks.

Track one: InP platform, 300 Gbps+ lanes

They are advancing an InP platform with SMART Photonics (foundry) and TU/e (Eindhoven University of Technology) via multi-project wafers (MPW). The reason for choosing InP is simple: good material properties and high-speed capability. Results: cutoff frequency above 110 GHz, with 320 Gbps PAM4 validated as feasible by TU/e; a 300 Gbps-class demo is in progress. On the driver side they work with Semtech, on the detector side with MACOM — MACOM's TIA integrates very well with its InP detectors.

He also told an old joke from the compound-semiconductor world:

I started doing InP epitaxy in 1996. Back then everyone said compound semiconductors were the semiconductors of the future, but they were always “the future”. Now it's really happening.

Track two: Quantum-dot comb laser × multicore fiber, 34.132 Tbps per fiber

This was the most impressive number of the event. The motivation is very practical: InP substrates may be in short supply, while quantum dots can be grown on gallium arsenide (GaAs), and GaAs is very mature at 6-inch and also good at 8-inch.

They used quantum-dot lasers to generate multiple wavelengths (a DWDM-like concept), combined with multicore fiber obtained from the University of Southampton in the UK, with these results:

212 Gbit/s PAM4 × 23 comb lines × 7 cores = 34.132 Tbit/s per fiber
After 2 km of single-mode fiber transmission, average TDECQ below 2.02 dB
Laser transmitter power about 0.76 pJ/bit
Published in Optics Express 34(13), 2026, pp. 24318–24328

Muhigana asked which side of the network this would be used on. Hao-Chung Kuo answered that transmission experiments at 2 km still gave very good results, and they believe it will be very helpful for future Optical I/O or scale-up applications — but he was candid: there is currently no standard for this, and no standards body is overseeing it.

Hong Hou followed up with a more commercial question: he mentioned recent news — a Santa Barbara quantum-dot laser company reached an agreement with an epitaxy supplier, claiming quantum-dot lasers are ready for commercial use; everyone is excited because quantum-dot lasers don't need an isolator, which can greatly reduce optical packaging and coupling complexity. So how far is commercialization?

Hao-Chung Kuo's answer was conservative, and that conservatism is itself important information:

Quantum-dot lasers still have reliability and integrability issues. And in the end it all depends on whether the system side adopts them. We think it will take another three to five years to become relatively widespread. But for specific applications, maybe usable within two to three years.

(STT fact-check: the news likely refers to the purchase agreement between quantum-dot laser company Quintessent and epitaxy supplier IQE; IQE announced in September 2026 that the quantum-dot laser technology had entered customer sampling, and Quintessent closed an oversubscribed US$40 million Series A in August 2026. This is consistent with Hao-Chung Kuo's “two to three years for specific applications.”)

Wei-Ping Huang then asked a question only someone who has done epitaxy would ask: you've built a great comb laser, which I'd guess is based on the inhomogeneous linewidth broadening of quantum dots — different dot sizes correspond to different colors. But if we want a single-wavelength quantum-dot laser, we need very uniform dots. How well can you control dot size now?

Hao-Chung Kuo's answer included a very human anecdote: quantum dots have been a research topic for more than twenty years; as a graduate student he grew lots of quantum dots, but they all went into a drawer because his advisor Milton Feng would scold him, “you're making ugly stuff” — too much strain, island structures, poor surfaces. But he believes quantum dots have unique properties and can be grown on GaAs, and now some people even grow them on silicon (though that's a very long-term goal).

He also made an important distinction: comb sources aren't exclusive to quantum dots — infrared materials can do it too, or you can use a DWDM multi-wavelength approach. He noted that friends from Lightmatter were in the audience, and Lightmatter is doing 16-channel DWDM — he believes multi-wavelength is the future direction, for both scale out and scale up.

7.7 The Road to 400G/lane: Who Can Get There

The moderator asked a question many wanted to ask: at CIOE and last year's shows, everyone was showing 400G modulators — silicon photonics, InP. Beyond the current solutions, what else will be commercialized?

Hao-Chung Kuo: coherent approaches all have a chance — silicon modulators, InP modulators, InP EMLs. But the best performance today is still InP EML, thanks to better integration; and some have shown very good TFLN results, so both have a chance. He flagged two key issues, though: first, 400G DSP is still a problem; second, the maturity of drivers and TIAs — which Semtech and MACOM can fill.

Muhigana responded immediately: drivers and TIAs are already there — from an analog IC interface standpoint, they're ready. On the modulator side, a hundred flowers are blooming: TFLN, InP, and silicon photonics are all being pushed.

Hong Hou gave a timing commitment and a technical judgment:

I agree. We won't let the ecosystem down; we will deliver 400G drivers and TIAs on schedule.

