Uncategorized

The optical networking industry is entering a period of architectural uncertainty as artificial intelligence (AI) reshapes bandwidth demand inside and between data centres.

“We’re reaching a point where we can’t just keep doubling the speed and keep the same architectures,” says Tom Williams, Vice President of Marketing at Cisco’s Acacia group. “Some of that is because AI is driving different traffic flows.”

In particular, AI is creating new bandwidth requirements for its various scaling networking domains that enable AI supercomputers. These networking domains are known as scale-up, scale-out, and scale-across.

Scale-up refers to high-bandwidth connectivity within tightly coupled AI accelerators or xPUs. Scale-up typically spans a rack but that is about to change. Scale-out is the networking that links these tightly coupled xPU racks and spans rows and rows of such racks, a data centre’s worth. Scale-across takes scaling even further by linking xPUs across data centres.  This is the huge scale that is now needed to support modern AI workloads.

Today, says Williams, scale-across often uses the same optical approaches as data-centre interconnect (DCI). But that may not be sufficient for the next phase.

“To really get where we want to go, we have to think about building things differently,” he says.

Architectural fragmentation

The industry is now debating several competing approaches: continuing with front-panel pluggable optics including the new XPO multi-source agreement, and bringing optics closer to the silicon in the form of co-packaged optics and near-packaged optics.

Williams expects these options to coexist for a period rather than for one architecture to immediately take over.

“There’s a tension,” he says. “AI has been driving a narrative that the innovation cycle needs to accelerate. But at the same time, the industry is trying to work through new architectures and new form factors.”

The historical cadence of Cloud data centres has been set by Ethernet switch silicon such as Broadcom’s Tomahawk chip family. Every two years a new generation of Tomahawk silicon doubles capacity. Now AI is pushing an even faster, annual cadence. Yet moving to new optical architectures slows things down because each approach requires ecosystem work, standards activity, and design trade-offs.

That tension is already visible at 200 gigabit-per-lane electrical interfaces. Although Nvidia and a few others are deploying such interfaces, broader adoption is taking longer than some expected. The move to 400 gigabit-per-lane host interfaces will take longer still.

This helps explain the sudden proliferation of multi-source agreements. Groups are forming around front-panel pluggables, co-packaged optics, and near-packaged optics, each seeking to define a path forward.

Williams sees XPO as especially interesting because it asks what becomes possible when liquid cooling is assumed from the outset.

“The main point of XPO is to say: what can you do if you always assume you’re going to have liquid cooling?” he says.

That could allow higher density than today’s pluggables, which must still fit within air-cooled constraints. But Williams stresses that many details remain unresolved, including yield, reliability, and whether a single form factor can span copper, short-reach optics and coherent interfaces.

“It stirs some good discussion,” he says. “I don’t know if it’s the right answer.”

Supply is the immediate issue

For all the focus on architectures, the most urgent industry challenge is supply.

Williams says the industry is facing multiple bottlenecks, including lasers, memory, and semiconductor foundry capacity.

Williams notes that the capacity of the largest chip foundries was always seen as almost infinite but that now, even here, industry concerns are being raised.

“Overall, we are seeing supply constraints we never considered before,” he says.

Laser capacity is another pressure point. Nvidia’s investments in Coherent and Lumentum are widely seen as moves to secure a more reliable supply of optical components, especially lasers.

“You can spend the money to bring up more laser capacity, but it takes a couple of years to bring a laser fab online,” says Williams. “There’s no way to just snap your fingers and have more laser capacity.”

The issue becomes more acute if optical links move deeper into scale-up AI systems. Such architectures could dramatically increase the number of lasers required.

Acacia’s coherent volumes

Acacia has been highlighting shipment numbers for its coherent digital signal processor (DSP) products.

The company’s embedded Jannu DSP and CIM8 coherent modem 1.2- terabit product has now reached significant volume. Williams says Jannu began sampling in late 2021, went through qualification in 2022 and began ramping in 2023. Higher-volume shipments followed in 2024 and 2025.

Today, Acacia is shipping about 25,000 Jannu units per quarter. That, says Williams, makes it the highest-volume, performance-optimised coherent product in the industry.

“We’re shipping at 25,000 now, and it’s still growing,” he says. “In an environment where the biggest growth is all in pluggables, performance-optimise [coherent modems] are still shipping in significant volumes.”

Acacia has also added an L-band version of the CIM8 to the initial C-band design.

Acacia’s earlier 400-gigabit coherent pluggable modules have cumulatively shipped 750,000 units over five years. Meanwhile, the company’s Delphi DSP for 800-gigabit coherent pluggables is earlier in its ramp, with 25,000 units shipped mostly in the last one or two quarters.

“It is still early in the ramp cycle,” says Williams.

Williams expects the 800-gigabit ramp to be more aggressive than the 400-gigabit ramp, though he avoids overstating the outlook.

“There’s plenty of opportunity for all of us if we can execute,” he says. “It’s a question of who executes and who can ramp effectively.”

Kibo and 1.6-terabit client-side optics

Acacia is now sampling Kibo, its 3nm DSP for 1.6-terabit client optics, with general availability of the IC expected later this year.

The 1.6-terabit client-optics market is already competitive, with established players in PAM4 [4-level pulse-amplitude modulation] DSPs. Acacia’s pitch is that performance and reliability matter more than ever.

The critical metric for AI operators is GPU utilisation. Any optical failure, intermittent link drop or unexplained link flap can reduce the utilisation of extremely expensive AI infrastructure.

“What people care about is uptime,” says Williams.

Acacia argues its coherent heritage and experience with 400ZR coherent-optics deployments give it credibility with hyperscale customers. Williams says some end customers have reported fewer link flaps from Acacia devices than from leading PAM4 competitors.

Reliability is not just one thing, he adds. Link flaps can originate on the electrical side, the optical side, or from system-level issues. Kibo includes diagnostic capabilities to help identify problems, including multipath interference detection that can reveal reflections caused by dust or dirt on connectors.

Acacia is also supporting transmit-retimed optics (TRO). Kibo remains a full DSP, but in a transmit-retimed-optics configuration, the receive-side DSP can be bypassed while still used for diagnostics and loopback functions.

100ZR: a quieter but long-lived market

Acacia has also introduced a 100ZR DSP, its first public announcement of such a product.

This is a different market from the high-speed pluggables for hyperscalers. Instead of a hockey-stick ramp, Williams sees 100ZR as a long-lived migration opportunity, supporting carriers moving from 10-gigabit tunable optics to 100-gigabit coherent links in access aggregation and metro DWDM networks. The product is expected to be generally available in a few months.

Key requirements include launch power, power budget, cost, commercial and industrial temperature support, and the ability to coexist with legacy optical systems. The zero-dBm launch-power option of the 100ZR is important for metro dense WDM replacement applications.

1.6-terabit coherent and beyond

The OIF’s 1.6-terabit coherent modem work is progressing, says Williams. The 1600ZR project is defined enough for DSP developers to proceed, while 1600ZR+ is slightly behind but progressing.

The 1600 Coherent Light (1600CL) is less clear because it spans many possible use cases, each with different requirements and uncertain volumes.

Williams expects 1.6-terabit coherent to be a major node. “I think 1.6 terabit is going to be huge,” he says.

