In Part 3, the penultimate part, four more executives, including Benny Mikkelson, Chris Cole, and Antonio Tartaglia, share their learnings from ECOC 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 the 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-niobate 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.