As for modulation: differentially driven EMLs have already demonstrated 400G performance, but you'll run into capacity constraints. So sometimes it's not just a technology question; it's about availability, value, and everything else. So we're trying to get silicon photonics modulators running at 400G — performance is still a bit short, but through co-optimization of driver and modulator we're confident we can get there. Then you can reuse the existing ecosystem and manufacturing infrastructure and ramp volume quickly. I think the other key components will be ready in about a year, and 400G projects are still a few years out, so the timing works.

Wei-Ping Huang poured some very valuable cold water:

I agree that all three solutions — silicon photonics, thin-film (TFLN), and InP — will eventually demonstrate reasonable performance. The first question is: technically, what price do you pay for each? The second question is capacity, and capacity has never been the scarce part; that's a different issue.

So in my view, if companies want to ramp volume at 400G, the ecosystem really needs to change significantly. We don't know how much investment is needed, or what price people are willing to pay.

I still believe in hybrid integration — like Intel integrating InP with silicon photonics, because InP has all the functions, not just the laser source but also high-speed modulation. But that brings us back to the foundry question. I attended a forum at OFC called “The Road to 400G,” but I still can't see a clear road, because the ecosystem isn't ready — except for silicon photonics.

STT Take: The signal from this question is clear: the 400G/lane bottleneck has shifted from the modulator to the DSP and ecosystem economics. Both driver/TIA vendors say “ready within a year,” all three modulator paths are “eventually feasible,” but no one can answer Wei-Ping Huang's two questions — what it costs, and who pays. MACOM's track record in data center and InP can be compared in Earnings Highlights: MACOM (MTSI) | FY2026 Q3; the materials and InP capacity side can be compared in Earnings Highlights: Coherent (COHR) | FY2026 Q4.

7.8 The Closing Line: Please Help Us Solve the Material Shortage

With under three minutes left in the panel and no audience questions, moderator Dr. Thomas Liu wrapped up himself. His words deserve to be quoted in full, because it was the only moment all day that someone stated the most practical problem directly:

Today we talked about AI, transceivers, even CPO. I look at it every day, and demand is very strong. So are the key materials. So to everyone here — I sincerely ask suppliers, manufacturers, semiconductor fabs, even component makers: please help us solve the material shortage.

I think that's the key right now. Technology is moving forward, whether TFLN or silicon photonics. But the real dilemma today is that we don't have the materials to build optics, to build AI infrastructure.

An event hosted by a connector and systems company, featuring the leading laser maker, Coherent's former CEO, NTT, ams OSRAM, Semtech, MACOM, and Hon Hai Research Institute, ended not with a verdict on which technology will win, but with “please give us materials.”

8. STT Insight: What This Conference Was Really Saying

Put the eight speakers' content together, and five things are worth taking away — two of which nobody said explicitly on stage.

8.1 “Coexistence” Is the Consensus, but Nobody Has Priced It

Mattera, Seyedi, Hong Hou, and Wei-Ping Huang all agreed that the four form factors will coexist. It sounds harmonious, but coexistence means: someone has to sustain all four ecosystems — pluggable, NPO, CPO, and OIO.

Each form factor needs its own connector specs, its own test methods, its own alignment processes, its own failure-analysis flows. These costs don't vanish just because “they'll coexist anyway” — they get spread across a market that has no standards yet.

Seyedi's line, “someone has to go out and build the ecosystem, and then stand up the entire supply chain,” sounded like thanks to his peers, but it states a harsh truth: before standards take shape, whoever funds the ecosystem bears the trial-and-error cost; after standards take shape, that cost may become someone else's moat — or someone else's free asset.

For Taiwanese companies, this is a strategic choice, not a yes-or-no question: whether to be the one building the ecosystem.

8.2 2028 Is the Only Checkable Date This Conference Left Behind

All day, only Wupen Yuen gave a year: 2028 is year one of optical scale-up, 2030 is full optical scale-up.

You can work backward from it: if it holds, 2H 2027 should see scale-up optical interconnect design freezes and pilot-production pull-ins — meaning orders for connectors, optical engines, alignment equipment, and multicore fiber should start showing up in book-to-bill and capex in 2027. If they don't, that year should be pushed out.

Notably, NTT's roadmap gives a sequence consistent with this year: Gen 1 (scale across) is commercial, Gen 2 (scale out) commercializes this year, Gen 3 (in-package) is still in R&D. When two independent timelines point to the same thing, credibility is higher than any single vendor's roadmap.

8.3 The Real Ceiling Is Materials and Alignment, Not the Technology Path

This was the most-mentioned yet least-featured theme of the day. Put four independent remarks side by side and it becomes clear:

Dr. Thomas Liu's closing: please help us solve the material shortage.
Hao-Chung Kuo's motivation for choosing quantum dots: InP substrates may be in short supply, so go GaAs.
Wei-Ping Huang's geopolitical concern: two separate ecosystems.
Seyedi's capacity confidence: GaAs/GaN on 8-inch, scalable to 12-inch — consumer-grade capacity.

These four point to one conclusion: in this round of AI optical interconnect, what decides who can ship isn't whose pJ/bit looks best, but who can get substrates, whose wafers are big enough, and whose alignment yield holds up.