That is because users of both 400ZR and 800ZR-class pluggables are likely to converge on 1.6 terabit. AI demand will further increase the scale of the opportunity.

By contrast, 3.2-terabit coherent remains uncertain. A single-carrier 3.2-terabit coherent interface will be technically challenging, and future architectures may decouple coherent line-side capacity from front-panel client-optics form factors.

The industry is entering a phase in which speed, packaging, cooling, supply, and system architecture are all changing.

Customers want partners that can help them “get where they want to go”, he says, as optical networking moves deeper into AI infrastructure.

• • • • Oriole’s next-generation platform, the Prism Ultra, will scale to 1 million AI accelerators (xPUs).
      • xPUs with up to 51.2 terabits of input-output (I/O) will be supported by using four new XPO modules.

AI models continue to advance, but this progress requires increasingly capable hardware—from xPUs to the networks that connect them.

AI supercomputers already deploy up to 100,000 xPUs across racks spanning entire data centres. Even that is proving insufficient, as AI service providers begin linking multiple data centres to pool resources for large-scale workloads. But scaling compute depends on scaling the network: within racks, across racks, and increasingly, across data centres.

Electrical switches today form the backbone of these systems, arranged in hierarchical network topologies.

Switch silicon continues to improve, doubling in capacity roughly every two years. For example, the latest Broadcom Tomahawk 6 delivers 102.4 terabits per second of switching capacity. However, xPU I/O bandwidth is growing even faster. Processors with tens of terabits per second of I/O are already being discussed.

Even with future 400-terabit switch chips, this would support only a limited number of fully non-blocking xPU connections—on the order of single digits. The result will be deeper network hierarchies, more switching layers, higher power consumption, and increased latency.

Google is already using optical circuit switches in its data centres to link its TPU AI accelerator chips. An optical circuit switch creates a temporary light path between two network endpoints, with the data sent as optical signals. Google’s first used optical circuit switches was to replace the top (Spine) layer of electrical switches resulting in significant power and cost savings. Google also uses optical circuit switches to tailor clusters of TPUs to AI workloads. Indeed, TPUs with optical circuit switches is now the dominant AI system deployed by Google.

Meanwhile, Broadcom and Nvidia have started adding optical interfaces to their switches, an approach known as co-packaged optics (CPO). Co-packaged optics-based electrical switching platforms is an alternative to traditional switch platforms that use pluggable optical transceivers. Using co-packaged optics saves power and has demonstrated greater system reliability but it does not address the issue of switching capacity.  To date, deployments of co-packaged optics-based switching have been limited.

Oriole is taking a more radical design approach. The UK start-up has developed an interconnect architecture that removes electrical switching altogether. Oriole’s argument is that as the xPU count grows, and the data each xPU sends, electrical switching will not keep up. But optical switching will.

But electrical switching dominates data centre traffic for a reason: it supports all types of packet flows including short ones. Conventional optical switching is good for very long packet flows – 100 gigabytes, says Oriole – because of the time it takes to set up a path i.e. when data is sent between the two end-points for a duration much greater than the switch path set-up time.

But Oriole claims its architecture tackles this issue of setup latency versus flow duration.

“Our complete solution allows you to run an optical circuit switch with the packet granularity of electrical switches,” says James Regan, CEO and co-founder of Oriole Networks.

Prism system

Oriole’s Prism architecture connects each xPU to an optical module, dubbed XTR by Oriole, using a PCI Express (PCIe) network interface card (NIC).

Oriole’s XTR module uses a tunable laser to generate optical signals at different wavelengths. The XTR adds further scale by using several fibre outputs, a form of space switch. The XTR also uses time-multiplexing sending data in 100-nanosecond (ns) windows and can switch – set up a new optical path – in 10ns.

The XTR modules connect to Oriole’s second system element, the photonic routing platform that acts as a passive optical network. The combination of the XTR and photonic routing platform results in a nanosecond-reconfigurable optical circuit switch-based network.

Design challenges

Oriole has had to tackle two issues to replace electrical packet switching with optical circuit switches.

The first issue is switching fast enough such that the time needed to create a signal path is a tiny fraction of the time the connection is active. Otherwise, xPUs remain idle while the path setup. AI inference traffic, for example, can involve small, frequent exchanges—often on the order of kilobytes, says Oriole. If each path incurs a millisecond-scale switching penalty, the network becomes inefficient. Since Oriole’s network can switch in 10ns, the overhead is insignificant.

A second challenge is transceiver synchronisation. Conventional optical systems require re-locking after each reconfiguration, which can take milliseconds. Oriole claims to maintain synchronisation across switching events, eliminating this penalty—though details remain undisclosed.

Together, these capabilities allow optical switching to operate at packet-like granularity, a key requirement for replacing electrical networks.

The implication is significant. Instead of routing traffic through multiple layers of electrical switches, data can move directly between endpoints over dynamically configured optical paths.

Scaling to one million xPUs

Oriole’s Prism product is now running in test environments, validating the core switching and synchronisation technology. The XTR optics is being manufactured using Tower Semiconductor’s commercial heterogenous process using technology from OpenLight. And Oriole is working on a full rack inference system trial with AMD, using the chip company’s AI processors. Prism is expected to become available from 2027.

Now Oriole is working on its second-generation Prism Ultra product. With the Prism Ultra, Oriole is no longer using a NIC card. Instead, it is connecting the processor directly to the optical switch. This is because the PCIe bus cannot support terabits of I/O traffic xPUs will soon generate.

“You can’t run that amount of traffic through PCIe; it will require an insane number of sockets,” says Regan.

The Prism Ultra will use the industry’s new XPO (eXternal Pluggable Optics) pluggable form factor announced by Arista and others at OFC in March.

An XPO module support up to 12.8 terabits with its multiple electrical and optical lanes. Oriole will use co-packaged copper and flyover cables to link the xPU to the XPO pluggables. Four xPOs will support 51.2 terabits-per-second of AI processor I/O.

The Prism Ultra still needs networking intelligence given the NIC will be removed from the design.

“The intelligence that was inside the NIC or FPGA now becomes part of the xPUs,” says George Zervas, CTO and co-founder of Oriole.

Oriole says its Prism Ultra will directly connect thousands of xPUs. But to scale further, a way is needed to link between the ‘scale-up’ clusters. Here, an intermediate switching step using an electrical engine of the xPU is used. The engine selects a wavelength and path to address the entry XPO residing in another scale-up cluster hosting the XPU to receive the data. Accordingly, only one hop is needed for the system to link any two xPUs in, effectively, a one-million 2D array of xPUs.

Oriole’s approach promises a significant reduction in power consumption by eliminating large numbers of electrical switches in huge XPU superclusters. Latency will decrease while reliability will improve as the number of active components in the network decreases, says Regan. The total number of transceivers required will also fall, simplifying system design.

Oriole has not said when the Prism Ultra will be available.

Latest funding round

Oriole has already raised $35 million and is now undertaking a new funding round. Its challenge is to translate a technically compelling concept into a deployable, widely adopted solution.

