Uncategorized

Marvell Technology announced it will buy optical input/output specialist Celestial AI for $3.25 billion. The deal’s value could rise to $5.5 billion if specific sales targets are met.

“We are playing offence in this company,” said Matt Murphy, Chairman and CEO of Marvell, on a bullish earnings call that opened with the acquisition announcement, adding: “Our future is very bright.”

The announcement marks the end of a notable year for optical interconnects and co-packaged optics, driven by the need to keep scaling AI clusters.

In March, Nvidia unveiled its first co-packaged optics-based Infiniband and Ethernet switch platforms. Broadcom then detailed the TH6-Davidsson, its third-generation co-packaged optics design that adds optical input/output (I/O) to its 102.4-terabit Tomahawk 6 Ethernet switch chip. And in October, Ciena acquired the co-packaged optics start-up Nubis Communications for $270 million.

Founded in 2020, Celestial AI has always targeted its Photonic Fabric technology to eight key players: four major hyperscalers and four chip players undertaking xPU development. Now, semiconductor firm Marvell has assessed the optical I/O marketplace and chosen Celestial AI. And it is willing to pay billions for the start-up.

LightCounting Market Research points out in a research note on the Celestial AI deal that Marvell’s lead customer is Amazon. Marvell separately revealed a related Amazon Stock warrant. Amazon has used such warrants with other vendors, such as Astera Labs and Credo Semiconductor, to discount the products it purchases.

By adding Celestial AI, Marvell’s data centre and AI strategy is strengthened through the integration of optical interconnect with its existing data centre chip portfolio. The acquisition will also reassure Amazon and other hyperscalers that may be working with Celestial AI – a start-up, albeit a well-funded one (Celestial AI has raised close to $600 million in funding) – that it now will have the backing of a key semiconductor player.

AI scale-up

Celestial AI has been developing its Photonic Fabric for AI scale-up networks, where multiple xPUs and memory are connected to enable linear scaling.

Current scale-up networks comprise 72 xPUs in a rack, but the number will keep growing to 144, 512, and 1,024 xPUs. Scale-up networks will also expand beyond a single rack. Connecting racks will require optical interconnect as copper will not be able to cope with the distance – tens of meters – and traffic: tens of terabits coming out of an xPU package.

At the core of Celestial AI’s technology is its Photonic Fabric chiplet, designed to link xPUs, xPUs to memory, and xPUs to a scale-up switch (see diagram below).

Celestial AI’s first-generation Photonic Fabric chiplet supports 16 terabits per second (Tbps), while the second-generation design increases this to 64 Tbps, a factor of four improvement using the same number of optical channels.

The second-generation design will be available sometime next year. The chiplet is added to implement a photonic fabric link on a multi-die xPU package.

The photonic fabric link uses an electrical IC implemented in 5nm CMOS and a separate photonic integrated circuit (PIC) that uses electro-absorption modulators, which Celestial AI claims are thermally stable and compact. By placing the electrical IC above the modulator, the driver-to-modulator path is short, reducing capacitance and improving signal integrity. No digital signal processor is needed at the receiver, reducing latency and consuming a several picojoules per bit.

The protocol Celestial AI uses for the optical link is flit-based. Flits are short, fixed-size packets that improve traffic latency and enable efficient forward error correction. The flit concept was introduced with the PCI Express 6.0 bus. Celestial AI says the latency for GPU-to-GPU comms is 128ns, and flits are an elegant way of managing latency, argues Celestial AI.

The start-up also provides complete link management and a protocol-adaptive layer that maps protocols to the flits, such as AXI (Advanced EXtensible Interface), HBM/DDR, UALink (Ultra Accelerator Link), CXL (Compute Express Link) and ESUN (Ethernet for Scale-Up Networking).

OMIB, memory modules, and the photonic fabric appliance

Celestial AI has also detailed its Optical Multi-Die Interconnect Bridge (OMIB), an optical equivalent of Intel’s Embedded Multi-Die Interconnect Bridge (EMIB) for electrical 2.5D interconnects.

By using OMIB, high-speed interfaces can be moved to the optical bridge, freeing up key space around the module’s edge. So, for a multi-xPU module where all the xPUs need to be connected, Celestial uses a separate plane for photonics (OMIB) to handle the I/O. The optical I/O can support UCIe, Max PHY, or a proprietary die-to-die interface. By freeing up the node’s periphery – Celestial AI uses the ‘beachfront’ solely for high-bandwidth stacked memory (HBM) and DDR memory.

