Roy Rubenstein

Part 2: Start-up funding

Scintil Photonics is betting that to keep scaling AI compute systems, integrated laser arrays will be needed alongside AI accelerator chips.

Scintil Photonics, a spin-off from French research lab CEA-Leti, has developed a heterogeneous integration photonics platform that combines indium phosphide lasers with silicon photonics.

The Grenoble-based start-up’s focus is to deliver light sources to feed co-packaged optics (CPO) in data centres. But its ambitions go beyond that.

“I’m convinced that we have absolutely the best heterogeneous integration technology platform in the world,” says Matt Crowley, who joined Scintil as CEO a year ago. “It was developed for a long time, it’s very difficult for others to replicate, it’s been scaled at [foundry] Tower Semiconductor, so it’s proven its manufacturability.”

Scintil’s task is to replace the piece-part manufacturing of numerous discrete optical components with a monolithically integrated design. “At Scintil, we want to take that to the next level by taking silicon photonics and bringing III-V and other more exotic materials into that integration flow,” he says.

Crowley’s background is in MEMS and semiconductors. He founded Vesper Technologies, a company specialising in MEMS microphones and accelerometers, which Qualcomm later acquired. Previously, he helped scale the start-up Sand 9, which was acquired by Analog Devices. That experience—turning custom wafer processes into high-volume production—is Scintil’s next challenge. “At my last company, we scaled to 60 million units with design wins at Samsung and Amazon,” says Crowley.

For most deep-tech start-ups, particularly wafer-based ones, transitioning from a few hundred prototypes to a manufacturing process that can produce millions of units is challenging.

“You have to convince customers you’ll be a reliable supplier, even assuming your specifications are better,” he says.

Structure and scale

Scintil recently raised €50 million in its Series B funding round. The backers include Yotta Capital Partners and NGP Capital, with participation from Nvidia and earlier investors.

“It was great to get that validation,” says Crowley. “Now we have to figure out how to ramp the product to production.” The funding will help the company expand its 50-staff and sites globally.

The company’s headquarters and core engineering are in Grenoble, France, complemented by designers in Toronto and the UK. A California office is planned as a customer-support lab, while Crowley is based in Boston.

“The primary location will be Grenoble, where engineering and operations sit,” says Crowley. California will likely be the second-largest office, where Scintil will work with customers to get systems up and running.

First bet

When Crowley joined Scintil a year ago, the start-up had two product directions. One was a generic photonic integrated circuit (PIC) platform, and the other was the external light source. The first significant decision he took was to focus solely on laser sources, where the company has seen strong customer interest. Crowley refers to this as being ‘laser-focussed on lasers”.

“In my experience, one of a start-up’s greatest advantages is focus,” he says. “A small group with high talent and great teamwork can out-execute larger groups.”

The goal is to develop Scintil’s LEAF Light product, a dense wavelength division multiplexing (DWDM) external laser source designed for next-generation co-packaged optics.

Accordingly, Scintil’s goal is to launch the light-source product, focussing the start-up on that. “To prove our platform and to prove the value of our IP, we have to launch a single product,” says Crowley.

Here, what is required are laser designs with high power, high reliability, and low cost and size. “There are two specs that are really important: wall-plug efficiency, but even more critically, channel spacing and consistency of manufacturing,” says Crowley.

Scintil believes that traditional distributed feedback (DFB) laser manufacturing won’t scale to the tens of millions of dense WDM array chips that will be needed starting in 2028.

Leaf light: precision and power

Scintil’s external light source is a monolithically integrated array of indium phosphide distributed-feedback (DFB) lasers, in configurations of 8 or 16 wavelengths, on a silicon photonics chip.

Each light source chip also features integrated waveguides and on-chip multiplexers, which combine the wavelengths of multiple lasers onto a single fibre. The design also integrates photodetectors and thermal-tuning elements to stabilise wavelength drift. “We take detectors, waveguides, all from the silicon-photonics toolkit at Tower Semiconductor, and put them on one chip with our DFB array,” says Crowley.

Scintil can support the CW-WDM multi-source agreement frequency grid where customers require it. “We are looking at what the customer wants,” says Yannick Paillard, Scintil’s chief commercial officer. “If they want the CW-WDM frequency grid, we can deliver that.”

Scintil  can deliver 8-wavelength implementations at 200GHz spacings or 16-wavelength implementations at 100GHz spacings. And the company’s product will support more than one fibre output—eight wavelengths times eight fibres, for example.