Oriole’s proposal runs counter to the established industry’s direction. Considerable investment continues to flow into scaling electrical switching and into co-packaged optics-based switch designs. These approaches extend familiar models rather than replacing them. Oriole’s architecture, by contrast, suggests a more optical approach to next-generation AI infrastructure.

Whether the industry continues its incremental path or adopts more radical approaches is unclear. Should Oriole’s nanosecond optical switching prove practical, it will offer an alternative way to scale AI supercomputers.

In the final OFC 2026 reflections, Cisco’s Bill Gartner, Vivek Raghuraman of Mixx Technologies, and Vincent Fraisse at STMicroelectronics share their thoughts

Bill Gartner, Senior Vice President, Optical Systems and Optics, Cisco

The AI buzz was palpable at OFC. It is super exciting to be in the optical networking industry and be part of the AI infrastructure buildout.

If you came to our booth, you would have seen the socks we gave out that say “Optics are Sexy Again,” which I think sums it up.

Scale-across networking was a hot topic at the Optica Executive Forum, where executives shared and shaped their perspectives. Scale-across differs from traditional data centre interconnect (DCI) in that it enables Graphics Processing Unit (GPU) connectivity in AI back-end networks, whereas traditional DCI provides connectivity between front-end networks.

Benefiting from scale-across networking, we expect the deployment of 800ZR/ZR+ coherent pluggables to be far larger than previously anticipated.

Acacia announced that it had already shipped 25,000 800-gigabit digital signal processor (DSP) ports, contributing to its coherent technology market leadership.

Another optical enabler of scale-across is the line system, and several new multi-rail line systems were announced, including Cisco’s Open Transport 3000.

To support the Petabyte-scale bandwidth demands of scale-across, more fibres will be needed, and each fibre pair requires in-line amplification. The multi-rail open-line system combines multiple fibre pairs in parallel on the same link over ultra-long-haul distances, which enables our customers to deploy additional fibre capacity with reduced amplifier power and space needs.

Optics for scale-out were also demonstrated with 1.6-terabit optics delivering ultra-high bandwidth connectivity and 800-gigabit linear pluggable optics (LPO) to reduce optical module power consumption by half compared to retimed optical modules.

Next-generation architectures were a key topic this year, with continued focus on co-packaged optics (CPO) development and four new MSAs – 400G Optical, XPO (eXtra-dense Pluggable Optics), Optical Compute Interconnect (OCI), and Open CPX. We’re seeing AI demand driving aggressive development timelines and new architectures. Technology transitions like co-packaged optics, liquid cooling and optics in scale-up networking are building momentum in the industry, but will co-exist with existing approaches over the next generation or two.

OFC is always my favourite week of the year because I get to talk to so many customers and industry colleagues. This year did not disappoint.

Vivek Raghuraman, CEO, Mixx Technologies

Last year, we left OFC with a prediction: that 2026 would be where system-level transformation takes over. Having spent the week in Los Angeles, I can say the industry got closer, but the path is complicated.

The shift that stood out most was not a product announcement or a new form factor. It is a change in tone. The co-packaged optics conversation, which dominated 2025, has matured from architectural vision to engineering accountability.

What unsettled me was a parallel current running through several sessions — a gravitational pull toward complexity. Thermally sensitive multi-wavelength designs, denser laser integration, and manufacturing yields that are unknown.

Technically ambitious, yes. But complexity has a cost that doesn’t show up in a lab. It shows up in the forward-error correction (FEC) overhead — the latency, the power, the operational burden carried not by the engineers who designed the system, but by the ones who run it at scale. That is not a scaling strategy, it is scaling debt.

The discussions that cut through the noise were the ones asking a simpler question: when you move from a controlled demo environment to thousands of deployed nodes, does the architecture still hold?

The answer to complexity is not more complexity.

Scale-up will follow scale-out architectures, and it will do so on the back of electrical SerDes, DR optics, and 200-gigabit /400-gigabit linear interfaces — proven, deployable, and built to run at scale by the engineers who operate them.

Deterministic performance kept surfacing — not as a feature, but as a deployment prerequisite. Inference workloads don’t forgive jitter or variance. That reality is beginning to shape architectural decisions in ways that were not visible even at OFC 2025.

A fundamental shift is underway — and it starts with radix. More endpoints connected directly means fewer switches in the path and fewer hops between nodes. The network flattens. The fabric breathes.

In AI infrastructure, every hop is a latency you cannot afford and power you cannot reclaim. That is not a marginal gain, it is a structural one.

That, ultimately, is what OFC 2026 revealed most clearly: an industry at a genuine inflection point. The ambition is real. The investment is significant. But alignment is still missing — around standards, around manufacturability, and around what truly matters at the system level. Until that alignment emerges, complexity will continue to masquerade as progress.

I left Los Angeles with a clear conviction: the demonstrations are real, and the deployments are measured, exposing the gap between what demos well and what scales.

Vincent Fraisse, Executive Vice President, RF and Optical Communications Sub-Group at STMicroelectronics

With my STMicroelectronics foundry hat on, I first looked at technology breakthroughs at OFC. I found a few things: first, the 400 gigabit-per-lane is a hard barrier to overcome due to limits on modulator bandwidth, driver power, optical loss, and signal integrity.

Modulator technology contenders were demonstrated at the show – Thin-film Lithium Niobate, Indium Phosphide, plasmonic, EML and even pure silicon photonics – but the path to industrialisation and large volumes remain challenging.

Still, the appetite is big to pursue “fast and narrow” design improvements, and the first technology to deliver enough performance margin and volume readiness will win big.

Meanwhile, the industry continues to densify bandwidth through more lanes, more wavelengths per lane, and tighter integration, such as the new XPO multi-source agreement. Integration and volume readiness will matter even more than before, and my takeaway is that silicon photonics is the right technology to bet on for the near future because it offers a strong path to scale, best functional integration and manufacturing leverage.

However, the “slow and wide” signalling camp also showed great progress at the show. With the OCI MSA, with links based on silicon photonics or more innovative solutions like microLEDs, the industry is trying to reduce the energy per bit transferred with a more de-serialised architecture.

This is much like what computing did years ago: it stopped chasing gigahertz clock speeds and focussed on parallelism. For optics it means increasing bandwidth density to terabits-per-second-per-millimetre-edge. The main technical challenges remain device efficiency, coupling loss, thermal budget, and link calibration.

The biggest momentum, of course, came from the optics boom for the scale-up network. As copper reaches its limit in reach and bandwidth density, many demos focused on making co-packaged optics a nearer reality rather than a distant dream.

Most implementations would actually be classified as near-packaged optics, but what matters is that industry seems to be moving quickly to enable volume deployment within a couple of years. This will require improving advanced packaging, increasing wafer-level test scalability, and addressing system serviceability through fibre-attach connectors, among other issues.

Again, the show exposed a new CPX MSA, many proofs-of-concept, a clear focus on the problems to solve, and many innovations driven by start-ups. But now, all the big companies have plans. This will become a reality sooner than most people think.

Gazettabyte has been asking industry figures for their thoughts after attending the OFC 2026 conference held in Los Angeles in March. In the penultimate post (Part 3), Nokia’s David Heard, Vipul Bhatt of Coherent, and Adtran’s Jörg Peter-Elbers share their thoughts.