Celestial’s business model has been to design Photonic Fabric chiplets customised to meet a chip player’s requirements, followed by unit sales. For OMIB, a hybrid model is possible: selling IP if the chip player wants to integrate the design in its xPU, but also selling a product in the form of a PIC.

Earlier this year, the start-up detailed a memory module that it claimed was the first system-on-chip with optical I/O at its centre, showing how the modulator can sit beneath the electrical driver at the die’s centre.

“Nobody’s ever built anything like this,” said Preet Virk, Co-Founder and COO at Celestial AI, earlier this year. “Because optics doesn’t like being in the middle of the die, it likes being at the side.”

Celestial AI is using the OMIB to free up the beachfront to interface with HBM and DDR memory. The high-performance memory module combines 48-72GB HBM3e and 2TB of DDR5 of memory along with 7.2Tbps bandwidth in each direction to the module. The HBM memory acts as a write-through cache to the slower DDR5 memory. From the GPU’s perspective, it has terabytes of high-speed memory, even though only a tiny fraction of that is HBM.

Connecting 16 of these modules in a 2-rack-unit (2RU) chassis results in a photonic fabric appliance with 33TB of unified memory coupled with a 115Tbps Flit-based electrical network switch. In the example cited, 16 xPUs connect to the ports such that each xPU, and each port, can access the entire 33TB of memory. Penguin Computing is building the 2RU appliance, and Celestial AI will provide the devices.

The start-up is also offering a module for a network interface card for server-based applications that are not AI but do need a large, unified memory.

Performance benefits

Celestial AI claims that, for deep learning recommendation models, a model might typically needs to be spread across 56 GPUs, not because of compute but because of the memory capacity required. Using its photonic fabric memory, the entire model can be loaded into the fabric and be accessed by a much smaller number of 16 GPUs, outperforming the 56-GPU configuration. The start-up cites a 12.5x performance improvement while cutting GPU count – and therefore capex and power – by nearly 70 per cent.

For large language models, the picture is different. A GPT-4-class system with 16 GPUs might have around 4TB of total memory, of which almost 1.8TB is used by the model weights. Having 33TB of memory fabric, the same 16 GPUs have an order of magnitude more memory. Celestial isn’t claiming fewer GPUs here; instead, it can offer longer context windows, higher batch sizes, and higher revenue per deployed GPU.

For both cases, the bottleneck shifts away from memory capacity and communication overhead back to compute, where xPU vendors are more comfortable competing.

Marvell’s gain

Buying Celestial AI, Marvell gains an optical interconnect technology for next-generation AI scale-up architectures. The technology will complement Marvell’s existing family of data centre chips and strengthens its hand of its custom ASIC unit, which develops core designs such as xPU for hyperscalers.

Marvell gained silicon photonics expertise through its acquisition of Inphi. Celestial AI now will add silicon photonics know-how at the chip and die-to-die levels.

The acquisition can also be viewed more broadly. While much is happening at the xPU level and between xPUs and memory, the system rack is becoming the new ‘compute’ unit, with scale-up architectures linking multiple such ‘nodes’.

This is what buying Celestial AI gives Marvell: Marvell already provides many of the technologies inside the rack, but with the Photonic Fabric technology, it can start addressing system-level issues and play a fundamental role at this higher level of system integration and co-design optimisation.

What next

Celestial AI expects to deliver its first product for scale-up connectivity to a hyperscaler customer in the first half of next year with volumes scheduled for early 2027.  In particular, its Photonic Fabric memory modules and Photonic Fabric network interface cards will sample in the first half of 2026 while the Photonic Fabric chiplet integrated with xPU & switches is targeted for the second half of 2026. The Penguin memory chassis is also expected in  2026.

Celestial AI started generating revenues earlier this year by undertaking three chiplet designs.

Meanwhile, Marvell believes the acquisition will close in the first quarter of 2026. It expects the acquisition to deliver meaningful revenues from 2028.

LightCounting expects scale-up Ethernet and NVLink switches with co-packaged optics in 2026, but it does not expect them to be deployed until 2028. “More advanced co-packaged optics for scale-up switches is behind in maturity, but Marvell does have a window if it can execute,” says LightCounting. Marvell is unlikely to see a return on investment until 2030, but only $1 billion of the deal’s value is in cash, adds LightCounting.

