top of page

ECOC2026 | No Technology Revolution for a Decade — Just a Manufacturing War: What Lumentum, Coherent, Marvell and Ciena Agreed On

Oct 10
10 min read

Updated: Oct 30


The closing Market Focus fireside chat at ECOC 2026 put Lumentum, Coherent, Marvell and Ciena on the same stage and asked one question: how does optics keep up with AI over the next three to five years? The four answers converged, and none of them were romantic. Demand is not a bubble. Technology is not the bottleneck. The war is in manufacturing. Coherent went furthest, arguing there will be no major technology inflection for the next ten years because the entire industry's energy will be consumed simply keeping up with growth. For the supply chain, that is both good news and bad news: the roadmap will not be flipped over — but the winners will be decided on yield, process and automation, the least glamorous parts of the business.

1. Why This Panel Mattered: Someone Finally Answered the Bubble Question

The session ran 55 minutes on Day 3 of ECOC 2026, co-moderated by Jon Pugh (Optica) and Loukas Paraschis (Lumentum), with Giovanni Barbarossa (Coherent, senior advisor to the CEO and former chief strategy officer), Rafik Ward (Lumentum), Xi Wang (Marvell) and David Rothenstein (Ciena) on stage. A systems vendor, a silicon vendor, a components vendor and a materials-and-manufacturing view, all in one room — a combination you rarely get at ECOC.

The first question was the one everyone actually wants answered: AI is pushing enormous pressure into the optical ecosystem, but is that demand durable?

Ciena's answer was to break demand into three parts instead of waving at "AI is strong."

The first part is non-AI traffic. Cloud backbone and network applications did not stop growing when AI arrived; on their view there is still roughly 30% of growth left in that layer through 2030. The second is distributed AI training — because no single data center can get enough power, training workloads are being split across multiple sites, and by 2030 roughly a third of workloads are expected to run that way. The third is inference, and that one has barely started: "The past few years were about building and training large models. The next few are about distributing and monetizing that investment."

He quoted a line he had heard the week before at an AI infrastructure summit, and it was the most quotable sentence of the session:

Compute is what started the AI race. Networking determines who scales.

▲ The three demand engines Ciena described — non-AI traffic, distributed training and inference — and the line that framed the whole panel. Graphic: Simple Tech Trend

2. The Order Got Reversed: Connectivity Used to Wait for Compute

Marvell took the systems-architecture angle. The biggest change in moving from general-purpose computing to AI computing, they argued, is not the raw compute number — it is that the order of innovation has been inverted.

The old rhythm was hardware and network infrastructure first: submarine cables, terrestrial networks, data centers, all built out, with compute catching up afterwards — "compute was always two generations behind." Now it is the opposite. AI and compute run out in front, and connectivity is the side chasing. That inversion is exactly why optical vendors are suddenly in the spotlight: not because optics got better, but because optics became the limiting factor.

Within that frame, interconnect splits into scale-up, scale-out and scale-across — and a fourth domain was flagged: scale-in, pulling optics inside the package, straight to compute and memory. The recurring line: light is the carrier of data, electronics is the language of compute, and how you build that bridge at AI scale is the industry's one big question.

Worth noting: the unit volumes differ by orders of magnitude across those domains. The further in you go, the higher the unit count, the lower the ASP, and the smaller the tolerance for power and yield problems. That is also why the same companies can back seemingly contradictory approaches at once — they are not competing for the same socket.


▲ Four interconnect domains and the technologies mapped to each — coherent and multi-rail fiber for scale-across, pluggables and LPO for scale-out, CPO / NPO / wide-and-slow for scale-up, and scale-in further out. Graphic: Simple Tech Trend

3. "Which One Wins" Is the Wrong Question

The moderator asked what everyone in the room wanted to ask: with CPO, NPO, external light sources, pluggables and wide-and-slow all in play, which one wins?

Lumentum's answer was honest: the industry periodically converges onto a shared roadmap, and this is clearly not one of those moments. This is a divergent phase — many competing solutions, overlapping, with no clear winner, and no settled answer even on whether NPO or CPO comes first. The fast-and-narrow versus wide-and-slow debate forks again inside the wide-and-slow camp itself. The line only a veteran delivers: looking back, the answer is usually obvious; looking forward, it never is.