“Because Scintil uses advanced semiconductor lithography, our lasers have better than ±10GHz precision,” Crowley says. “Competition struggles to get better than ±50GHz. That’s architecturally important because if your channels are too close, they start to interfere with each other downstream.”

Power output and efficiency are also on the roadmap. “We’ve achieved up to 20 milliwatts per carrier,” he says. “Market demand for higher power is increasing as customers want to split signals and generate more carriers.”

As for energy efficiency, Scintil cites Nvidia’s published results. “They’ve shown that the dense WDM co-packaged optics approach can get to sub-4 picojoules per bit today, with a path below one picojoule per bit,” says Crowley. “At that point, optics become more power-efficient than copper.”

Framed against copper, the objective is to achieve power per bit comparable to, and ultimately better than, copper at relevant distances.

Manufacturing partnerships

Production leverages Tower Semiconductor’s PH18 silicon-photonics platform, with Scintil performing post-processing to form the indium-phosphide lasers.

“Tower manufactures the silicon-photonics wafer,” Crowley explains. “Then it goes to Scintil, where we have a wafer probe station with a custom probe head and optical measurement capability that we developed. We can do optical measurement of every die on a wafer.” The goal is to transfer that flow to high-volume assembly partners, or OSATs, as volumes increase.

“We’ll take our custom probe head and install it at an OSAT,” he says. “That’s how we scale. I can collect statistical data, feed it back to the foundry and design teams, and get into a continuous-improvement cycle.”

This known-good-die approach also offers flexibility. “Large customers may want to do their own assembly or co-package with other chips,” Crowley adds. “We’re open to selling them known-good dies or full modules.”

Scintil has already given companies samples of its product. The expectation is that it will make several thousand chips in 2026.

Speaking about the challenge of a start-up getting into the biggest accounts, Crowley says it is key to make life as easy as possible for partners.

“Do as much work for them as you can — build the full module, qualify it, give them the reliability data, the audit reports. That’s how you get designed in,” says Crowley.

Reliability

Coming from the MEMS world, Crowley brings a distinct perspective on reliability targets. “My last company had a failure rate of 0.2 parts per million,” he recalls. “In this industry, when someone says 0.7 per cent failure rate, there’s incredible room for improvement.”

He calls reliability “the hidden spec” in photonics. “We treat it as another design parameter,” he says. “If a metal trace is too thin or a layer isn’t laminating correctly, we expect designers and process engineers to fix it. Once wafer-level technology is working, you get a virtuous cycle: costs go down, performance and reliability go up.”

Scintil’s push to industrialise heterogeneous integration is one of many elements that will determine how the optics industry keeps pace with AI’s compute appetite.

Click here for Part 1: Start-up funding 

For Bill Gartner, Cisco’s senior vice president and general manager of optical systems and optics, AI’s rapid rise is not just driving bandwidth demand, it’s forcing a rethink of design, reliability, and component integration.

In conversation, Gartner outlines the logic and direction behind Cisco’s evolving optical strategy.

Bill Gartner’s remit encompasses Cisco’s photonics portfolio: inside the data centre, inter-data-centre transport, and technologies such as coherent modems and chips that Acacia, a Cisco business unit, develops.

“Basically, my job covers anything that is optics or photonics,” says Gartner. “It’s a wide brief that increasingly revolves around a single, urgent theme – AI – that is changing everything.”

One consequence of AI, according to Gartner, is a surge in demand for pluggable coherent optics for data centre interconnect (DCI). “For several hyperscalers, coherent interfaces between data centres – DCI – have had to grow significantly to support expanding AI models across sites,” he says. The requirement is boosting demand for coherent pluggables and, in some cases, optical systems and transponders. For Cisco, this is a boon — especially Acacia, which has become central to the company’s optical strategy.

However, AI is also causing a shift with regard to optical modules used inside the data centre, specifically optics at 400 gigabits and 800 gigabits.

“We have been a relatively small player inside the data centre for optics,” says Gartner. “This is an area where Cisco has great potential to deliver.”

For AI, reliability matters

AI’s hunger for parallelism exposes something that traditional data centre networking could hide. In classical IP networks, the TCP/IP protocol can quietly retransmit lost packets. In a graphics processor unit (GPU) cluster, however, every device must remain in sync. “If one link has a burst error or a link flap,” Gartner explains, “it impacts all of them, and the workload has to stop, requiring a back-up to a checkpoint and a restart.” This networking development means that reliability is a significant issue for AI.