Julie Eng, CTO of Coherent

One memorable book I read this year was The Thinking Machine: Jensen Huang, Nvidia, and the World’s Most Coveted Microchip, by Stephen Witt. This book is worth reading, as it covers the formation and evolution of Nvidia, a company that is obviously very influential in photonics for optical networking in AI data centres.

I also read The Lost and the Found: A True Story of Homelessness, Found Family, and Second Chances, by Kevin Fagan, a San Francisco Chronicle reporter who wrote about homelessness in San Francisco. I also enjoyed The First Ladies by Marie Benedict and Victoria Christopher Murray, which was a fascinating portrait of Eleanor Roosevelt working with Mary McLeod Bethune for civil rights in the US.

There is also The Unforgiving Minute: A Soldier’s Education by Craig Mullaney. Craig is my colleague, and this book summarises his path through West Point to active duty.

I really appreciated this book, both because I learned more about my colleague’s life and experiences, and because my brother was in the military during wartime, which helped me gain better insight into his experiences.

I’m reading There’s Got to Be a Better Way: How to Deliver Results and Get Rid of the Stuff That Gets in the Way of Real Work by Nelson Repenning and Donald Kieffer. It’s about dynamic work design and how to lead a modern company to better results. I’ve started this one, so maybe I can report on it next year.

On re-reading this list, I’ve read mostly non-fiction or historical fiction this year. Maybe I’ll read more fiction in 2026.

Helen Xenos, Senior Director, Portfolio Marketing, Ciena

I chose a career in engineering some 30 years ago because I wanted to understand how things work. (Unsurprisingly, my favourite books are mystery books.)

More recently, my curiosity has expanded from how things work to how people work.

I have become fascinated by the science of human behaviour and what drives positive, lasting change.

This interest has blossomed with the wealth of podcasts now exploring these topics, which have introduced me to insightful books in this field.

My favourite book of 2025 is Changeable: How Collaborative Problem Solving Changes Lives at Home, at School, and at Work, by J. Stuart Ablon, PhD. It offers a mindset shift and presents an alternative to traditional reward-and-punishment approaches for addressing challenging behaviour in children—and adults. One of its interesting ideas is the notion that people who struggle with problem behaviour “lack skill, not will”, and “if they could do better, they would”. Ablon outlines how Collaborative Problem Solving can not only reduce conflicts but also strengthen the underlying executive function skills needed for long-term success, positively transforming interactions at home, in schools, and in the workplace.

At a time when developmental and behavioural challenges are rising dramatically, I found this book’s insights relevant and profoundly hopeful.

Hojjat Salemi, Chief Business Development Officer, Ranovus

There is much noise these days about US–China relations and who will dominate the world with AI. We get bombarded with headlines, hot takes, and opinions. This year, two books stood out to me. One is Breakneck: China’s Quest to Engineer the Future, by Dan Wang, and the other is Reshuffle: Who wins when AI restacks the knowledge economy, from one of my favourite thinkers, Sangeet Paul Choudary.

In Breakneck, Dan Wang explains China’s rise simply and powerfully. He describes China as an ‘engineering state’ — a place where the focus is on building, executing, and iterating quickly (i.e., Engineers are in charge). That mindset has enabled China to roll out high-speed rail, factories, and infrastructure at a pace the rest of the world watches with envy. He contrasts this with the US, which he calls a ‘lawyerly state,’ where even a single energy or rail project can get stuck for years in environmental assessments, zoning debates, and legal challenges. These systems exist for a reason, but they slow the country’s ability to lay the foundations for the next era of technology and industry. This was clear to me when I landed at St. Louis Airport and attended the recent Supercomputing 25 (SC25) conference and exhibition. This comparison gives a clean lens for understanding why some nations move quickly while others struggle to keep up. The book also highlights the downside of the China’s approach.

In Reshuffle, Sangeet Choudary examines a different transformation — the restructuring of global value chains through platforms, data, and ecosystems. He makes a strong case that the centre of gravity is shifting from companies that “own everything” to companies that orchestrate networks and interactions. Scale now comes from ecosystems, not vertical integration.

When you read both Breakneck and Reshuffle, you see two parallel forces shaping the world: China racing ahead in physical and industrial build-out, and global companies reshaping competition through digital platforms and ecosystem advantage. One book explains the speed of building; the other explains the logic of reorganising. Together, they offer a framework for understanding where technology, strategy, and global competition are heading.