David Heard, President Network Infrastructure, Nokia

The biggest OFC takeaway was the sheer scale of growth across networks, coupled with an accelerating diversification of connectivity applications.

The industry is experiencing unprecedented demand driven by AI, cloud, and next-generation digital infrastructure, and this is driving innovation in optical technologies and architectures – primarily for high-capacity, power-efficient solutions that can be supplied in substantial volumes.

Against this backdrop, what also stood out for me is the growing importance of co-innovation efforts with customers and partners. This was reflected in the positive reception of the vision we announced for solutions optimised for our customers’ applications. This is a strategic priority for Nokia, as is staying disciplined in focussing capital on areas where Nokia has a clear technology advantage and can differentiate.

This creates significant opportunity across the ecosystem, from AI cloud operators and data centre providers to service providers and infrastructure vendors.

From my conversations with customers, it is clear that the ability to bring expertise across fixed, IP, and optical domains creates real value as AI reshapes traffic demands and patterns across every layer of the network.

In a supply-constrained environment, that value extends to expertise in critical material sciences and having your own semiconductor manufacturing and advanced packaging, which increasingly matter for scale and resilience.

Demand for 800-gigabit coherent pluggables continues to outpace even aggressive forecasts, with industry volumes expected to exceed three million units annually by 2030.

There is also strong momentum around 1.6-terabit coherent technologies. Multiple approaches are under discussion, including subcarrier-based implementations. These aim to extend reach for long-haul transmission.

Another noteworthy OFC theme is the growing focus on multi-rail line systems. The industry is prioritising higher amplification density and greater service capacity to maximise infrastructure footprint, highlighting the criticality of efficiency and density as networks scale. One takeaway from OFC was how rising volumes and growing application diversity are changing what scale and innovation mean for the industry.

Customers no longer want one-size-fits-all architectures. They seek solutions tailored to specific use cases, such as AI-driven data centre interconnect, metro, or long-haul transport. They demand suppliers meet capacity requirements that, in some cases, exceed historical norms by an order of magnitude.

There was also significant industry interest in the new XPO (eXtra-dense Pluggable Optics) module, which is emerging as a compelling alternative to co-packaged optics (CPO) and near-packaged optics (NPO).

XPO has the potential to address some of the engineering and manufacturability challenges that have slowed the adoption of those approaches.

Hollow-core fibre continues to generate interest, especially for its latency benefits. However, it remains several years from large-scale deployment. Challenges around manufacturing cost, operational readiness, and ecosystem maturity still need to be addressed.ugh.

What struck me at the show wasn’t a single “new” idea, but rather how quickly the industry has aligned around a few non-negotiables: delivering new levels of scale and efficiency as AI drives unprecedented demand, addressing the resulting power consumption challenges, and translating cutting-edge innovation into real-world business purpose.

Vipul Bhatt, Vice President of Strategic Marketing at Coherent

What stood out at OFC was not just that optical networking continues to grow. By now, that is expected. Instead, it was the breadth of the expansion on display. I saw an industry widening across technologies, architectures, companies, and alliances.

First, the development envelope is broadening. We saw multiple 400-gigabit demonstrations even as 200 gigabit is still ramping. There was a diverse set of form factors and device platforms. For example, VCSELs (vertical-cavity surface-emitting lasers), silicon photonics, EMLs (externally modulated lasers), and indium phosphide Mach-Zehnder modulators were all represented.

Second, innovation is unfolding across all three AI network domains. Optics in scale-up came into much sharper focus. Multi-rail approaches for scale-across gained visibility. The industry also showed steady progress in scale-out.

Third, there are newer players.

In optical circuit switching alone, there are more than a dozen entrants.

Lastly, the industry has formed several new alliances. These include three major multi-source agreements — XPO, Open CPX (co-packaged ‘x’), and Optical Compute Interconnect (OCI) — a telling sign that collaboration is scaling alongside the technology.

When an industry is expanding on multiple fronts, it does not just get bigger. It becomes more generative. More entrepreneurial bets are placed, more unconventional ideas are tested, and more consequential innovations emerge.

The rate of useful surprise increases. In the mid-1980s, Ethernet went through such an expansion. OFC 2026 suggests that optics for AI is entering a moment like this. It was a great time to be alive then, as it is now.

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

OFC underscored how quickly optical networking is accelerating under massive AI infrastructure investments. Technology developments that long sat on the industry roadmap are now moving rapidly towards deployment.

AI model training, advanced reasoning, and increasingly autonomous AI agents are driving capacity demand at a pace that makes higher interface speeds and tighter electro‑photonic integration unavoidable.

As the industry moves from 800 gigabit-per-second (Gbps) to 1.6 terabit-per-second (Tbps), power efficiency and manufacturability have become as critical as reliability and performance. This shift creates opportunities for VCSEL‑based scale‑out architectures, as well as innovative coherent scale‑across solutions that extend from data centre interconnect into metro and regional networks.

Meanwhile, network operations are being reshaped by agentic AI.

AI systems that can analyse, reason, and act promise a more intuitive operating model than traditional network management approaches. Engineers will increasingly interact with their networks using natural language, enabling real‑time, closed‑loop responses across open and disaggregated optical infrastructures.

Hollow‑core fibre is another technology crossing an important threshold. As practical installation challenges are addressed and manufacturing capabilities expand through industry partnerships, it is now approaching deployment at a meaningful scale. Its combination of ultra‑low latency, reduced nonlinear effects, and favourable dispersion characteristics paves the way for a new generation of high‑capacity dense WDM systems, enabling fewer, higher‑power amplifiers to efficiently extend reach and capacity.

Lastly, the sky is no longer the limit for optical networking. Coherent optical technologies originally developed for terrestrial networks are being adapted for space‑based systems, enabling resilient, high‑capacity connectivity where fibre is unavailable—or where additional physical diversity is required.

Emerging standards are enabling interoperability across ground, airborne, and space assets, allowing optical networks to seamlessly extend to in‑orbit satellite constellations.

The broader picture is clear. With hyperscalers investing at unprecedented levels in AI infrastructure, scale itself has become the dominant driver of innovation. Optical communications and networking have never been more exciting.

Gazettabyte is asking industry figures for their thoughts after attending OFC 2026 in Los Angeles. In Part 2, LightCounting’s Vladimir Kozlov,  Chris Cole, and Aloe Semiconductor’s Chris Doerr share their thoughts.

Vladimir Kozlov, CEO of LightCounting

The resurgence of interest in near-packaged optics (NPO) surprised me the most at OFC. Meta and Oracle hinted that they will deploy near-packaged optics first, but co-packaged optics (CPO) will be the ultimate solution.

There are rumours that suppliers are placing large-volume orders for near-packaged optics. The Consortium of On-Board Optics (COBO) was ahead of its time, but only by a few years. Will Microsoft deploy near-packaged optics this time? And why did they give up on COBO? We could have been on the 3rd or 4th generation of COBO modules by now, and this is what it takes to get the technology right.

New technology continues to surprise as well. Polarisation management in silicon photonic chips, shown by Silith, and the low-voltage method for tuning micro-rings, shown by NewPhotonics, are good examples here.

With all the attention on co-packaged and near-packaged optics, the pluggable community, led by Arista’s Andy Bechtolsheim, is struggling to impress.

