Summary
- The ECOC demonstrator combines an addressable 850nm microVCSEL array with BizLink multi-core fibre packaging and connectivity.
- The architecture spreads traffic across many parallel lower-speed optical channels.
- The partners are targeting lower power and latency for short-reach AI scale-up connectivity, with industrialisation still ahead.
BizLink and ams OSRAM are developing a highly parallel optical interconnect architecture for AI scale-up systems, taking a deliberately different route from pushing ever-higher data rates through a smaller number of channels.
The companies are demonstrating a connectorised multi-core fibre link at ECOC 2026 in Málaga. The prototype combines a two-dimensional addressable microVCSEL array operating at 850nm with BizLink fibre-to-fibre connectivity and optical packaging.
The partners describe the approach as “wide-and-slow”: bandwidth is distributed across a larger number of parallel lower-speed channels rather than concentrated into fewer channels running at the highest possible signalling rate.
That trade-off is aimed at AI scale-up domains where large numbers of accelerators need extremely low-latency connections over relatively short distances.
Faster individual electrical and optical lanes increase bandwidth, but they can also raise signal-processing complexity, power consumption, thermal load, and equalisation requirements. Increasing parallelism offers another route if the additional fibres and optical channels can be packaged economically.
Bandwidth is only one constraint
The most demanding AI networks are now being designed around four competing requirements: bandwidth, latency, energy per bit, and physical density.
Optimising one can make another worse. Very high serial speeds may reduce the number of lanes required but demand more sophisticated electronics. Large numbers of slower parallel channels can simplify some signalling requirements while increasing packaging and connector complexity.
BizLink and ams OSRAM are effectively testing where that balance should sit for scale-up links.
The microVCSEL array is central to the concept. VCSELs are widely used for short-reach optical communication, and an addressable two-dimensional array provides multiple optical emitters in a compact package.
BizLink contributes fibre-array assembly, packaging, test, connector technology, and high-volume manufacturing expertise. Those capabilities are particularly important if a laboratory photonics architecture is to become something that equipment manufacturers can build repeatedly at commercial scale.
The packaging problem moves forward
As optics move closer to accelerators and switching ASICs, packaging becomes as important as the optical device itself. Mechanical tolerances, fibre alignment, thermal behaviour, manufacturability, connector serviceability, and yield all influence whether a photonics architecture can survive outside a demonstration.
The ECOC system remains a demonstrator rather than a generally available data centre product. The companies are using the event to validate the architecture and develop an ecosystem around it.
That distinction is important because the optical-interconnect market currently contains multiple competing approaches: faster pluggable transceivers, co-packaged optics, optical circuit switching, expanded-beam connectors, external lasers, and increasingly parallel fibre architectures.
There is unlikely to be one physical solution across every part of an AI data centre. Short scale-up links between tightly coupled accelerators have different requirements from switch-to-switch connections across a hall or long-distance links between buildings.
The wider infrastructure consequence is power. Every watt consumed by networking becomes electrical load that has to enter the facility and heat that has to be removed. At small scale, differences in link efficiency appear minor; repeated across thousands of accelerators and optical channels, they become part of the site’s power and cooling budget.
BizLink and ams OSRAM have not yet demonstrated that wide-and-slow will become the dominant answer. Their work does show why the question is being reopened: AI systems are pushing enough traffic over short distances that the industry can no longer optimise bandwidth without simultaneously accounting for latency, power, density, and the physical reality of packaging the fibres.