Stephen Hardy, former editorial director of Lightwave

While it would be an exaggeration to state that half of the optical communications community plays the guitar, it would be difficult to swing a patch chord of decent length within a session at OFC and not strike someone who does. As I would be one such person were I still going to OFC, it is perhaps no surprise that I eagerly read I Want My MTV: The Uncensored Story of the Music Video Revolution, an oral history compiled by Rob Tannenbaum and Craig Marks that recounts the first dozen years of that onetime music video channel.

Scores of executives, recording artists, video directors, former Monkees, producers, etc., offer their often-conflicting viewpoints on MTV’s creation and evolution. Of course, there’s plenty of sex, drugs, rock and roll, rap (eventually), inside stories, and big hair. Also discussed are a video so bad it destroyed a successful rock star’s career, why MTV stopped playing videos, and the phenomenon that was Tawny Kitaen. The book is laugh-out-loud funny and entertaining.

Those of a more serious bent will want to investigate David Grann’s latest, The Wager. The subtitle describes the book as “A tale of Shipwreck, Mutiny, and Murder”; it is a tale of Pestilence, Cannibalism, Superstition, Dumb Luck, and Grub Street. The Wager was a man-of-war that foundered attempting to round Cape Horn in 1740. The aforementioned shipwreck, mutiny, and murder ensue as part of a harrowing tale of how a handful of crew members and officers survived such catastrophes and made it back to England – just not at the same time nor with the same account of what happened.

Grann offers insight into the everyday challenges of seafaring in the 1740s. It wasn’t for the faint of heart.

Lastly, as a big fan of the Expanse book and TV series, I was excited to discover that the two gentlemen who created that world under the combined nom de plume of James S.A. Corey have launched another sci-fi book series. The Captive’s War series begins with The Mercy of Gods, in which an elite team of Earth scientists can forestall the destruction of the planet by their new alien overlords if they can demonstrate their “usefulness” as humanity’s representatives. The team soon begins to wonder whether the definition of “usefulness” extends beyond completing the task they’ve been assigned. As with the Expanse, the novel contains interesting, flawed characters and ambitious world-building. That world includes far-fetched elements (including sentient parasites that are infecting – and killing – members of the scientific team), but somehow it holds together interestingly.

I can’t wait for the next instalment.

Neil McRae, Chief Network Strategist at Juniper Networks.

The Last Man on the Moon: Astronaut Eugene Cernan and America’s Race in Space tells the extraordinary story of Gene Cernan—Commander of Apollo XVII and the last human to walk on the Moon.

I’ve read this book more times than I can remember, and each time it delivers such a surge of inspirational energy that you could charge your iPhone with it.

Having been fortunate enough to meet Gene many times, I can say without hesitation that he was one of the most charismatic, brilliant, and genuinely inspiring people I’ve ever known. He was a role model in every regard—not only as a leader, but as someone who constantly pushed boundaries and challenged assumptions. His guiding philosophy echoes throughout the book: “How do you know how good you are unless you try?”

The memoir traces Cernan’s journey from his early life to his unexpected place among the Gemini and Apollo astronauts, joining the program surprisingly late yet quickly becoming indispensable. What stands out is his unwavering commitment to experimentation, to stretching human potential, and to paying attention to the details that matter—not just as an astronaut, but as a husband, father, leader, and, importantly, follower. (Not nearly enough is written about followership, in my view.)

Though the book isn’t meant to be a technical manual of the space program, it manages to weave in just the right amount of engineering, mission training, and operational insight to satisfy both casual readers and space enthusiasts.

Reader beware: this book may make you cry with joy. That’s why I keep returning to it. Ultimately, *The Last Man on the Moon* is far more than an astronaut’s memoir—it’s a testament to human ambition, teamwork, and the extraordinary risks behind the Apollo program. I’ve read many biographies and accounts of the era, but Cernan’s stands among the most compelling personal narratives of the space age.

Whether you’re fascinated by space exploration or drawn to stories of perseverance, adventure, and history in motion, this memoir offers a powerful and unforgettable journey—and maybe even a new set of coordinates if we ever build that time machine.