Ciena went further and called the "copper or photons, pluggable or co-packaged" framing an almost dogmatic debate. Their view: all of it happens, and multiple technologies coexist for years. As lane rates move from 100G to 200G to 400G, and modules stack from 800G to 1.6T, 3.2T and 6.4T, what changes is the fuel mixture — not which fuel gets switched off. Pluggable optics stay for a long time. CPO and NPO intercept the market over the next few years, and then they stay for a long time too.


▲ The 2026–2030 interconnect mix — copper shrinking in reach but not disappearing, pluggables persisting, CPO and NPO intercepting, wide-and-slow and OCS each finding a slot. Graphic: Simple Tech Trend (illustrative, not a forecast)

4. The Real Case for Wide-and-Slow: A Production Line Someone Else Already Paid For

The most valuable stretch of the session was Lumentum and Coherent trading remarks on GaAs, because it pulled "wide-and-slow" out of the technology debate and back into supply-chain reality.

Coherent started with Face ID: that supply chain went from a few million units a year to hundreds of millions, one of the very few cases of compound semiconductors reaching true consumer-scale volume. Lumentum finished the story — that is GaAs, not InP, a line both companies invested in; but handset volumes have not grown in years and every generation brings another die shrink, so that line now has spare capacity, and it was built to generate photons in enormous quantities in the first place.

Put those two halves together and the appeal of wide-and-slow becomes obvious. It is not only lower per-lane rate, lower power and lighter DSP burden. The more practical driver is that it can route around InP as a chokepoint, riding a line that consumer electronics already scaled and that still has headroom today.

Marvell made the same point more bluntly: the industry is hitting a materials-and-capacity ceiling, and if you are willing to architect for "wide and slow," you can use large numbers of micron-scale devices that were built for consumer markets and already proven in volume, then bridge them with electronics.

The other side of that argument deserves equal airtime. GaAs spare capacity sits on 3D-sensing, short-reach device specs. Converting it into dense data-center arrays means redoing wavelength, reliability, temperature grades and test methodology. Capacity is not yield, and it is certainly not automotive-grade reliability.

5. The Bottleneck Is Not Light — It Is Manufacturing

Coherent spent a long stretch on something deeply unfashionable: wafer size.

Electronics moved from 4-inch to 6-inch in the 1980s, 6-inch to 8-inch in the 1990s, and on to 12-inch after that. Compound semiconductors for photonics are still working their way toward 6-inch today. The conclusion was sharp: on the manufacturing side, photonics is a long way behind electronics, which is exactly where the room to innovate is — but there is no shortcut. There is no replacement for material science, and no replacement for process technology, because that is ultimately where real differentiation comes from.

Marvell added the same point from the other end. Their approach is to treat packaging itself as the optimization target: ask whether packaging can be removed so devices ship as bare die, and make assembly look more like building an iPhone — fewer stations, less manual labor, automation taking over — so customers can ramp and ship fast. That is not a cost argument. It is a capacity-elasticity argument.


▲ Wafer-diameter timelines for silicon electronics and compound semiconductors, roughly three decades apart; Coherent’s position is that differentiation ultimately comes from materials and process. Graphic: Simple Tech Trend (timeline approximate)

Ciena turned it into an acceptance test. Every solution discussed on that stage is fundamentally about reducing power, cost, space and latency while increasing capacity, density and signal integrity. But if manufacturability is not designed into the architecture, the prettiest product in the market is meaningless — if a hyperscaler cannot get volume, it will not buy.

6. The Most Underrated Passage: Customers Are Buying "Nothing Goes Wrong"

If only one passage from this session survives for product managers, it should be Coherent on decision psychology.

The observation: the number of people who can actually approve a new technology in this industry is startlingly small — on the order of ten to twenty decision makers worldwide, and at the hyperscale end some of those calls concentrate in a single person. Their shared behavior is simple: they always pick the lowest-risk option. Not the most advanced, not the cheapest — the one least likely to blow up in their face.

So the advice to new-technology suppliers was not "make the spec best in class." It was minimize the risk the customer has to carry when adopting your product. This industry has a long memory for failure; one large-scale quality event makes a whole generation of purchasing decisions conservative.

Lumentum gave the venture-side version of the same point. Funding conditions for photonics startups are nothing like they used to be; they have met on the order of a hundred of them, and the founders fall into roughly two camps. One camp is genuinely strong at technology development — and there is still another gap to fill. That gap is the distance from technology to volume manufacturing to a customer willing to place an order.

Put the two halves together and you get the panel's most practical sentence: good technology does not mean it gets bought; manufacturable, repeatable and auditable does.