Gartner cites a study by SemiAnalysis suggesting that, in a 100,000-GPU cluster, even when the optics having a five-year mean-time-to-failure, the first failure will occur within 26 minutes.

Meta has published data that an AI cluster can experience a 40 per cent cut in computing performance due to such failures.

This is where Cisco is well placed. Gartner points out that Cisco has platform-breadth with its Silicon One ICs, optics, system platforms, and software. With this, Cisco gains insight into how the various elements interact. “The platform approach gives us insights that others, who play only in one of those silos, lack,” says Gartner.

Cisco also examines the components making up the link and factors in ageing effects to ensure sufficient margin over the product’s lifetime. The company also uses rigorous stress testing to push optics beyond formal compliance. In one internal test, Cisco tested 20 third-party optics; and not one of the modules passed Cisco’s qualification regimen.

The implication is simple, says Gartner:

“If you’re buying generic optics, you’d better be sure they’ve really been through a significant qualification process, because in AI, the cost of failure is enormous.”

Cisco has historically qualified third-party optics. However, that is no longer an economical model given hyperscalers ask: “Well, who are you buying that from? I’ll go to them,” he says. And that has been a big part of Cisco’s business. More recently, Cisco has been developing its own technology, including photonic integrated circuits and digital signal processors (DSPs). “That allows us to compete head-on with the suppliers of those optics,” says Gartner.

Co-packaged optics

Some two years ago, at OFC 2023, Cisco showed a co-packaged optics (CPO) prototype in a 25 terabit-per-second (Tbps) switch, an early sign of intent that met, in Gartner’s words, a range of responses, from a yawn to outright negativity.

The criticism, he says, came from suppliers who saw co-packaged optics as collapsing value between the switch silicon and optics. “We put it on the back burner for a bit,” he admits. “But Nvidia’s [co-packaged optics switch] announcement earlier this year has brought it back to the forefront.”

Co-packaged optics’ original driver was power reduction, but linear-drive pluggable optics (LPO) have emerged as another route to efficiency. Cisco is pursuing both, with co-packaged optics activity underway, but measured.

Gartner sees AI workloads as the natural home for co-packaged optics, once manufacturing ecosystems and standards mature: “Only a few companies have all the pieces — optics, switch silicon, systems, and software — and Cisco is one of them.”

Optical circuit switching

Optical switching has been discussed for decades. “I have two patents on optical cross-connects — both expired — which tells you how long this technology’s been looking for a problem to solve,” quips Gartner. However, AI may finally provide a solution to this problem.

In GPU clusters, workloads can require semi-static optical paths between processors. “You could manage that with Ethernet, or you could manage it optically,” he says. Cisco is in “monitoring mode,” assessing how optical circuit switches might complement electrical switching.

Inter-data centre connectivity

When it comes to inter-data centre connectivity, the bulk of the coherent pluggables shipping are the OIF-defined 400ZR and the 400-gigabit ZR+ versions, which offer greater reach.

Hyperscalers account for the bulk of deployments, linking data centres and scaling AI workloads across geographically distributed sites. However, Cisco claims that it also has 350 service providers deploying the optics, with almost all of them using the 400ZR+ standard.

Meanwhile, deployment of 800ZR and 800ZR+ has begun. Cisco had alpha samples one year ago and now is in full production. “It’s a mixed bag,” says Gartner, “with two hyperscalers transitioning to 800-gigabit coherent pluggables, while other hyperscalers are waiting for the OIF-defined 1600ZR and 1600ZR+ specifications to be completed and will skip 800ZR and 800ZR+ to deploy 400ZR and 400ZR+ in the meantime.”

“The innovation has been beyond belief,” says Gartner. “I worked on the very first dense wavelength-division multiplexing (DWDM) system that was ever deployed, and it had a capacity of 20 gigabit-per-second,” he says. “When I came to Cisco, the total capacity on a system was 400 gigabit, and now we are putting 400 gigabit in a pluggable.”

The OIF is currently working on defining 1600ZR and, for the first time, a specification for enhanced ZR+ (1600ZR+). The industry organisation is also developing a pared-down 1.6-terabit version, known as ‘coherent-lite’, for up to 10km.

The decision for coherent players is how best to address the applications and how to design the coherent DSP chips needed. At the recent ECOC 2025 show, companies discussed various possibilities: developing a single DSP for all three standards; two DSPs, one for the 1600CL and 1600ZR, and one for 1600ZR+; or even a distinct DSP for the 1600CL and one addressing 1600ZR/ZR+. Given that coherent DSPs will be implemented in 3nm or even 2nm CMOS, these are very costly undertakings.