Rebecca K. Schaevitz, Co-Founder & Chief Product Officer, Mixx Technologies

I read quite a bit and could share a stack of books. But instead, I want to highlight something different: the podcast Acquired, hosted by Ben Gilbert and David Rosenthal. They dive deep into the stories behind the world’s most influential companies—Nvidia, TSMC, Trader Joe’s, Costco, the Indian Premier League, Nintendo, and their latest, a three-part saga on Google.

What makes the show special is the way Ben and David bring these stories to life. They make you fall even deeper in love with companies you already admire (looking at you, Trader Joe’s and Costco) while revealing the unexpected decisions and creative pivots that shaped their success.

Even though the episodes run long (4+ hours!), I listen in 20-minute segments on my commute—tiny windows of inspiration between all the roles I juggle as a co-founder and parent.

Reflecting on how others built enduring businesses is meaningful as we grow Mixx from the ground up. And who knows—maybe one day we’ll get the Acquired treatment ourselves. A co-founder can dream.

Chris Cole, Optical Communications Advisor

Earlier this year, I read Supreme Commander: The War Years of Dwight D. Eisenhower by Stephen E. Ambrose, who wrote the definitive biography of Eisenhower (Soldier and President), which I had read previously.

This is the ultimate management challenge of all times in terms of scope, difficulty, and unpredictability. Eisenhower had to manage disparate teams over large geographic locations, with multiple bosses and subordinates holding a full spectrum of views, in the face of a formidable opponent.

We tend to see how history happened as inevitable, but individuals alter the course in dramatic ways.

A lesser commander would have prolonged the conflict for many more years. War brings good and bad management into stark relief because the consequences are so severe. The stakes we deal with are much less; however, the lessons of leadership are universal.

Scott Wilkinson, Lead Analyst, Networking Components, CignalAI

By the end of 2025, my reading has been consumed by Walter Isaacson’s masterful – and very long – biography of Leonardo Da Vinci. But since I haven’t finished that one yet, it’s not a valid choice. Maybe I’ll have it completed in time for the 2026 list.

Other books that were interesting enough to talk about this year included “Killers of the Flower Moon by David Grann, which is required reading for anyone who saw the movie, Wasteland: The Secret World of Waste and the Urgent Search for a Cleaner Future by Oliver Franklin-Wallis which will make you think more carefully about everything you throw away, and Seveneves by Neal Stephenson that Andrew Schmitt convinced me to read and I thoroughly enjoyed.

But the one book that I bothered people with the most at parties was “Then Everything Changed: Stunning Alternate Histories of American Politics: JFK, RFK, Carter, Ford, Reagan” by Jeff Greenfield.

Greenfield’s fascinating book offers three alternative US histories based on events starting in the 1960s, and it differs from other, lesser alternative histories in the expertise of its author. Jeff Greenfield has been a political reporter and author for ages and knows the personalities and temperaments of all of the affected parties. What results are detailed, well-considered, and very thorough alternatives.

The scenarios covered include what if J F Kennedy had been assassinated between his electoral win and his inauguration – something that came very close to happening. If his wife hadn’t come to the door to wish him goodbye on that morning in December 1960, Lyndon B Johnson would have been the president during the early days of the civil rights movement and, critically, the Cuban Missile Crisis. JFK and Lyndon B Johnson were very different men, with distinct personalities and backgrounds. Sometimes history chooses wisely, and sometimes not so much.

Other scenarios include what if Robert F Kennedy had turned in a different direction and avoided his assassin in 1968, and what if Gerald Ford hadn’t flubbed his debate appearance against Jimmy Carter in 1976. Each is investigated in historical narrative form to demonstrate how what we assume was inevitable in our history is often just the luck of the draw.

Every page of the book offers historical insights into names that most Americans know, in ways they may never have considered.

Appearances from John McCain, Gary Hart, and others make the stories seem very real. And the threads the author follows from event to event are logical, with some going quite well and others not at all. History is a series of small events with enormous consequences.

Apologies to those whom I bothered with alternative history stories this year. I promise that, for the next few months, I will limit myself to telling interesting stories about Leonardo Da Vinci.

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 bond the III-V material to the back side of the silicon photonics wafer to form the lasers.

“Tower manufactures the silicon-photonics wafer [up till the III-V processing],” 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

Benny Mikkelson, Senior Vice President and General Manager for Cisco’s Coherent Products (Acacia) team, Co-Founder of Acacia

With this year’s ECOC being one of the largest on record, it was an honour to return to my home country of Denmark to deliver a plenary talk during the opening ceremony.