This is also why InP export controls deserve long-term tracking — it isn't a news item; it's a structural variable in this supply chain. Hao-Chung Kuo betting on both InP 300 Gbps+ and quantum dots isn't technical greed; it's a supply-side hedge.

8.4 Seyedi's “Optics Is Only a Single-Digit Percentage” Rewrites CPO's Value Proposition

This is the correction most worth remembering from the event — and the easiest to overlook.

If interconnect accounts for only a single-digit percentage of data center power, then the system-level benefit of “CPO saves 70% of interconnect power” is 70% of a single-digit percentage — not irrelevant, but definitely not decisive.

So what is CPO actually selling? Layer Seyedi's and Yuen's arguments together and the answer is clear:

Reliability (Yuen's cited “at least 20x that of pluggables”)
Latency (one less electrical channel, one less DSP, one less FEC latency penalty)
Density (a 0.9RU tray needs 300+ scale-up lanes, which OSFP can't fit)
Converting saved power into compute (Yuen's “every watt not spent on compute is a loss”)

Point 4 is the correct way to describe “power savings” — it's not about saving on the electricity bill; it's reallocating power budget to GPUs. Viewing CPO's investment logic through this framework is far more accurate than looking at pJ/bit alone.

8.5 FIT's Role Has Changed — the Most Practical Takeaway for Taiwan's Supply Chain

The most concrete business fact of the day was hidden in one sentence from the NTT keynote: the detachable optical interface with 500 pins at 200G per pin was built by FIT together with NTT.

This matters on three levels:

It attacks CPO's biggest commercial obstacle. If optical engines can be removed, the service model works; if the service model works, the psychological barrier to customer adoption drops a notch.
It puts the connector maker at the design table. NTT said clearly that beyond connectors, external laser sources are also being co-developed. The connector maker no longer just receives specs and builds parts; it helps define the optoelectronic architecture.
It answers Seyedi's request. Wide-and-slow pushes the problem onto glass, ferrules, fiber, fiber ribbons, active alignment, and high-throughput tools — Seyedi said on stage that this is what Hon Hai is very good at, and publicly invited collaboration.

So the signal from this event for Taiwan's supply chain isn't “CPO is coming, go build CPO.” It's this: whichever path wins, the winner must solve “alignment” and “volume production” — and those two are Taiwan's home turf. For a complete industry-chain map, see Must-See for 2026 AI Infrastructure: The Complete Optical Communications and CPO Supply Chain Map.

8.6 Counterpoints: Three Risks the Conference Didn't Address

To be honest with readers, we have to cover the other side too.

First, this was the host's home turf. Two of the four keynotes closed by praising FIT's capabilities, and the panel made several positive references to Foxconn's manufacturing capabilities. These judgments aren't necessarily wrong, but they aren't neutral third-party assessments.

Second, capital market patience is a real risk. Wei-Ping Huang's slide was more direct than his spoken remarks: leading AI service providers are in a fragile position of negative cash flow and high debt. If the capital structure of AI service providers runs into trouble in 2027–2028, Yuen's 2028 will be pushed out — and what the optics industry has historically been best at is getting hurt collectively when demand forecasts fall short.

Third, wide-and-slow has no standard yet. Hao-Chung Kuo said multi-dimensional multiplexing has “no standard and no standards body overseeing it”; Seyedi said it needs “some common understanding of pitch and rate”; Muhigana said slow and wide “still has to be built and proven.” A path without standards has an unpredictable volume-production timeline. This is the biggest discount to apply when treating 2028 as an investment basis.

9. Conclusion

FIT TECH DAY 2026 had a beautiful theme, but the whole day was really answering the least romantic part of it: at Scale.

The optics industry is no longer debating “whether to use light.” The four keynotes and the panel reached a very consistent consensus — interconnect has graduated from component to architecture, light will move step by step closer to compute, the four form factors will coexist rather than replace each other, and 2028 to 2030 is the window when optical scale-up really happens.

Only two real disagreements remain. One is how long copper can hold on: Semtech says equalizers can push to three meters, Wei-Ping Huang says copper still leads by one to two orders of magnitude at meter scale, and MACOM says once you leave the rack, it has to be optics. The other is who pays to build the ecosystem: CPO is a bundled solution, NPO preserves room for third parties, and wide-and-slow doesn't even have a standard yet.

But the most honest line of the whole day came from moderator Dr. Thomas Liu in the final three minutes: today's real dilemma isn't technology — it's that we don't have the materials to build AI infrastructure.

That line should change how you track this supply chain. Technology roadmaps get updated every quarter, but substrate supply, wafer size, export controls, and alignment yield don't — they are slow variables, but they decide who can deliver.

For Taiwan's supply chain, what this event left behind isn't a technical conclusion but a position: whichever path light takes, it ultimately has to land as alignment, packaging, and volume production. With its 500-pin detachable optical interface, FIT has proven this position can move up to the design table; what remains to be proven is whether it can deliver volume every quarter.