“It’s natural for hyperscalers to ask, if I only need 20 kilometres, can I reduce power or cost?” he says. “Coherent-lite has promise there, but you have to avoid fragmenting the market.”

Cisco is committed to the OIF’s 1600ZR and 1600ZR+ standards.

“I think coherent-lite will have applications as well,” says Gartner. “We have not announced any product, but we are certainly investigating coherent-lite.”

Embedded coherent modems

Gartner maintains that Cisco’s early bet on coherent pluggables is being vindicated. When Cisco acquired Acacia, competitors dismissed pluggables as a niche for 400ZR. “Since then, we’ve seen 400ZR+, 800ZR+, and even ultra-long-haul pluggables,” he notes. “In every dimension — bit rate and reach — the application space has expanded.”

Even so, embedded coherent optics continue to offer spectral-efficiency advantages for long-haul and subsea systems. “We’ve hit Shannon’s limit,” he says. “The future innovation is about reducing cost, power, and size — all of which favour pluggables,” he says, but admits that embedded coherent modem designs will still serve the high-end use cases where fibre is scarce.

Pluggable form factors

Form-factor debates still animate the industry. Cisco championed the QSFP-DD pluggable module form factor for backward compatibility with 100-gigabit optics, even as the high-speed alternative form factor, the OSFP, gained traction at 800 gigabit. And now, work is being undertaken for a new high-density OIF pluggable.

But talk of yet another form factor makes Gartner wary. “Whenever someone proposes a new form factor,” he says, “we have to ask: does it support copper as well as optics? Is it backwards compatible? Are we retiring existing infrastructure prematurely?” The wrong choice, he warns, can strand customers.

Silicon photonics and 400-gigabit lanes

Cisco remains a believer in silicon photonics, but Gartner acknowledges the industry’s search for new modulator technologies as optical lane speeds approach 400 gigabits.

“People never thought we’d get silicon photonics to where it is today,” he says. “We are believers — but we are also exploring other approaches.”

The question is whether Cisco continues to push lane speed or moves to wider, slower buses. “Both options are on the table,” says Gartner.

Beyond optics

On his LinkedIn profile, Gartner lists his role as a kids soccer coach. Does he still coach? “My kids are grown,” he says, with a smile, “so I’ve stopped coaching soccer.” Since the COVID-19 pandemic, however, he has been learning to play the guitar.

Playing guitar requires controlled adjustments, as does Gartner’s role regarding Cisco’s photonics’ strategy. Somehow, though, mastering six strings and frets sound simpler than orchestrating a symphony of photonic technologies and business challenges.

Julie Eng, the CTO of Coherent, has received the Lisa Su Woman of Innovation Award, honouring women driving innovation in semiconductors.

She spoke to Gazettabyte about her career, women in engineering, and her perspectives on optical technologies and their developments.

It was the end of the first day at the ECOC conference, held in Copenhagen late last month, when Julie Eng arrived at the conference centre full of energy. All the more remarkable given that she had just stepped off a flight from the US and, after the interview, was going to a dinner engagement.

Eng spent two days at ECOC before returning to the US to receive the Dr Lisa Su Woman of Innovation Award, established by the Global Semiconductor Alliance (GSA) to honour women driving innovation in semiconductors. The award is named after the CEO of AMD, Lisa Su, the inaugural recipient, while Eng is the fourth woman to receive the award.

Eng has also been elected this year to the US National Academy of Engineering. Here, an existing member nominates a candidate who is evaluated on their life’s body of work. Academy members must then recommend the candidate for successful admission.

Early path to engineering

Eng’s career in photonics began with an aptitude for maths and an early example of pragmatic decision-making.

Growing up in the middle of the US, she had never met a scientist or an engineer. Her father was a businessman and her mother, an English teacher. Her school required students to take maths exams set by the local mechanical engineering society, and Eng excelled at it.

“Even when it wasn’t required anymore, I kept taking them because you could win money,” says Eng, laughing. “It was easier than babysitting.”

Eng remained unsure about engineering, so she took a 5-year university programme, three years studying liberal arts and two years at an engineering school. This led to a summer research programme at AT&T, at a time when fibre was being laid in the US Northeast corridor. “It was very exciting, I learned about semiconductor lasers and everything,” she says.

She joined AT&T Bell Labs but then lived through the telecom bubble and its aftermath. She then joined Finisar in 2003, where she eventually became responsible for its optical transceiver design, which generated over $1 billion in revenues. During her tenure, the unit released 270 products. Eng then headed Finisar’s 3D sensing unit, stepping away from communications for five years.