AI was once again a central theme, with significant discussion around ‘scale-across,’ a new category of AI compute fabric connectivity designed to link AI workloads across multiple data centres. This is a topic we have been engaging with our customers on for some time, and it is good to see industry alignment.

Another major focus at the show was 800ZR/ZR+ coherent pluggable optics. Driven by scale-across architectures, we now expect deployment of this generation of technology to be far larger than previously anticipated. In fact, many of our customers are asking us to accelerate the ramp of the 800-gigabit generation even faster than the 400-gigabit generation. We expect our silicon photonic-based implementation to play a key role in helping us meet the challenge of this rapid transition, just as it did in the case of 400 gigabits.

AI is also intensifying competition between technologies. Coherent optics continue to push toward shorter reaches, while optics are increasingly competing with copper in scale-up applications. New fibre types, such as multi-core and hollow-core fibre, are approaching commercial viability. It is exciting to see such transformative innovations happening at the physical layer, even if full-scale deployment will likely take decades.

As mentioned in my plenary talk, the past decade of coherent innovation has centred around speed, power efficiency, and cost. While the industry continues to tackle today’s challenges with baud rate, fibre capacity, and power consumption, I am confident that our community will keep pushing the boundaries of innovation to discover solutions we haven’t yet imagined. I look forward to seeing some of those breakthroughs come to life at next year’s ECOC in Spain.

Chris Cole, Optical Communications Advisor

The highlight of ECOC 2025 was Meta’s paper, which presented measurements of a Broadcom co-packaged optics switch, a first for a mainstream system based on integrated optics. These have dominated advanced optics discussions for nearly a decade, with a cumulative investment of three to four billion dollars in venture capital. Yet the supposedly nimble start-ups are nowhere near this milestone.

Broadcom initiated its program in 2019, based on a proposal by Alexis Black, Karl Muth, and Vivek Raghuraman, with Near Margalit overseeing the lengthy and challenging engineering process. Drew Alduino, a co-author of the Meta paper, deserves gratitude for leading hyperscalers away from secrecy about operations and publishing reliability data, as seen in this paper, so that vendors can develop solutions.

The Meta paper establishes a clear benchmark for determining the veracity of integrated optics claims; full-system-level measurements mean they are real. If system-level data is not available, a stepping stone was set by Broadcom at ECOC 2024 when they published waterfall curves for every link in a co-packaged optics switch sub-assembly. No measurements mean optics claims are wishful thinking. Ashkan Seydi has promised that Nvidia will publish extensive system-level measurements for their Spectrum-X co-packaged optics-based switch. Perhaps these developments will result in the optics industry no longer tolerating powerpoints with colourful optics cartoons and dazzling application simulations based on them.

During ECOC 2025, Ciena completed its $270 million acquisition of Nubis Communications and featured them in its booth. While Nubis had not published system or sub-assembly level measurements, this acquisition represents the first publicly visible examination of an integrated optics start-up by an organisation technically capable of tough and objective due diligence. Therefore, this represents a fair valuation of a solid and complete co-packaged optics development team and product, although with no deployment.

To put this in perspective, an aggressive co-packaged optics deployment ramp may reach around $1 billion in a few years, a fraction of the total datacom optics market. Most of this will be internal optics, like in Broadcom and Nvidia sub-assemblies, which means Ciena needs to capture a significant portion of the remainder.

On the exhibit floor, the dominant module vendors were showing some flavour of co-packaged optics, typically developed for a single large customer. This suggests that for the multitude of start-ups, the market is primarily technology components for one of the few closed ecosystems of the major players.

Further, after Ciena, there are few large optics companies left who have not made an acquisition in this space. This makes a high return-on-investment (ROI) exit difficult. A trick for sidetracking this stark reality is to declare oneself not just an optics company but a computing company with transformational optics, addressing a much bigger market. This has the added benefit of not needing to be either. When pressed about one area, it can be argued that the innovation lies in another. This may even lead to an acquisition by a non-optics player.

A refreshing development at ECOC 2025 was the waning of the linear pluggable optics (LPO) hype, which had dominated optics meetings for years and had been a triumph of wishful thinking over mathematics. At OFC several years ago, Google and Meta presented a rigorous analysis showing that linear pluggable optics was problematic in hyperscale applications. Yet, on the off chance that their maths may be wrong, the industry rushed headlong into development only to face the harsh reality of data centre operations. So far, there is little deployment, although some continue to have large, link engineering efforts for 100 gigabit per lane. This is at the expense of 200 gigabit per lane, for which linear pluggable optics are unusable.