Deployment of linear pluggable optics (LPO) remains limited. The 1.6-terabit Linear Retimed Optics (LRO) may see more deployments this year, but who knows?

Andy Bechtolsheim rushed in to introduce the XPO form factor, focusing intensely on marketing it and promoting its success.

But XPO only supports 200 gigabit-per-lane electrical signalling. I was hoping to see a 400 gigabit-per-lane solution as well. This is how the OSFP form factor gained market acceptance: it started as a 50 gigabit-per-lane alternative to QSFP-DD and became the dominant solution for 100 gigabit-per-lane modules.

Can XPO follow this strategy by starting with 200 gigabit-per-lane and blossoming to 400 gigabit-per-lane?

Lastly, several suppliers commented that they reached a point at which they could no longer accelerate.

We have been accelerating for three years now, and it shows no signs of slowing down. If anything, there is pressure to move even faster.  Are we reaching a breaking point when customers are going nuts and suppliers turn into zombies?

Are we starting to make irrational decisions? I hope not, but it is never too early to raise this question. Can we stop, take a deep breath, and take a sober look at the situation? Let us not get carried away with the excitement.

I feel that LightCounting is the designated driver at this party.

Chris Cole, Optical Communications Advisor

My favourite OFC quote was during Sunday’s AI interconnect Session from Anthony Torza of Cisco who commented that co-packaged optics integration is inevitable but not imminent.

My favourite insight was from Rebecca Schaevitz of MixxTech, which she first shared at DesignCon, that co-packaged optics will become mainstream like ASICs when optics packaging meets JEDEC standards, most importantly by eliminating epoxy.

Volume optics deployment became clear during OFC. Next is 200 gigabit-per-lane PAM4 full and half-retimed pluggable, with innovation in density and cooling coming from new form factors like XPO and OIF High Density.

Initial co-packaged optics deployment will be fully linear as part of complete box solutions, with the ramp in volumes likely to be modest. The next generation of pluggable and co-packaged optics will double the rate to 400 gigabit-per-lane PAM4 with half-retimed pluggable and linear co-packaged optics looking feasible.

It is not clear where the market is for the multitude of optics companies not in the few closed ecosystems like Broadcom’s and Nvidia’s.

Meta continued to set the bar for the minimum required to demonstrate that integrated optics are real. They published more comprehensive traffic and reliability statistics of Broadcom’s co-packaged optics integrated with a switch in boxes running in racks, adding to previous results presented at ECOC and OCP events.

The Optical Compute Interconnect Multi-Source Agreement (OCI MSA) represents a return to fundamentals for the optics industry by specifying non-return-to-zero (NRZ) modulation and dense WDM bandwidth scaling. Optical PAM4 was adopted despite Shannon’s capacity theorem telling us to do otherwise. However, given the enormous investment in ASIC SerDes, OCI is likely to remain niche until after the 400-gigabit PAM4 generation. Cisco’s Torza’s quote applies here; inevitable but not imminent.

The excess hype surrounding integrated optics confirmed that we are in an investment bubble. There are too many investment dollars chasing too few real optics technologies. Most investments have little chance of leading to useful products.

Transformational optics require new process development, which is frightfully expensive and takes a long time, a brutal reality that doesn’t make for an exciting pitch deck. Start-ups know this and spin fantasies for which they are richly rewarded, the less plausible the higher the investment dollars.

For investors, this is the correct strategy because a fantastic story is much more likely to get next-round investors excited. The trick is not to be the one left standing when the music stops. A perennial example continuing at OFC is optical computing, an illusion supported by bad mathematics and optical communication advantages irrelevant in computing..

Chris Doerr, CEO of Aloe Semiconductor

I had four take-aways from OFC, and they all start with the letter “c”: chemistry, coherent, co-packaged optics, and consortium.

Chemistry: materials with heavy elements are on allocation, such as the indium phosphide wafers used to make lasers and modulators. This is because of high demand for optical transceivers and broken supply chains due to trade wars. This world-wide shortage of heavy-element materials is driving strong interest in silicon photonics.

Silicon’s atomic number is only 14 and thus in very plentiful supply.

Coherent: in the constant battle between intensity-modulation direct-detection (IM-DD) and coherent, coherent is suddenly starting to win. This is driven by unexpected forces, such as optical circuit switches now used in data centres.

Interestingly, coherent is the antithesis of “slow and wide”, and coherent seems to be winning.

Co-packaged optics (CPO): CPO has become red hot. The CPO workshop on Sunday was in a huge room yet was so crowded that many people had to stand outside. However, CPO is more confusing than ever. It is not clear if it is for scale-up or scale-out or both architectures. Also, the introduction of giant, liquid-cooled pluggables, such as the XPO, make CPO less compelling.

Consortium: multiple new consortiums were introduced at OFC: the XPO, the Open CPX MSA, and the OCI MSA. This trend seems to show that needs are so urgent that traditional standards bodies have to be bypassed by self-aggregating groups of companies.

On other topics, a new phrase was breathed into life at this OFC: “scale-across”. It has become a triumvirate with “scale-up” and “scale-out”. Scale-across mainly addresses the disaggregation of computing centrrs due to power supply and/or cooling constraints.

Lastly, OFC 2026 showcased that 200-plus gigabaud optics is a reality, with multiple demonstrations in all platforms, including silicon photonics.

Gazettabyte is asking industry figures for their thoughts after attending OFC 2026 in Los Angeles. First contributions from Maxim Kuschnerov of Huawei, Broadcom’s Near Margalit, and Corespan’s William Koss.

Maxim Kuschnerov, Senior Director R&D at Huawei

OFC 2026 stood out as the most substantial and extraordinary OFC in recent times.

We saw a tectonic shift as topics that had been developed and discussed for years finally became reality. Optics for AI scale-up went from a possibility to a commitment with the Nvidia Feynman generation and will be standardised in a new multi-source agreement, the OCI MSA, that proposes 4x50G non-return-to-zero (NRZ) signalling using micro-ring resonators.

New form factors received significant momentum, with Nvidia firmly betting on co-packaged optics (CPO) for AI scale-up and scale-out architectures, while Arista spearheaded a next-order-of-magnitude increase in pluggable modules through its XPO MSA.

The XPO MSA features a 4x density increase and liquid cooling, potentially enabling operation beyond 400W per pluggable.

Lastly, 400G per lane optical PAM4 materialised with Broadcom’s Taurus DSP, with first-module vendors showing very good performance, supporting the assumption of KP4 forward-error correction (FEC).

The OFC technical sessions and exhibition showcased a wide variety of material options for optical modulators.

The firm Coherent demonstrated the first 400-gigabit transmission over a silicon photonics Mach-Zehnder modulator using a standard CMOS process, which on paper was supposed to signify a decisive potential win over the more niche thin-film lithium niobate and indium phosphide for high baud rates. However, the optoelectronic bandwidth was only 70GHz, and performance was worse than that of their electro-absorption modulated laser (EML).

Conversely, Nvidia’s 200G per lane DR8 module based on its micro-ring resonators demonstrated superior performance, with error floors down to 1e-14, surpassing the EML solution.