II-VI acquired Finisar in 2019 and had its own 3D sensing business, so Eng ran the two units combined. II-VI then bought Coherent, changed its name, and it was the CEO, Chuck Mattera, who asked Eng to become Coherent’s CTO.

The CTO role

Eng realised there was no clear job definition once starting her role, so she contacted other companies’ CTOs to hear of their experiences and advice. “One of the things I heard and liked was making sure that the CTO office is useful to the company’s business units,” she says. “So, I have tried to make the link tighter.”

Much of the advice she received made immediate sense: staying up to date with the latest technologies, introducing them to the company to ensure the right people are aware of them, and focussing on work of value.

“One thing I do is bring together people inside the company with an idea of innovation,” says Eng.

This year, Coherent held its first Innovation Summit. “To attend, you had to submit a paper, like at a regular conference,” says Eng. There was also a second track for R&D pitches. “A bit like [the TV shows] Shark Tank or Dragons’ Den,” says Eng. There was a pool of money, with the CTO’s Office selecting the finalists that then pitch their ideas to a panel of judges.

Eng says the CTO role is unlike her previous jobs. Before, she was responsible for delivering products which had its own frenetic pace. “I don’t have that anymore,” she says. “Instead, it’s unclear what you should do, so I had to coalesce around what I think the role really is, and what I think are my team’s deliverables.”

Her team’s work spans a vast breadth of topics and applications. For example, Eng has a team working on lasers for fusion energy generation. “It stretches very wide, but it doesn’t mean we have the same pounding pace that manufacturing and production jobs have,” she says.

Eng has now work for two CEOs during her time as Coherent’s CTO. The current CEO, Jim Anderson, joined Coherent in June 2024. Eng says one of Anderson’s strengths is ensuring R&D spending is directed to the highest growth and most profitable areas. “It sounds like motherhood and apple pie, but he has got a methodology for it,” says Eng, pointing to his semiconductor industry background. Semiconductors are a more mature industry than optical components, and Anderson has been at IC companies with their own fabs, so he understands manufacturing, she says. ”Now Michael Hurlston is CEO at Lumentum, so more semiconductor people are coming into our industry,” says Eng.

Technology Trends

Being seated with the CTO is a good opportunity to ask quick-fire questions about technology.

Do VCSELs have a future beyond 200 gigabit-per-second (Gbps) and what role will co-packaged optics play?

“I don’t have anything specific to announce yet, but I’m not willing to say that 200 gigabits is when VCSELs max out,” says Eng, who is also a big fan of VCSELS for co-packaged optics.

At OFC 2023, Coherent demoed, along with IBM, the use of compact VCSELs co-packaged optic modules for a government project. At the time, it was seen as a curiosity. Now, there is more momentum due to VCSELs having lower cost and lower power. Eng gave a talk at ECOC reporting 1pJ/bit using VCSELs. “There are some trade-offs, of course, because for any directly modulated laser in a fixed, high-heat environment,” she says. “The thermal solution has to work to ensure that you can hit the FIT (failures in time) rates.”  Accordingly, VCSELs’ most promising role will be for scale-up network applications, where cost and power are key, and where a reach of 30m is sufficient.

Eng expects to see silicon photonics-based co-packaged optics to be also deployed due to its longer reach using single-mode fibre that also makes it compatible with optical circuit switches. In this architecture, the heat issue is avoided by moving the laser externally.

As for 200Gbps per lane linear pluggable optics (LPO) modules, Eng believes they will find a role in applications over short, controlled links. “Linear retimed optics (LRO) will definitely work, but I think people will try really hard to make linear pluggable optics work too.”

Coherent has set up a 400Gbps/lane lab to assess its own components and work with other firms that provide parts. That way, Coherent can measure performance and form an objective view.

Coherent also has optical circuit switches that use its liquid crystal technology, which does not move up and down like MEMS.

“Optical circuit switches solve problems for our customers and that’s my favourite kind of technology,” says Eng.

Using optical circuit switches allows systems to be reconfigured, in the event of a rack failure or, more generally, as large jobs are completed. Using such switches, AI accelerator clusters can be reassigned to new jobs within the data centre.

“For optical circuit switches, we estimated a $2 billion market by 2030, but we think we might have under-called it,” says Eng. “I also think there’s not going to be as many competitors as there are for transceivers.”