Other end users are focused on deploying half-retimed (RTLR Standard) optics for 200 gigabit per lane, which have benefited from two years of detailed work in the OIF. More will follow. Thankfully, this is happening without much fanfare.

Antonio Tartaglia, System Manager and Expert in Photonics at Ericsson, Radio and Transport Engineering, Transport Systems

What impressed me at ECOC was the reported progress towards the maturity of hollow-core fibre. Fibre, offering lower transmission latency, is crucial not only for AI data centre interconnects and financial institutions, but also for radio access networks (RAN). As RANs adopt packet front-haul, everything between the baseband units and the radios—including packet nodes and fibres—must support low latency and precise time synchronisation.

I gained deeper insight into the progress of co-packaged optics (CPO) toward volumes. Nvidia’s plenary talk offered new insights into its motivation and implementation, while Meta’s presentation provided solid experimental data demonstrating the superior reliability of co-packaged optics compared to pluggable solutions.

As an engineer with a strong telecom background, I see co-packaged optics as a brilliant proprietary innovation. However, our industry relies on multi-vendor interoperability and initiatives such as the Mobile Optical Pluggable Alliance (MOPA) and IEEE 802.3 to support it. Increased standardisation efforts in the co-packaged optics space, as discussed in Meta’s talk, could help reduce barriers and encourage broader adoption within the telecom industry.

The increasing focus of the optical components industry on AI data centres wasn’t a surprise, but I did sense a subtle temptation to ‘let go’ of the telecom business. Telecom remains a stable market, and abandoning it could pose a long-term risk for component vendors. Additionally, AI relies on telecom networks to function; creating value demands energy, computing power, and—most importantly—user data. Where else is user data supposed to be generated (accessed) and delivered (transported) if not through telecom networks?

An executive at a stealth-mode company

At ECOC 2025, we learned that this continues to be a year of solidifying silicon photonics as the dominant technology of interest for datacoms going forward.

Notably, from multiple workshops, including “AI Interconnect Dilemma – which technology is doomed: VCSELs or Silicon Photonics?”, we learned that silicon photonics was the technology of choice, as directly stated by hyperscaler end-users. For the first time, there seemed to be an overall consensus among the speakers that silicon photonics is superior in the long term for a combination of system performance, reach, reliability, and integration, enabling high data densities. This was a significant development for the community.

Co-packaged optics was a widely discussed topic, not surprisingly, since Nvidia has been public on the subject, starting with Jensen Huang’s monologue on silicon photonics for co-packaged optics at the GTC summit last March.

A big highlight at ECOC, however, was Nvidia’s Edward Lee’s plenary talk. What was surprising was how open Nvidia was to the photonics community about their architecture, needs, and roadmap for co-packaged optics.

Meta also reported on a study using a co-packaged optics test system, which showed co-packaged optics provides a 65 per cent power savings compared to a system with traditional retimed pluggables, running the same configuration with the same workloads.

Both Meta and Broadcom reported achieving 5x better reliability with co-packaged optics, and the observed failures were associated with the remote laser source, rather than the actual laser or the co-packaged optics engines.

Based on presentations from the tier 1 module builders, Nvidia, and other AI hyperscalers, there was also a consensus that scale-up versus scale-out will drive a bifurcation in the transceiver requirements.

Scale-out requires ‘fast and narrow’ 400 gigabit per lane PAM-4 modulation, while ‘slow and wide’ requires WDM and lower lane rates, e.g. a maximum 64 gigabit per lane with a non-return-to-zero (NRZ) signalling format.

Regarding modulator technologies, both indium phosphide on silicon photonics and thin-film lithium niobate on silicon photonics were mentioned as directions of interest for scale-out co-packaged optics to achieve the necessary bandwidths, optical performance, and integration densities.

For scale-up, silicon photonics is well-positioned to be the technology of choice due to the high integration densities required, the need for single-mode fibre, and the suitability of silicon-based modulators. Here, no serdes is desired due to low latency, and micro-ring resonator modulators with 64Gbps non-return-to-zero and WDM appear to be the primary technical approach.

For pluggables, 1.6 terabit using 200 gigabit per lane is now commonplace, and 200 gigabit per lane silicon photonics looks to be projected to take dominant market share over time. The learnings from OFC 2024 and OFC 2025 are starting to come to fruition.

