As power and space constraints make concentrated data centre expansion increasingly difficult, Gonzalo Camarillo, Marketing Steering Committee Chair at the IOWN Global Forum and Head of Implementation Components at Ericsson, considers how all-photonics networks could support a more distributed model for AI infrastructure.
The UK’s AI strategy is moving from ambition into delivery. Investment is accelerating, and demand for AI-ready infrastructure continues to rise, but the physical constraints of data centre expansion are becoming harder to ignore. Power availability, grid latency and the limits of dense metropolitan build-outs are increasingly shaping where and how new capacity can be added.
These constraints matter because AI workloads are no longer static or centralised. They are scaling across distributed environments, where processing, storage and data sources are continuously in motion. In this context, network performance plays a defining role in determining what infrastructure can realistically support, rather than simply connecting fixed systems.
All-Photonics Networks (APNs) offer one possible way of addressing that constraint.
Moving beyond the limits of traditional networking
APNs transmit data using light end to end, avoiding repeated conversions between optical and electrical signals. This has the potential to enable data to move at higher speeds over distance and could change some of the assumptions that underpin infrastructure design.
Those assumptions matter because conventional network architectures are still largely built around proximity. Data centres cluster in dense regions partly because latency, jitter and bandwidth constraints can limit how effectively systems operate over distance. This has reinforced a model in which capacity is concentrated in a small number of hubs, often close to demand, regardless of whether those locations offer the best long-term efficiency or sustainability.
By improving network performance over distance, APNs could reduce some of the disadvantages associated with geographic separation. This could allow infrastructure design to move beyond proximity-led optimisation and take a wider set of variables into account, including energy availability and regional demand. Connectivity would become a more flexible part of the design framework rather than a constraint that determines it.
Enabling distributed AI at scale
This change in design logic could have direct implications for AI systems, which rely on the continuous movement of data between compute, storage and processing layers. Training large language models requires high-volume data transfers across distributed environments, while inference workloads depend on rapid responses at scale.
When networks struggle to keep up, architectures tend to centralise. Compute is concentrated in fewer locations to reduce performance loss, even when that increases energy consumption, puts strain on already constrained data centre hubs and limits flexibility in how infrastructure evolves.
APNs could support a more distributed model by enabling faster data movement across multiple locations while maintaining performance. Compute could be spread more evenly across regions, while workloads could be balanced according to energy availability, demand patterns and operational efficiency.
This would reduce dependence on geographic concentration and allow AI infrastructure to scale more closely in line with demand rather than location.
Strengthening resilience in financial services
A similar shift could be relevant to financial services, where infrastructure already operates in real time. Payments, fraud detection, trading systems and customer services all depend on continuous availability and sustained performance under load.
In this environment, resilience depends on consistency. Short periods of latency variation or degraded performance can have disproportionate operational and reputational consequences.
APNs could improve data movement between distributed data centre environments, helping primary and secondary systems remain more closely synchronised. Depending on the wider architecture, this could enable failover processes to operate with less of the performance loss typically associated with redundancy models.
That could change how continuity is delivered. Rather than relying solely on switching between active and passive systems under stress, financial institutions could operate infrastructure across multiple sites simultaneously. This may improve stability during peak demand and reduce disruption when systems are under pressure.
The result could be a more stable operating environment in which resilience is embedded within the structure of the network rather than managed primarily as a fallback capability.
A shift in how infrastructure is designed
Taken together, these developments point to a broader shift in how infrastructure decisions could be made as connectivity constraints begin to ease.
Industry initiatives, including the IOWN Global Forum, are exploring photonics-based networking and APNs as part of a wider move towards more distributed infrastructure models. This work reflects a growing view that connectivity can be treated as an enabling layer rather than a fixed limitation.
As data centre planning gradually moves away from concentration in a small number of dominant geographies, capacity could be distributed more widely, sustainably and strategically. Network limitations may no longer dictate location to the same degree, allowing infrastructure strategy to reflect a broader set of economic and physical realities.
