Network-as-a-Service: The Infrastructure Layer Beneath the AI Boom

The public narrative around AI infrastructure focuses almost exclusively on compute — GPUs, data centres, power. This is understandable: the GPU shortage of 2023 was the most visible bottleneck in the AI buildout, and the data centre construction boom is the most capital-intensive expression of AI demand. But there is a layer of infrastructure beneath the compute layer that is equally essential and considerably less discussed: the networking fabric that connects AI compute clusters to each other, to the internet, and to the enterprise systems that consume AI services.

Training a large language model requires the simultaneous coordination of thousands of GPUs, each communicating with the others at extremely high bandwidth and extremely low latency. A single training run for a frontier model can generate petabytes of inter-GPU traffic. The networking infrastructure required to support this — high-speed interconnects, programmable switches, intelligent traffic management — is a specialised market that is growing at rates that rival the GPU market itself. And unlike GPUs, which are produced by a handful of large semiconductor companies, the networking layer is served by a more fragmented ecosystem of specialist operators and equipment manufacturers.

Network-as-a-Service (NaaS) — the delivery of networking capabilities as a programmable, on-demand service rather than through dedicated hardware — is emerging as the dominant model for AI infrastructure networking. NaaS operators provide the software-defined networking platforms that allow AI workloads to be routed dynamically across distributed infrastructure, optimising for latency, bandwidth, and cost in real time. For hyperscalers and large enterprises running AI at scale, NaaS is not a convenience — it is a necessity.

At Lensbridge, our investment in PacketFabric — one of the leading NaaS operators globally — reflects our conviction that the networking layer is a critical and underappreciated component of the AI infrastructure stack. PacketFabric's programmable network connects over 60 data centres across North America and Europe, providing the high-bandwidth, low-latency connectivity that AI workloads demand. As the AI buildout continues, we expect the NaaS market to grow significantly — and we are well-positioned to benefit from that growth.