At the exhibit, many booths showcased 1.6-terabit transceivers that were maturing using silicon photonics as expected. Another significant development, which began at OFC 2025 and was validated at ECOC, is that 400 gigabit per lane will be deployed starting in 2028, likely using gearboxing from 200 gigabit per lane PAM-4 electrical to 400 gigabit per lane PAM-4 optical in the DSP. However, the standards bodies still have not reached consensus on the electrical interface. Direct-drive from serdes with no digital signal processor is preferred for near packaged optics/ co-packaged optics to reduce power and cost.

One surprising development was Huawei’s presentation on the last day of the conference, where they showcased what appeared to be a fully ready thin-film lithium niobate-on-silicon photonic platform for 400 gigabit per lane, including reliability data publicly shown for the first time. They demonstrated silicon photonic PICs featuring regions of heterogeneously integrated thin-film lithium niobate for the Mach-Zehnder modulator phase modulator sections, with the remainder of the circuit comprising traditional silicon or silicon-nitride waveguides.

Huawei reported the necessary 110GHz electro-optic bandwidth for the thin-film lithium niobate modulator, integrated high-speed germanium photodetectors, and excellent passive performance, including low-loss edge-coupled fibre attach.

The work demonstrates that thin-film lithium niobate on silicon photonics is now ready for prime time for pluggable transceivers and scale-out co-packaged optics.

Dubbed PCIe 8.0, the newest standard will double the bus’s speed per lane to 256 giga-transfers per second (GT/s), twice the transfer rate of the PCIe 7.0 standard ratified in June. First PCIe 7.0-based products are expected next year.

In the data centre, PCIe is used by general-purpose processors in servers and AI accelerator clusters, connecting processors to storage and network interface cards (see chart above).

The specification of the newest version of the PCIe bus poses significant design challenges. Yet despite the challenges posed by moving to the next speed doubling, Al Yanes, PCI-SIG president and chairperson, expects PCIe 8.0 to be completed on schedule in 2028. “We have very good confidence that we can do 256-gig on copper,” he says.

Design challenges

Doubling the data rate per bus lane means faster signals over a copper channel. The doubling of transfer speed shortens the achievable signal reach and places demands on the design of connectors, while requiring advanced and costly printed circuit board materials. “All the things required to make these interfaces run so fast,” says Yanes.

It is likely that more retimer ICs will be needed to satisfy reach requirements along with advanced copper cabling as an alternative to traces on a printed circuit board. In turn, PCIe-over-optical will play a role for longer reaches. PCIe 8.0 will also require new test and measurement procedures.

A strength of the PCIe protocol is its backwards compatible, all the way to the first PCIe 1.0 that operates at 2.5GT/s. “But do we really need to go that far back?” says Yanes. “There are always going to be these questions and challenges that the Work Group will face, trying to run at these crazy frequencies.”

Given the early status of the PCIe 8.0 work – now at Version 0.3, the first review draft – these remain open questions. Acknowledging the standard’s early development stage, Yanes concedes: “We are not as smart as we are going to be a year from now.”

PCIe and AI

PCIe used in AI systems typically connects the host processor (CPU) to AI accelerators (GPUs) and to network interface cards (NICs) to allow GPUs to connect to the scale-out network. The PCIe bus is also used as the physical layer when the CXL protocol is used with memory. However, with alternative schemes being adopted, does PCIe risk playing a diminishing role in AI?

“Tightly-coupled AI designs are already replacing PCIe with alternatives, such as Nvidia using NVLink to connect CPUs and GPUs,” says Bob Wheeler, principal analyst at Wheeler’s Network. “On the other hand, PCIe remains the underlying physical layer for some protocols such as AMD’s Infinity Fabric and Amazon’s NeuronLink. Increasingly, however, AI interconnects are adopting Ethernet-style physical layers that already operate at 200Gbps per lane.”

Yanes admits it is essential to be fearful and challenge the PCI-SIG’s work to ensure it remains relevant. He welcomes the recent wave of OpenAI investments in hardware companies that promote more AI system options. “That is good for us, we are a standards organisation,” he says. “It helps us compete when there are different solutions for AI.”

While PCIe 8.0 is scheduled for completion in 2028, Yanes says that if AI’s scaling trajectory continues, PCI-SIG members could press to accelerate its release.