The most likely realisation of co-packaged optics for scale-up will be silicon photonics micro-ring resonators, but VCSELs will put up a good fight.

High-temperature tolerance 2D VCSEL arrays were demoed at the show, and the supply chain is broad enough to support a solid alternative to silicon photonics. The 1060nm single-mode devices over multi-core fibre could indeed be an interesting solution, enabling longer reach.

Technical papers at the show detailed highly interesting results on 400G per lane barium titanate (BTO), lithium tantalate, and even graphene, which made a comeback at lower rates. Although these alternatives pose no immediate threat to the supply chains for thin-film lithium niobate (TFLN) and indium phosphide, they represent significant achievements.

Coherent transceivers also stand to benefit from next-generation form factors like XPO, enabling half-band modules in which four pluggables cover the entire C+L band. Ciena’s multi-rail concept, with dense inline amplifiers and full C+L-band transponder cards, pushes this concept even further, enabling a seamless rollout of long-haul capacity, fibre by fibre, without the need for wavelength-selective switching in these high-capacity, AI-driven backbone networks.

Optical circuit switches are the final technology to emerge from decades of research and niche markets. Optical circuit switches are becoming a high-impact product with significant backlogs and deployed across various use cases, such as spine switching, scale-up pods, and potentially even long-haul fibre-based protection switching.

Near Margalit, President and General Manager of the Optical Systems Division at Broadcom

At OFC, what stood out to me was how much Nvidia has invested in the co-packaged optics supply chain.

Additionally, I learned that lithium niobate is ready for 400 gigabit per lane. However, no one had a really good OFC 400 gigabit link demo.

William Koss, CEO of Corespan Systems

What a difference comparing my last visit, at OFC 2024, to this year’s OFC. There was so much more energy around this time.

Co-packaged optics or the death of copper was a big topic. Copper is not going away anytime soon, but the trend towards all-photonic solutions is emerging.

There is still a long way to go with co-packaged optics. I visited several suppliers, and there wasn’t much new, as they are still trying to get what they have into high-volume production.

Co-packaged optics is hard. There is science and tradecraft in knowing how to use it.

OFC 2026 was also the year of the optical switch. I visited seven companies with optical switches at the show: Coherent, Lumentum, iPronics, Huber-Suhner (Polatis), Calient, Accelink and Salience Labs. I even met Ming Wu at the show, founder of nEye Systems, so that’s eight optical circuit switch vendors I know of. All the various technology options were present.

I thought Arista’s Andy Bechtolsheim’s announcement of a liquid-cooled pluggable (XPO) was the most surprising news. AI has become so power hungry that we are using liquid cooling for pluggable optics!

Nvidia has had such a large impact on the optical industry that vendors are now quoting “scale-across” alongside scale-up and scale-out.

I clearly see the benefits of hollow-core fibre (HCF). It becomes a very nice fibre plant for the data centre and can be used for PCI Express (PCIe) as SerDes speeds range from 32G to 64G to 128G.

Distance is better with the loss budget, and hollow-core eliminates the need for coherent systems inside the data centre.

That might be a surprising revelation to watch: hollow-core fibre versus coherent optics in the data centre.

NewPhotonics has unveiled a near-packaged optics (NPO) product, broadening its optical engine portfolio.

Near-packaged optics refers to an optical engine placed close to the main processing chip (ASIC) on the host board, rather than on the rack’s front panel like traditional pluggable modules.

NewPhotonics, a fabless semiconductor start-up based in Tel Aviv, also demonstrated an early co-packaged optics (CPO) design at the OFC show held in Los Angeles earlier this month. Co-packaged optics refers to integrating an optical engine into the same package as the main processor chip, effectively adding optical input-output (I/O) to the chip.

The start-up views near-packaged optics as a way to enable customers to enhance system power efficiency and signal integrity while simplifying integration into existing architectures. In comparison, co-packaged optics requires more substantial changes to the hardware.

Focus on pluggables

Despite the industry’s growing interest in tightly integrated optical architectures for AI systems, NewPhotonics stresses the importance of pluggable transceivers.

“It’s an existing, very fast-growing market and a strong business to be in,” says Doron Tal, Senior Vice President and General Manager of Optical Connectivity at NewPhotonics.

NewPhotonics highlights the large deployment of OSFP pluggable modules by hyperscalers building AI clusters. Pluggables remain a standards-based market, defined by interoperability and established supply chains – all factors that favour steady, incremental improvements.

Now, by adding the near-packaged optics product and following the co-packaged optics demonstration, NewPhotonics is showing it is addressing the full spectrum of optical interconnects.

Integrated laser PICs

NewPhotonics has developed a toolkit of novel technologies to differentiate its photonic IC (PIC) products.

The company has adopted heterogeneous integration IP from OpenLight, now available as a process design kit at foundry Tower Semiconductor, to embed lasers into its pluggable module PICs.

Conventional 1.6-terabit optical modules typically use external lasers, modulators, lenses, isolators, and multiple optical coupling interfaces. Each element adds complexity, loss, and manufacturing overhead, says NewPhotonics.

With a PIC optical engine, there are fewer assembly steps, lower coupling loss, reduced system power consumption, and streamlined manufacturing.

“In semiconductors, integration is key, and that is exactly our focus,” says Tal.

NewPhotonics has two product PICs for 1.6-terabit optical modules. One, the NPG10204, is for digital signal processor (DSP)-based optical modules. It integrates eight lasers and eight modulators for use in a 1.6-terabit DR8 module, with eight electrical input channels, each at 224-gigabit PAM4.

A full DSP-based module consumes 28W, while a half-retimed module (also known as Retimed Transmitter Linear Receiver, RTLR) consumes 23W. A RTLR module using NeoPhotonics’ PIC, the power is 15W.

The second 1.6 terabit PIC is the NPG10203, for a 224-gigabit signalling linear pluggable optics (LPO) design that uses no DSP. These toolkit capabilities are the basis for NewPhotonics’ LPO+ product, which embeds NewPhotonics’ optical signal processing engine to address signal impairments. The resulting 1.6-terabit pluggable module consumes fewer than 10W.

Centera Photonics is the first announced customer to use the LPO+ engine.

Optical domain equalisation

In conventional high-speed links, equalisation is performed electronically using a DSP for feed-forward equalisation (FFE) and continuous-time linear equalisation (CTLE). Decision feedback equalisation (DFE) can also be used to tackle non-linear distortions.

“What is common for all these methods is that they are all some kind of filter,” says Professor Yosef Ben-Ezra, CTO and co-founder of NewPhotonics.

NewPhotonics, on the other hand, performs signal repair in the optical domain. The merits of the approach, says Ben-Ezra, are that it is very fast, reduces power consumption, and is not limited by the data rate.

Tellingly, the optical signal processing compensates for impairments across the channel, not just electrical impairments but also those introduced by the optical front end.

“With electronic equalisation, only part of the channel is visible,” says Ben-Ezra. “With our optical signal processor, the entire channel is equalised.”

NewPhotonics says its compensation approach restores up to 8dB over the longest channels.

The optical signal processor is programmable and adaptive, with monitoring and control circuitry based on another of the start-up’s novel technologies.