Coherent also has thermal materials such as silicon carbide, diamond, and diamond-loaded silicon carbide ceramics that are becoming increasing of interest for thermal management in the data centre.

Beyond datacom and telecom

Eng is also in charge of technologies used for non-telecom and datacom applications. One such application is semiconductor wafer inspection. Here, lasers are increasingly used to scan wafers to identify defects as the industry moves to 3nm and 2nm CMOS process nodes, with Coherent providing the lasers to semiconductor capital equipment makers.

Another laser application is for OLED displays. Here, a large laser creates a line beam for annealing, by sending heat into the semiconductor process used in the display.

Biomedical is another area albeit volumes are still low. One application is for brain imaging. By optically activating the brain, researchers can study degenerative diseases using Coherent’s lasers.

There is also activity in fusion energy start-ups. One type, known as magnetic fusion, uses superconducting tape inside the magnet. The superconducting material is deposited using pulsed laser deposition. The other approach, laser inertial fusion, uses giant lasers, optics, and crystals which Coherent makes.

“And anything quantum: quantum computing, quantum networking, and quantum sensing,” says Eng.

Women in Engineering

Eng gives talks to students promoting engineering, as well as to women in the field. Her own experience in engineering and the optical industry is positive; it is meritocratic, she says.

“I’m a very driven person and have not felt treated differently,” she says. “But you always notice when you are the only person who is different to everyone else in the room.”

Eng is keen to have diverse opinions in meetings, as it broadens the perspective and delivers better outcomes. To this aim, having people with different backgrounds helps. “It could be they grew up in a different country, or it could be that they think a different way, or they may be a different gender,” she says.

Optical industry challenges

For Eng, this is a notable period in photonics, and with it comes challenges.

One is a supply challenge due to the growth in demand. A second issue is the technical challenges posed by growing data rates. Data rates have been constantly increasing, but their rate of growth – the period of time between speed hikes – is getting shorter.

Eng has investigated how long it takes the entire industry to ship its first 10,000 units, and the time it takes to ship 10 million units in one year. In the past, it was a 10-year period; now it is two to three years.

“So, data rates are growing faster, the new technology is getting absorbed faster, but also you’ve got to ramp up the curve a lot faster,” she says.

Another issue she notes is the advantages and disadvantages of a standardised world.

With standardised transceivers, it is known what to make next. “I know without anybody asking me that the next thing I should do is work on 400-gig modulators, and I can make that decision and understand that completely independently of anyone else,” says Eng. But with co-packaged optics and near-packaged optics, it is becoming a non-standard world. That can also have advantages. It can’t be easily swapped in and out and does not have as many competitors.

However, the disadvantage is that multiple approaches may be needed. “Maybe you do some stuff, and then the customer realises it doesn’t work, and then you have to start over,” says Eng. Moving from that standardised to this non-standardised world has its pros and cons.

The AI opportunity

AI and the hyperscalers are now driving the photonics industry, but there are those concerned that this period may be another bubble.

“We only have the visibility we have, but all we can say is our visibility looks robust,” says Eng, noting that it is still the early innings when it comes to AI.

“There might be some ups and downs, for sure, but so far, what we see matches what the hyperscalers have said publicly about their capex spend.”

Polina Bayvel, Royal Society Research Professor & Professor of Optical Communications & Networks; Department of Electronic & Electrical Engineering, UCL.

ECOC this jubilee year (1975-2025) felt like a party. The large number of exhibitors reinforced the excitement, as did the technical conference attendees numbering approximately 2,000.

Unfortunately, I had to leave halfway through the conference and follow the rest online. The Sunday workshops were packed – especially the one on modulator technologies, with people lining the gangways, along the walls, and sitting on the stairs.

The availability of reliable, high-frequency modulators, especially for coherent transmission, is definitely a challenge. Thin-film lithium niobate is winning as a technology, but there are too few suppliers of low-loss commercial devices capable of bandwidths well beyond 100GHz. Broadband modulators, spanning several wavelength bands, are also clearly missing.

In terms of new developments, work on hollow-core fibre continues to impress, with several low-loss results and the promise of broadband operation over hundreds of nanometres. The post-deadline papers on hollow-core fibre by the Microsoft, Southampton, and Linfiber Technology teams were impressive – especially the long-haul transmission with a fully loaded C-band from the team led by Ben Puttnam (Microsoft).

One of the key promises of hollow-core fibre is the absence of non-linearities and the possibility of using high-power lasers. However, given operators’ caution about allowing high-power (above 20dBm) in their networks, I wonder whether high-power will ever come into practical use? And what about the availability of really high-power amplifiers across various bands? Something for researchers to think about!

