AI workloads are changing what organisations need from their infrastructure. Much of the public conversation focuses on compute (or computing power): GPUs, accelerators and the power to run them. But keeping those systems productive also depends on something less visible: the networks that move data between processors, storage and, where required, cloud environments. Fibre is a critical part of those networks. As AI workloads grow, networks built for earlier requirements can struggle to deliver the bandwidth and low latency they demand, limiting the performance of expensive compute infrastructure. Understanding where legacy networks fall short, and why, is now part of any serious AI infrastructure strategy. And it’s bringing dark fibre into focus.
What makes AI workloads different
AI workloads don’t behave like the enterprise applications most networks were built to carry. Large-scale distributed training and real-time inference can place demanding requirements on network bandwidth, latency and consistency.
Data movement at scale
Training and running AI models means moving large datasets between GPU clusters, storage platforms and cloud environments, often continuously. Within a data centre, high-performance interconnects such as InfiniBand link GPUs and nodes together. Between facilities, that job falls to fibre. In a global survey of more than 1,300 data centre decision makers, 81% said they expect large language model training to take place across distributed data centre facilities.
Data gravity adds to the pressure. As datasets grow, they become harder and slower to move, so the network connecting them has to deliver the throughput to keep AI pipelines running without constant re-engineering.
Inference is becoming the dominant workload
Training gets the headlines, but inference is where most AI compute is heading. McKinsey projects that by 2030, inference will represent more than half of all AI compute. Gartner forecasts that in 2026, global spending on AI-optimised cloud infrastructure for inference will overtake spending on training.
That changes the demand on the network. Training runs are large but scheduled. Inference is continuous, distributed and latency-sensitive: AI systems and AI agents answering requests in real time, from wherever users and data happen to be. The network needs to deliver predictable performance around the clock, not just capacity for occasional peaks.
Hybrid deployments are the norm
Few organisations run AI in a single environment. Gartner predicts that 90% of organisations will adopt a hybrid cloud approach through 2027. AI workloads increasingly span public cloud platforms, private data centres and edge locations, which means more data moving between more sites, and more points where the network can slow everything down.
GPU utilisation depends on the network
GPUs and other accelerators are among the most expensive hardware in an AI deployment. In distributed training, GPUs need to exchange data frequently as well as access training datasets and storage. Insufficient bandwidth, congestion or excessive latency can leave accelerators waiting, reducing utilisation and slowing results. Where workloads depend on connections between facilities, the performance of those links also matters.
Where legacy fibre networks fall short
Much of Australia’s existing metro fibre was built long before AI workloads existed. FibreconX’s assessment of the fibre landscape found that wholesale fibre options have barely changed in 20 years, with a lack of capacity where it’s most needed: between data centres. That gap shows up in four ways.
Capacity planned for a different era
Legacy networks were sized for the traffic patterns of their time. Demand between data centres is now growing far faster. In the Ciena global survey, data centre decision makers anticipated at least a sixfold increase in data centre interconnect bandwidth demand over the five years to 2030. More than half (53%) expected AI workloads to place the biggest demand on that infrastructure over the following two to three years. Networks with limited capacity headroom can require upgrades before they can support growing AI demand.
Longer, indirect routes add latency
Retrofitted networks often follow the paths of older infrastructure rather than the most direct route between facilities. Every extra kilometre adds latency, and for latency-sensitive AI workloads, those delays add up across every request, every training step and every data transfer. The shortest path between facilities is a design decision, and it’s hard to add after the fact. For distributed training in particular, where nodes in different facilities need to exchange data constantly, route length becomes part of the architecture rather than a background detail.
Shared paths create resilience risk
When multiple services share the same physical corridors, a single incident can take several links down at once. For AI workloads that organisations now treat as mission-critical, availability depends on genuinely diverse routes with minimal common path, not backup capacity that runs through the same duct.
Manual provisioning slows deployment
Adding capacity on many legacy networks still means manual requests, long lead times and pricing that changes along the way. Wholesale buyers have described dark fibre as hard to request, procure and provision. When compute can be added quickly, a network that’s slow to scale becomes the limiting factor.
What this means for AI infrastructure decisions
The network is now a strategic part of AI infrastructure, not just an operational utility. In practice, that means bringing network engineering into AI infrastructure design from the start, rather than after the platform, hardware and deployment plan are locked in. Teams that treat connectivity as part of the AI architecture, alongside compute and storage, avoid discovering the bottleneck once workloads are already live, and build a more sustainable foundation for scaling AI capability over time.
For infrastructure leadership teams planning AI capability, a few requirements stand out when evaluating connectivity:
- Capacity headroom that can scale with GPU and storage growth without re-engineering
- Direct routes between facilities that reduce latency and improve throughput
- Path diversity that protects availability for mission-critical workloads
- Automation and visibility, with provisioning, monitoring and fault response that keep pace with deployment
- Flexibility to integrate public cloud, private data centres and hybrid deployments end to end
Getting these right early reduces complexity later, and helps GPU investment deliver the outcomes it was bought for rather than sitting idle waiting on data.
A network built for what AI workloads need next
FibreconX is Australia’s first new fibre network in over 20 years, purpose-built rather than retrofitted onto legacy infrastructure. FibreconX Core connects major data centres across Sydney and Melbourne with new fibre, shorter paths and direct connections between facilities, designed for resilience through redundancy and minimal common path. Capacity headroom is built in from day one and through FusionXpress, can be provisioned in as little as 40 seconds.
It’s infrastructure designed for the AI workloads running on it now, and whatever comes after them.
Plan AI infrastructure around the network
AI workloads will only be as fast, reliable and scalable as the network connecting them. Organisations planning their next stage of AI deployment need fibre that can keep up.
Talk to the our team about network coverage, capacity and connection options for AI workloads.