Called Niox, it enables the monitoring and control of optical circuits without using photodetectors to tap off light signals. Using Niox, the system can compensate for channel variations in real time, making maintenance easier and improving system stability.

NewPhotonics says its optical signal processor is so efficient at restoring margin that its LPO+ modules will work with all other vendors’ LPO modules. No ‘bookended solution’ is needed, says Tal.

“With the first generation of LPO [at 100Gbps] it was ’remove the DSP and pray,’” says Tal. “What we’re doing now [at 200Gbps] is making LPO interoperable.”

NewPhotonics has also announced a 3.2-terabit DR8 PIC that supports 448 gigabit per second (Gbps) signalling, with sampling scheduled for year-end. The PIC can be used for modules and for pluggable sockets.

Near-packaged optics: a bridge architecture

NewPhotonics’ move into near-packaged optics reflects a broader industry search for architectures that reduce power and improve signal integrity without the full disruption of co-packaged optics.

For Tal, near-packaged optics are a pragmatic solution on the pathway to co-packaged optics.

The key advantage, he says, is that near-packaged optics can be used without redesigning the entire system. Unlike co-packaged optics, which require tight co-design of optics and the ASIC, near-packaged optics maintains much of the existing electrical architecture, enabling easier upgrades and faster deployment.

NewPhotonics says it already has customers using its near-packaged optics solution, dubbed the NPC50503, a 1.6 terabit photonic engine.

The PIC builds on the same platform as its pluggable products. It combines integrated lasers, modulators, and its optical signal processing capability into a compact optical engine. Multiple near-packaged 1.6-terabit optics engines can be used alongside a switch or an AI accelerator.

While the electrical environment is simpler than in pluggable systems, challenges remain, including signal integrity across substrates and interconnect layers. Here, NewPhotonics also uses optical signal processing.

Luma Opticore demonstration

NewPhotonics demoed at OFC a co-packaged optics engine, offering a glimpse of its longer-term development work.

Co-packaged optics promises improvements in bandwidth density and power efficiency, but it also introduces significant challenges in packaging, thermal management, and system design.

NewPhotonics’ co-packaged optics demonstration differs from its pluggable and near-packaged optics products in that the engine uses an external laser source.

This reflects a broader industry view that external lasers are more practical in co-packaged optics systems, where thermal constraints and reliability requirements differ from those in pluggable modules.

The company says the design aligns with the newly announced Optical Compute Interconnect MSA (OCI MSA) introduced at OFC by founding members Meta, Microsoft, OpenAI, AMD, Broadcom, and Nvidia. The OCI MSA supports multiple wavelengths per fibre.

NewPhotonics’ design uses micro-ring resonator modulators to support five wavelengths per fibre.

Moreover, the micro-ring modulators do not require thermal control, simplifying the overall design. “The resonators are controlled but not thermally,” says Tal.

The co-packaged optics engine’s electrical input supports 53-gigabit non-return-to-zero (NRZ) and 112-gigabit PAM4 data rates, while its next-generation design will support 224Gbps and 448Gbps rates.

For the optical output, there are five wavelengths per fibre and twenty fibres overall, referred to as a tile. Each engine has five tiles, a total of 100 ring resonators. If each wavelength carries 112 Gbps, that is 11.2Tbps in total. NewPhotonics refers to its co-packaged optics engine as the 10-terabit Luma Opticore.

In the OFC demo, the start-up showed the Luma Opticore on a glass substrate with five fibres in and 20 fibres out.

It also demonstrated its micro-ring resonator modulator operating at 53 Gbps, with error-free operation despite temperature variations.

NewPhotonics is essentially betting that it will take longer for co-packaged optics to become established than many expect. For the start-up, the opportunity lies in improving what exists today while quietly preparing for what comes next.

The OIF has a range of demonstrations lined up for this week’s OFC show in Los Angeles.

• 448-gigabit electrical interconnect, and vendors showing more module offerings and interoperability at 224-gigabit electrical signalling.

• Coherent optics: 800-gigabit pluggable modules operating over the L-band.

• The first multi-core fibre modules that are part of an OIF OFC demo: 800G DR, and coherent ZR sent over multi-core fibre.

CEI-448G steps into view

A main OFC showpiece will be a live 448 gigabit-per-second (Gbps) electrical signalling demonstration, part of the OIF’s Common Electrical Interface specification work (CEI-448G).

The goal is to showcase 448-gigabit electrical transmissions using advanced test-and-measurement gear and 448-gigabit silicon.

The test and measurement equipment for 448 gigabit is huge, power-hungry, and noisy, whereas the 448-gigabit silicon uses the latest CMOS process node and is tiny.

“It is like a refrigerator versus a penny, and they’re both doing the same thing,” says Mike Klempa, the OIF Physical and Link Layer (PLL) Interoperability Working Group Chair, and of Qualcomm (Alphawave Semi).

The 448-gigabit demo will showcase different signalling options, including PAM4, PAM6, and PAM8, along with measurements such as bit-error rates and eye diagrams.

At 224 gigabits, the demos will emphasise the ecosystem’s broadening and growing maturity. The emphasis is on growing interoperability across a range of implementations: silicon, modules, linear-drive pluggable optics, direct-attach copper (DAC) cable, active copper cable (ACC), and retimed and unretimed approaches.

Coherent optics, L-band and multi-core fibre

For the OIF’s 800-gigabit coherent optics interoperability demonstration, a notable addition will be support for the L-band for the first time.

Multi-core fibre will also be part of the demos.

“We will also use the multi-core fibre in one of our [coherent] ZR links,” says Karl Gass, OIF Physical & Link Layer Working Group Optical Vice Chair, another first.

In turn, 800-gigabit DR (Data Centre Reach) intensity-modulation direct-detection (IMDD) transmission will be shown working across multi-core fibre.

The OIF does not define multi-core fibre but hyperscaler interest means it will be part of the OIF demonstrations.

First CEI-448G projects now underway

The OIF has also announced that it has kicked off the first two CEI-448G specification projects: 448-gigabit very short reach (VSR) and long-reach (LR) interfaces.

“We have officially started work, but we have not got to where we are defining specific parameters, like loss budgets, modulations, and jitter targets,” says Nathan Tracy, OIF President and Technologist, System Architecture Team at TE Connectivity.

The VSR interface refers to the chip-to-optical-module link that spans anywhere on the host board to a chassis faceplate. Or as Tracy puts it: “Getting the signal out of the box.”

The LR interface refers to the link between line cards via a rack’s backplane and even between racks. In AI architectures, this is referred to as the scale-up architecture.

“LR needs to be at least one meter, but we’re fighting with all these extra package losses and higher insertion loss channels, and that is challenging,” says Klempa, adding that CEI-448G LR will require additional techniques and tools compared to what is used for CEI-224G LR.

White papers aim to narrow the design space

The OIF has also recently published two White Papers. One details the OIF’s energy-efficient interfaces; a framework for categorising the interface types.

The second White Paper, a 50-page document, is titled “Compute Optics Interface (COI): Energy-Efficient Photonic Interconnects for AI Compute Scale-up.”

The purpose of the documents is to guide the industry in identifying more energy-efficient hardware architecture solutions for AI clusters. The OIF also wants feedback to inform its future specification work.

