As AI data centers scale, the way computing resources connect is becoming increasingly important to infrastructure performance and business value.

High-performance XPUs are being deployed to support AI training and inference, but their effectiveness depends not only on compute capacity. The network connecting those resources can influence the performance and scalability of distributed processing, the effectiveness of model training and use, the efficiency of compute resources, and how quickly organizations can turn AI infrastructure into business value.

“Intra-data center connectivity is now a strategic decision rather than a purely technical one,” said Teresa Monteiro, Director of Product Marketing at Nokia.

That decision involves trade-offs. Data center operators must balance performance and latency against power consumption, cost, flexibility, ease of deployment and scale, and serviceability. As AI systems grow, no single approach can optimize all of these dimensions.

Power is an increasingly important consideration because connectivity infrastructure consumes energy, drives cooling requirements and can affect overall data center scalability. In power-constrained environments, power saved on connectivity can instead be allocated to XPUs.

Cost also encompasses both upfront investment and ongoing operational expenses, while flexibility determines how easily infrastructure can adapt to changing requirements, traffic patterns and multi-vendor environments. Ease of deployment and serviceability can similarly affect rollout complexity, maintenance and upgrades.

These trade-offs are shaping the choice between pluggable and more integrated optical approaches.

Pluggable optics remain the most widely deployed connectivity solution inside data centers. Their modular design allows components to be upgraded or replaced independently and supports flexibility, ease of deployment and serviceability. Industry standardization also enables multi-vendor environments.

More integrated approaches, including near-packaged optics (NPO) and co-packaged optics (CPO), move optical connectivity closer to the switching silicon. This can reduce losses and inefficiencies, improve power consumption and enable greater connectivity density. However, these approaches also change how optics and switches are designed and built and reduce some of the flexibility associated with modular designs.

The broader point is that connectivity architectures involve different compromises rather than a simple progression from one technology to another.

Even within pluggable optics, different approaches balance power efficiency, latency, performance and operational requirements differently. Emerging extra-dense pluggable optics (XPO), meanwhile, seek to increase bandwidth density while retaining the modularity, serviceability and deployment model associated with pluggable optics.

More integrated architectures also introduce tighter relationships between optics and switching components. NPO and CPO can improve power efficiency and signal integrity, but they are initially more complex and can be more challenging to service because of tighter component coupling.

This points to coexistence rather than convergence. “Some environments will prioritize modularity, multi-vendor compatibility, flexibility and serviceability, leaning toward approaches that are easier to operate and evolve, even if they are less efficient at the extreme. Others will prioritize performance and efficiency at scale, potentially using tighter integrated environments for the largest AI data centers where every watt and every microsecond counts,” said Monteiro.

Those priorities can also differ among enterprises, cloud providers and specialized AI operators, depending on their cost structures, time-to-market requirements, and operational models.

As a result, new connectivity technologies are adding to the available options rather than simply replacing existing approaches. Pluggable optics continue to evolve in speed, density, and efficiency, while connectivity is also moving closer to computing and switching silicon.

“The challenge is not choosing a winner architecture but selecting the right fit for a given scenario. Success will depend on making informed trade-offs that align with business goals and operational realities,” Monteiro added.

Intra-data center connectivity is consequently becoming more than a technical layer in AI infrastructure. As AI systems scale, it is becoming a strategic lever shaped by the specific requirements of each environment. Find out more.