7. Build, Partner, Invest, Acquire: How Four Companies Allocate Bets

Pushed to the strategy layer — when do you build, when do you partner, when do you invest? — Ciena gave the cleanest framework, usable as a decision table:

  • Build when it sits on top of an existing capability.

  • Partner when the market is still uncertain, or when you want a hedge before making a more substantive strategic commitment.

  • Acquire based on what the market needs and whether you can get there on your own within the window the market allows — time is the real variable in M&A.

  • Invest, through corporate venture capital, to understand emerging, disruptive technology in adjacent markets. It is a ticket to see the future.

He also said the honest part out loud: bet long enough and some of those bets go wrong. That is the nature of the business. Lumentum added a second acquisition test — strategic fit is hard to describe, but you know it when you see it — and noted that partnerships usually happen when each side holds half of a capability.

Coherent laid out the competitive reality: a customer that wants the same thing can go to four, five or six suppliers. In that structure, a supplier without differentiation is only ever used for price discovery. Marvell's answer is to go back to the foundations — build reusable technology blocks deep enough that one set serves both scale-up and scale-out — and to treat the ecosystem as an asset the industry has to maintain together. Lumentum summarized the whole segment in one line: the market is very big, so knowing what you are good at matters more than anything.

8. What Got Named Next: OCS, and Moving Into the Package

Asked what they find most exciting architecturally, the panel named two things.

The first is OCS (Optical Circuit Switch). Lumentum's description was vivid: it is a relatively new product category, and when you walk a customer through what it can do, you can almost watch the gears turn — because it is open-ended and general enough to be dropped into many different places in the network. They expect the list of OCS use cases to keep growing. Coherent added a concrete technical wish: whoever pushes OCS switching speed up opens a new space for innovation.

The second is Marvell's "break the walls down": if light can be extended all the way to compute and then to memory, network tiers can be flattened dramatically and the system can finally be optimized as a whole. They also pointed at the moving boundary — copper's sensible reach is shrinking inward, down to a few millimeters, and everything beyond that belongs to photons.

Ciena's pick was the most operational: multi-data-center interconnect. Power load at a single site has topped out, so hyperscalers must distribute training workloads across sites; back-end GPU-to-GPU traffic runs in the tens of Pb/s, far beyond what a single fiber can carry, which forces multi-fiber, multi-rail designs. Their framing: the fiber rail is becoming the new unit of data-center capacity — the point is not digging another duct, but densifying existing fiber assets so operators can upgrade without rebuilding line amplifier sites.

9. What It Means for the Component and Packaging Supply Chain

Collapse the four views and there are three things worth acting on.

First, the materials and substrate timeline is now confirmed. When Coherent says there is no replacement for material science and process technology, that is an admission that InP and GaAs capacity, wafer size and yield are the real gate for the next few years. Substrate and epitaxy expansion schedules will decide how much module vendors can actually take. (Related, in Chinese: InP substrates as the real chokepoint.)

Second, packaging and test are gaining structural value. Marvell's bare-die shipping and automation, Ciena's manufacturability acceptance test, and Coherent's risk-minimization argument all point at the same process steps. For companies positioned in coupling, packaging and test, that is both an opportunity and an obligation. (Related, in Chinese: what breaks CPO is the test nobody wants to talk about.)

Third, do not bet on a single route. The panel's most consistent message was coexistence. Betting everything on CPO misses the long tail of pluggables; betting everything on pluggables misses the scale-up ramp. The safe position is capability that all four domains consume — light sources, coupling, passives, test and mechanics.

The counterargument belongs here too: all four companies are incumbents. "Multiple technologies coexist" is both their analysis and their safest public position. Genuinely disruptive change is never announced by the companies on stage.

10. Takeaway

No new specification was announced in this session and no new product was shown, yet it described the real state of the optical industry in 2026 more clearly than any technical program: the divergent phase in technology is not over, but the industry has already moved its energy from inventing to being able to build.

Coherent's line — no major technology inflection for ten years, because keeping up with growth will consume everything we have — sounds conservative. It is actually the highest-grade warning the supply chain will get this year: over the next few years, the companies that make money will not be the ones with the prettiest technology, but the ones that can deliver stable volume at the exact moment the customer is most afraid of something going wrong.

Translated into one action: put resources into yield, automation and test — not into the next acronym.

This article is technical and industry trend analysis only and does not constitute investment advice.

Related Reading

Comments

Rated 0 out of 5 stars.
No ratings yet

Add a rating
bottom of page