“AI is moving very quickly, and there are many factors, of which energy efficiency is just one,” says Jeff Hutchins, OIF’s vice chair of the Energy Efficient Interfaces efforts, and of Ranovus, and one of the reports’ authors.

Despite the hyperscalers’ need to move quickly, they also recognise the value of standardisation. The OIF has published the reports to highlight key issues and help implement standards.

The Compute Optics Interface White Paper details the trade-offs in optical-scale-up interconnect but does not declare a winning approach.

What surprised the authors

The documents took a year to create. Did the work lead to any surprising insights?

Hutchins highlights bi-directional optics, saying he had not fully appreciated its benefits in terms of cost savings and reduced fibre count.

Also, the importance of latency: not just raw one-way latency, but what happens when communication goes wrong, such as retransmissions and error handling. Here, turnaround time is crucial in tightly synchronised AI accelerator systems.

One debate to be resolved is the choice between slow-and-wide links and fast serial ones for scale-up. What should be asked here is what can be built in volume, reliably, and in time, says Hutchins.

The OIF hopes the documents will lead to smaller-group convergence and even a multi-source agreement (MSA), followed by a more formal standards activity once the most practical path becomes clearer.

CMIS security additions

The OIF also recently approved security additions to the Common Management Interface Specification (CMIS).

The goal is to add trust to the links that CMIS uses to set up pluggable modules for authentication, authorisation, and resistance to interception as part of the CMIS commands.

AI is forcing both pragmatism and engineering excellence

Tracy says this period, driven by AI’s scaling needs, is remarkable but not unique. There have been other times when market dynamics have driven developments that meet the needs yet may not use the most elegant engineering solutions.

“We could talk for months about the most optimal way to do things, but there is some brute force going on where that is good enough,” says Tracy. “That said, there is some very exquisite engineering going on to solve some really hard problems that this market faces.”

Tracy cites as an example the CEI-448G work: “It will absolutely be exquisite by the time we’re all done with it.”

The new signalling rate is going to bring with it challenges that, for Tracy, represent an intersection of several factors.

AI cluster sizes are getting bigger, requiring longer reach, yet signalling rates are doubling and higher bandwidth densities are required.

All this raises challenges: rising energy consumption, thermal management issues, and delivering greater functionality over management buses.

“It is the intersection of all the things that the OIF does,” concludes Tracy. “And when we bring all of those challenges together, that is where very careful engineering must occur.”

Adtran has announced an 800-gigabit linear pluggable optics (LPO) transceiver designed for short-reach data centre links.

The LiteWave800 is an OSFP 800G-DR8 module with a reach of 500 metres that consumes less than 1 watt of power, enabled by the use of 100-gigabit single-mode VCSEL lasers.

“We have partners that we have been working with on single-mode VCSELs, and the devices are going to be part of our LPO platform that we’re announcing at [the upcoming] OFC [show],” says Saeid Aramideh, Vice President of Business Development, Optical Engines business unit at Adtran.

AI workloads and the push for LPO

“AI workloads are driving everything, and these require much higher interconnect bandwidth than what we are typically used to seeing,” says Aramideh.

Not only are AI clusters driving the need for higher interconnect speeds, but the designs are also raising issues of link reliability, latency, system cost, and energy efficiency.

“Power consumption has become the most critical limiting factor in large-scale data centres, and is directly impacting rack density: how many pluggables you can put in there, cooling requirements and so on,” says Aramideh. “Ultimately, the issue of power is going to impact the economic viability of the infrastructure.”

The attraction of LPO is that it uses an optimised electrical-optical interface with the host IC driving the optical link. The approach eliminates the need for the pluggable module to have its own digital signal processor (DSP), a key contributer to the overall power consumption.

First-generation LPO modules consume between 5–8W, compared with well over 10W for conventional DSP-based 800-gigabit pluggables, says Aramideh.

Adtran set itself an energy consumption target of 1pJ/b for its LiteWave800 design. At 800 gigabits-per-second (Gbps), an energy efficiency of 1pJ/b corresponds to roughly 0.8W total module power, significantly lower than conventional DSP-based modules.

Design approach

To achieve the ambitious power target, Adtran used its in-house electronics IC design team, along with using single-mode VCSEL technology operating in the 1310nm band for the optics.

“The choice of single-mode VCSELs in the module is the most essential defining factor when it comes to reducing the power,” says Aramideh. Using EML lasers or a silicon photonics modulator approach yields much higher energy numbers, ranging from 5-15pJ/b.

Multi-mode VCSELs are an established industry technology. Developing single-mode VCSELs is challenging, as is creating a supply of such devices. But the laser device has advantages such as its low drive voltage and good linear performance.

Adtran has also been using single-mode VCSELs for its 10×10 gigabit MicroMux pluggable product and says it has two supply partners for its VCSELs.

Adtran also decided to make its own electronics as typical driver and trans-impedance amplifiers in the pluggable consume up to 3pJ/b.

“But low power is essentially meaningless if you cannot have good signal integrity,” says Aramideh. Given that signal integrity is the biggest challenge with an LPO design, Adtran’s focus was on how to close the loop between the optics and the host driver IC.

Making use of the OIF’s Common Management Interface Specification, or CMIS, Adtran created a set of controls known as the Versatile Control Set (VCS).

“VCS lets you manage certain attributes on the host DSP, and these attributes set the operating margin and tune parameters to coordinate the signal integrity between the module and the host, and control them dynamically,” says Aramideh.

LiteWave800 architecture

The 800-gigabit module has eight 100-gigabit channels, each operating at a 53-gigabaud symbol rate and the PAM4 modulation format.

The drivers interface to the VCSELs which also use a thermal-electric cooler. However, based on testing results, such coolers may be optional depending on the pluggables’ environmental conditions. On the receive side of the pluggable, there are photodiodes and two quad trans-impedance amplifiers (TIA).

The expertise Adtran brings is in controlling the single-mode VCSEL, interfacing the drivers and the VCSELs, and the overall link control.

Roadmap beyond 100 gigabit-per-lane

One question facing the industry is whether LPO architectures can scale to 200 gigabits-per-lane, where tighter electrical margins may require alternative approaches. Is Adtran’s LiteWave design thus a one-generation product?

“There is a roadmap, but it may not necessarily be LPO,” says Aramideh. He cites the use of near-packaged optics for next-generation AI clusters, where the optics are brought closer to the host IC.

Using near-packaged optics based on single-mode VCSELs and linear electronics and addressing the challenges of serviceability and thermal coupling might prove a better approach, argues Aramideh.

“Can we get to 200 gigabit per lane? The answer is absolutely, and we must. This is what industry is asking for,” says Aramideh. “Would the answer be an LPO? At this point, that is a question mark and maybe near packaged optics would prove a better solution.”

Market implications

Aramideh hopes there will be an uptake of single-mode VCSELs, leading to wider adoption of the technology in the data centre.

He also believes that the LiteWave800 design benefits pluggables overall by offering an attractive alternative to the low-power argument used by co-packaged optics proponents. Such designs typically consumes 5pJ/b.

“Having a product at 800 gigabit that hits one picojoules per bit, that sets the industry standard,” says Aramideh.

The product will be available as samples in the third quarter of this year and will be in production early 2027.