AI infrastructure: Planning for long-term growth

Author: Joe Peck

This article from Antonio Castano, Global Market Development Director at AFL, highlights how a combination of technologies, architecture, and deployment models is driving the next generation of physical infrastructure.

In today’s landscape, several infrastructure approaches are becoming increasingly relevant:

• Neoclouds provide specialised compute capacity

• Brownfield deployments accelerate AI implementation within existing facilities

• DCI extends AI infrastructure across multiple facilities

• CPO brings optical interfaces closer to compute and networking components

• VSFF supports higher-density optical connectivity

Each approach addresses a different requirement. However, the consistent, underlying planning challenge remains that AI deployments can change significantly between hardware generations, whilst optical cabling, pathways, and connectivity infrastructure typically remain in service for much longer periods. This means physical infrastructure needs sufficient capacity and flexibility to accommodate technologies that may not yet be widely deployed.

Designing around a single AI architecture can limit future options, particularly where increasing fibre density, distributed connectivity, or new optical interfaces require changes to the physical layer. The long-term objective is to establish an infrastructure foundation that can support successive generations of AI systems without requiring extensive physical redesign.

AI is not scaling through a single infrastructure model. As neoclouds, brownfield deployments, DCI, CPO, and VSFF connectivity reshape the data centre landscape, the physical layer must be designed as a flexible foundation that can adapt across multiple generations of AI architectures.

Approaches to the infrastructure of tomorrow

Neoclouds provide dedicated access to accelerated computing, allowing organisations to scale AI capacity without building equivalent facilities. This model can shorten deployment timelines whilst increasing demand for high-density power, cooling, networking, and optical connectivity. For infrastructure planners, the key consideration is ensuring supporting physical infrastructure can accommodate rapid changes in compute requirements.

Brownfield deployments can accelerate AI capacity by reusing existing power, cooling, pathways, and facility space. However, infrastructure designed for conventional workloads may not accommodate the fibre density required by modern AI systems. For example, an NVIDIA NVL72 rack can require up to 1,152 fibre connections. Retrofitting requires careful planning for capacity and future upgrades.

Data centre interconnect (DCI) allows AI environments to operate across multiple buildings, campuses, or locations. This approach can provide greater flexibility when capacity, power, or resilience requirements exceed what one facility can support. However, longer connections introduce additional considerations around latency, optical performance, power, and network architecture that must be addressed during infrastructure planning.

Co-packaged optics (CPO) places optical interfaces closer to compute and networking components, reducing electrical transmission distances within systems. The architecture can support higher bandwidth while changing how fibre connectivity is presented around equipment. Physical infrastructure, therefore, needs sufficient flexibility in cable routing, fibre management, and connectivity capacity to accommodate evolving optical architectures.

Very small form factor (VSFF) connectivity enables more optical connections within limited rack and panel space. Higher connection density becomes increasingly important as AI systems require greater numbers of fibres for high-speed networking. The benefit depends on adequate pathway capacity, patching space, and cable management, making VSFF part of a wider physical infrastructure strategy.

Long-term AI growth: Building flexible foundations

AI infrastructure will continue to combine different deployment models, facilities, optical technologies, and connectivity architectures. Because physical infrastructure remains in service longer than compute and networking hardware, capacity and flexibility are critical. A modular optical foundation allows operators to accommodate future AI requirements whilst protecting existing infrastructure investments.

With AI infrastructure demands scaling fast, bringing new considerations that challenge traditional data centre design, AFL’s ‘AI Infrastructure’ whitepaper series, including Architecting AI at Scale and Building AI Training Clusters at 16K Accelerators, examines these requirements in greater detail. For a deeper dive, read AFL’s blog, What Does Sustained AI Growth Mean for Data Center Fiber Infrastructure?

For more from AFL, click here.