As AI models continue to grow in size and complexity, so does the infrastructure required to support them. Higher-performance GPUs, larger clusters, and ever-increasing bandwidth demands are pushing traditional network architectures toward their physical limits.
Long-standing copper interconnects, for instance, are increasingly struggling to deliver the speed, distance, and power efficiency that next-generation AI environments require.
This is where co-packaged optics (CPO) is beginning to reshape the conversation.
By integrating optical and electronic components more closely together, CPO significantly reduces the distance signals need to travel, enabling lower power consumption, greater bandwidth density, and improved reliability.
As momentum builds behind CPO, however, important questions remain: is the technology mature enough for large-scale AI deployments? Can manufacturing processes support the volumes that hyperscale data centers will require? And is the broader ecosystem ready?
Against this backdrop, in a recent DCD>Broadcast episode, Benoit Fleury, PCS commercial director at Corning, and Leo Lin, business unit director at EXFO, explore how glass-based waveguide technology and advanced wafer-level testing are helping answer these critical questions, while highlighting why collaboration across the optical ecosystem will ultimately determine CPO’s success.
Moving beyond traditional fiber coupling
Today’s optical engines typically rely on fiber array units (FAUs) to connect optical fibers to photonic integrated circuits. While the technology has served data centers well for many years, increasing channel counts and tighter integration requirements are driving innovation in this space.
Modern CPO architectures require higher fiber densities, multiple fiber types, and increasingly precise alignment. As a result, manufacturing these systems at scale becomes progressively more difficult as performance requirements continue to rise.
“The real challenge going forward is achieving extremely high levels of density at very high scale,” explains Fleury.
To address this, Corning is developing a glass substrate solution that could eventually replace conventional fiber coupling for future generation systems with densely packed glass waveguides created through an ion-exchange process.
Rather than relying on active alignment techniques, Corning’s GlassBridge enables passive alignment between the fibers and the photonic integrated circuit. The result is a solution capable of delivering both exceptional precision and a manufacturing process better suited to high-volume production.
“It’s one thing to demonstrate this technology in a laboratory or in moderate production volumes,” says Fleury. “Doing it consistently and reliably at scale is where the real opportunity lies.”
Because the waveguides are manufactured with extremely high positional accuracy, they can align directly with their silicon counterparts without extensive adjustment during assembly.
Engineering for scale
Delivering higher density is only one part of the challenge. Each and every element of the glass bridge must also meet high performance and reliability requirements.
The waveguides themselves must minimize propagation loss while maintaining strong return loss and polarization extinction ratios for external laser sources. They must also support increasingly complex routing without compromising signal quality.
Equally important is the glass composition itself: “The glass has to remain extremely robust over long operating lifetimes and under elevated temperatures. Optimizing the material itself is just as important as optimizing the waveguides,” says Fleury.
Corning has also deliberately adopted mature, well-established ferrule technology to simplify fiber attachment and improve long-term reliability rather than introducing unnecessary complexity elsewhere in the system.
Taken together, these individual engineering decisions create an optical platform designed for both performance and manufacturability at scale.
Testing as a competitive advantage
As optical systems become increasingly integrated, testing is shifting beyond a quality assurance exercise into a critical part of the manufacturing process itself.
According to Lin, testing now spans every stage of a product’s lifecycle, from design validation and wafer production through packaging, manufacturing, deployment, and ongoing operation.
For CPO in particular, identifying defective components as early as possible dramatically reduces downstream costs.
“As higher-value components are added during manufacturing, the cost of failure increases at every stage,” explains Lin. “That’s why identifying known-good dies as early as possible, ideally at wafer level, becomes extremely important.”
Wafer-level testing prevents defective photonic chips from progressing further through expensive assembly processes, improving both manufacturing yield and overall product reliability.
Unlike conventional electronic testing, however, photonic testing introduces an additional layer of complexity. Engineers must precisely couple light into and out of the photonic circuit before any measurements can be taken.
Increasingly, the solution lies in Edge coupling. This approach most closely reflects the optical interfaces used in finished products, creating stronger correlation between wafer-level testing and final system performance.
“This correlation is extremely valuable,” says Lin. “It allows defects to be identified much earlier while reducing redundant testing later in the manufacturing process.”
Achieving this consistency requires highly repeatable positioning and measurement accuracy throughout production, something increasingly enabled by specialized wafer-level test platforms.
Leveraging AI
As well as accelerating complexity across the network, AI is helping to transform how photonic devices are tested. Rather than relying solely on traditional rule-based inspection, machine learning algorithms can rapidly identify chip features, detect defects, determine orientation, and accurately position devices under a wide range of lighting conditions.
“In our photonic wafer test station we’ve demonstrated AI-based chip identification with accuracy of around 99.5 percent,” says Lin.
This improvement directly increases testing throughput while enabling more robust automation across manufacturing.
AI is also helping engineers uncover relationships hidden inside large volumes of test data. One example involves predicting RF bandwidth using relatively simple DC electrical measurements. By training machine learning models on these correlations, many expensive and time-consuming RF tests can potentially be eliminated without sacrificing reliability of the device.
Lin explains that laboratory demonstrations suggest as many as 75 percent of RF bandwidth tests could ultimately be replaced by simpler electrical measurements.
Going mainstream
Although CPO remains relatively early in its adoption curve within data centers, commercial deployments have already begun.
Initial implementations have largely focused on scale-out architectures, where optical links connect racks across large AI clusters. These environments benefit immediately from higher bandwidth density, lower power consumption, and improved transmission over longer distances.
The next phase, however, promises more challenges. Scale-up architectures, which rely on direct optical interconnects between GPUs inside individual racks, introduce an order of magnitude more optical connections.
Supporting these densities will require even greater integration while maintaining manufacturing efficiency and long-term reliability. Highly integrated systems inevitably require highly integrated development.
“This is where density starts to become huge,” says Fleury. “And where scalability becomes incredibly important.”
Across both deployment models, innovation is taking place throughout the optical ecosystem – from foundries and component manufacturers through to testing specialists and system integrators.
An ecosystem effort
Corning and EXFO view CPO as a technology that no singular company can deliver alone. Instead, successful deployment will depend on close collaboration between material suppliers, photonic foundries, testing specialists, packaging experts, equipment manufacturers, and system integrators.
Data sharing across these organizations is becoming increasingly important, particularly when correlating wafer-level testing with later manufacturing stages.
“The partners have to work together,” says Lin. “Only by sharing relevant information across the ecosystem can we begin reducing unnecessary testing while improving yield.”
Fleury agrees that ecosystem collaboration will become even more important as CPO matures: “We’re already seeing a lot of collaboration today. And I think that’s only going to increase going forward,” he concludes.
For both companies, this collaborative approach extends beyond individual products to construct a manufacturing ecosystem capable of supporting AI infrastructure at scale.
While enthusiasm surrounding CPO continues to build, both Fleury and Lin stress that the technology remains at an early stage within the data center market, with significant innovation still ahead.
For Corning, advances in glass-based waveguide technology offer a practical route towards delivering the precision, density, and manufacturability future deployments will require. Increasingly, intelligent testing powered by AI will be equally essential in delivering quality while keeping manufacturing costs under control.
The technology may still be in its early stages, but as AI continues to accelerate and the data center ecosystem aligns in preparation, the foundations for widespread adoption are rapidly taking shape.
To hear more about wafer-level testing for next-generation co-packaged optics with Benoit Fleury and Leo Lin watch the full DCD>Broadcast episode, here.
For more information on CPO technology click here. And for more from EXFO, visit the website here.
More from Corning