Despite the excitement in the field, I felt disappointed and somewhat shocked by how the entire field is being dominated, and its direction shaped, by the hyperscalers.

This development has an impact on the development of optical technologies and on fibre companies. The increasing dominance of coherent pluggables over embedded systems overlooks that both are the result of an enormous amount of research on digital communications and digital signal processing (DSP), and that they are not separate technologies but represent a continuum of optimisation.

The plenary talk by Edward Lee from Nvidia brought this into sharp focus. The push to develop optical networks for AI factories overlooks the fact that these networks consume at most 10 per cent of the power in these data centres. The push to reduce this by halving it to 5 per cent is having an effect, changing the direction of the entire field to serve these goals. The end-goal is to use the savings in power to monetise the graphics processing units (GPUs) with the GPU-as-a-service model. More GPUs will make more money but for whom, though?

Money talks, of course, but are we missing opportunities to develop groundbreaking science, devices, and algorithms in the future if the entire field is now being encouraged to work on co-packaged optics and its applications for short-term gains? At least Nvidia seems to have discovered wavelength-division multiplexing, citing it as an exciting ‘future alternative’ technology!

In the Sunday workshops, numerous speakers from fibre companies cited cost as the biggest challenge. Optical fibre companies, who have led the developments enabling the communications over four orders of magnitude in distance (1km to 10,000km), now appear as ‘poor relatives’ to the hyperscalers, being demoted to the level of ‘commodity’ providers. This was shocking to see.

This may stem from the short-term approach to optical networks, which are made up of fibre links. These are viewed as ‘dumb pipes’ or ‘optical plumbing’ rather than offering intelligent access to bandwidth that more adaptive, intelligent and educable networks could provide. These networks will, in the future, operate on land, in space, and under the sea. Maximising network throughput and focussing on the design of networks resilient to changes in topology, traffic patterns, and service requirements could bring much greater benefits in the long term.

I encourage all optical comms researchers & professionals to think beyond the current co-packaged optics & AI hysteria, as hard as it is to resist the AI tsunami. But all tsunamis eventually pass.

 

Dr Sanjai Parthasarathi, Chief Marketing Officer at Coherent

ECOC 2025 once again confirmed the incredible momentum driving the data centre and communications industries, and highlighted how photonics innovation is evolving to meet the scale and complexity of the AI era.

The prime time for 1.6 terabit is coming! The demand is booming in both scale-out and scale-across applications: we saw a lot of effort at all levels of the value chain (optics, electronics, and integration) to be ready for the expected ramp-up in 2026.

The appetite for bandwidth remains relentless, despite some concerns about long-term uncertainties; demand shows no signs of easing. This sustained growth is driving innovation in areas such as higher bit-per-lane transmission. One area of particular focus was 400 gigabit-per-lane transmission, which underscores the industry’s commitment to pushing the boundaries of speed, efficiency, and cost-effectiveness in next-generation optical interconnects.

Co-packaged optics (CPO) was one of the most discussed topics on the show floor. The conversation has clearly shifted from concept to manufacturability. Turning co-packaged optics into a large-scale reality will require a coordinated effort across the entire ecosystem, encompassing demand from cloud operators and solutions not just at the laser and integration levels, but also down to the passive optics element and fibre assembly.

Alternative architectures are also generating strong interest. Our “slow-and-wide” approach, demonstrated through a 2D VCSEL array and detector configuration, drew considerable attention and thoughtful questions. We see this as a promising pathway for scalable implementation: one that could complement the traditional co-packaged optics roadmap.

A key theme emerging from many discussions was scale-across: a concept gaining traction as data centres evolve into distributed, geographically diverse engines powering AI. The enthusiasm surrounding this idea points to a new wave of market potential as global AI infrastructure matures.

I was also pleasantly surprised by the strong interest in quantum key distribution (QKD).

The joint demo we participated in attracted more attention than expected, boosted by growing momentum from government programmes.

ECOC 2025 highlighted how innovation in photonics, from co-packaged optics manufacturability to new scaling paradigms, will continue to shape the future of AI connectivity and high-performance networks.

Lisa Huff, Chief Analyst, Optical Components, Data Centres and Connectivity at DC Tech Analysis

Just like OFC 2025, ECOC was all about AI networks and how the optics industry can support them. I expect this trend to continue for years to come as cloud service providers struggle to build their massive AI infrastructure.

At ECOC 2025, I observed two significant technological advancements.

• Progress towards the realisation of co-packaged optics/ near-packaged optics from both Meta and Broadcom. The reliability data shown by both companies highlight that the promise of the technology may be realisable. While co-packaged optics has seen slow, incremental advancements, this seems to be a much larger one. Whether this can be commercialised and deployed into production data centres remains to be seen, but it’s a monumental milestone along this path.
• 400 gigabit-per-line electrical interfaces are progressing much faster than anticipated. Just six months ago, at OFC, most companies were struggling to show any kind of signal integrity for any type of PAM encoding, and especially PAM4. At ECOC 2025, the OIF demonstrated a 400-gigabit open eye using PAM4 – a significant milestone, marking great progress in six months.

From one of the workshops, I learned that the fight between VCSELs and silicon photonics is alive and well. I believe it will continue into 1.6-terabit devices and probably beyond that.

 

Jörg-Peter Elbers, Vice President, Advanced Technology, Standards and IPR, Adtran

ECOC 2025 was held in the beautiful city of Copenhagen. The event was very well organised and attracted over 8,000 visitors from across the world.

For me, the most exciting themes were “AI for Optics” and “Optics for AI.”

The former theme focused on leveraging agentic AI and large language models to automate and improve efficiency in network operations.

The vision is to apply a divide-and-conquer strategy, breaking down complex tasks and workflows so that specialised agents—think of AI-based ‘junior engineers’ trained through knowledge input and feedback—can solve them more effectively.

The theme of Optics for AI resonated throughout the conference and show-floor discussions. A highlight was Nvidia’s Edward Lee keynote, which showed the massive scale of next-generation AI factories and explained why Gigawatts are the new currency of data centres.

While copper remains dominant in scale-up configurations, optical technologies are essential for scale-out and front-end connectivity and are expected to play an increasingly important role in future. Near-packaged optics and co-packaged optics promise reduced power consumption at the electrical interface, though pluggable transceivers continue to be vital for interface flexibility and inter-data centre reach.

There is growing industry consensus around OSFP as the preferred format for pluggable 800 gigabit and 1.6 terabit transceivers, which demands equipment capable of dissipating 40W per OSFP in a compact equipment form factor. The 200 gigabit per lane speed currently represents the sweet spot at which electrical and optical lanes can operate comfortably without the need for ‘gearbox’ ICs. The jury is still open on whether next-generation 3.2Tb/s interfaces will adopt 400- gigabit lane speeds or require increased parallelism.

Looking ahead to a data centre landscape where AI accelerators and compute clusters demand petabit-per-second connectivity and ultra-compact, low-power optical short-reach interfaces, our Starfall project anticipated a post-IPoWDM era in which transponders would experience a renaissance.

We envisioned “one box – one band” solutions, where pizza-box-style integrated terminals provide interface conversion from short-reach to dense WDM data centre interconnect optics in a rack-and-stack configuration for a full wavelength band. It was gratifying to see this concept echoed in Benny Mikkelsen’s (Cisco/Acacia) keynote talk, which proposed a full C-/L-band transponder in a box using multi-channel integration.

Another hot topic was the advancement of hollow-core fibers. Pre-show announcements, such as the collaboration between Microsoft/Lumenisity and Corning, signal that the industry is becoming serious about addressing scalability challenges in production and making these fibers more widely available. Insertion losses continue to improve, and operational challenges are being actively addressed.

ECOC 2025 will likely be remembered not only for its 50th anniversary (51st edition) but also for the drone sightings over Copenhagen Airport prior to the event and the European leaders’ summit at the Bella Center on the final conference day.

Gazettabyte was launched in 2009 covering developments in photonics and chips. Looking back, it has been a period of remarkable developments and technological change.

Gazettabyte will continue covering the technological progress and interviewing key industry executives. Two such interviews coming soon include Coherent’s CTO, Julie Eng, and Bill Gartner, Senior VP and General Manager, Optical Systems and Optics at Cisco.

In turn, the editor of Gazettabyte, along with Dr. Daryl Inniss, Principal Market Analyst at LightCounting Market Research, are embarking on the second edition of our book, Silicon Photonics: Fueling the Next Information Revolution. You should expect more, deeper coverage of photonics and chips as well as interviews with luminaries as we progress with the research and findings.

Gazettabyte would like to take this opportunity to thank its sponsors, many of whom have backed the site since its start. Much has changed in the media over the last 16 years. Without Gazettabyte’s sponsors, the online publication would not exist.

Roy Rubenstein
October 29th, 2025