IT, in its current form, relies entirely on networks. From office LANs to transatlantic fibre-optic cables, to WiFi and cellular connections, no business can really carry out its day-to-day tasks without using at least one network.

Yet, with all the attention focused on AI and its vociferous appetite for GPUs, CPUs, and even storage, the network risks being overlooked.

Without both investment and new ways of engineering the network, connectivity could derail progress in digital transformation, the user and customer experience, data science, and, of course, AI.

Recent research from vendor Cisco found that three in four enterprises expect to hit the limits of their campus and branch networking capacity within two years, and they are seeing network traffic double year on year. Already, some enterprises say they are holding back on scaling up AI investments because of network constraints.

Unlocking the problem

“You have to look at where bottlenecks are,” Paul Stringfellow, CTO at IT consultants Gardner Systems, told ITPro.

“This is less an enterprise network issue, but more a backend issue. So, looking at how you ensure high capacity [connections] between CPUs, GPUs, and back-end storage.”

As Stringfellow points out, there have been advances in high-speed connectivity, including faster Ethernet speeds and RoCE (RDMA over Converged Ethernet), which allows remote direct memory access over Ethernet technology.

These, though, along with Infiniband and 400Gbps and even 800Gbps Ethernet, are mostly targeting data centres and cloud providers.

For client devices, developments in WiFi should improve both connectivity for users and for IoT devices. Technologies such as WiFi 6 and WiFi 7 – and in the future, WiFi 8 – look to increase bandwidth and provide reliable connections, even as more users or devices connect to the WLAN.

“WiFi also needs to be included in this conversation, so many users now use WiFi as the primary office connection. The idea of cabled connections is foreign in a lot of places,” says Stringfellow.

The move to cloud and SaaS-based applications, as well as widespread home working, means that much of an enterprise’s data traffic has shifted to WiFi, with users never reaching for an Ethernet cable. Effective WiFi has become essential, with poor connections directly affecting productivity.

“Enterprise Wi-Fi has evolved significantly and now carries far more than simple user connectivity,” Phil Huang, business development and field application manager for the UK and Ireland at vendor D-Link, told ITPro.

“Modern Wi-Fi networks routinely support collaboration platforms, voice and video applications, IoT devices, and a wide range of cloud-based services,” he says. “In that sense, Wi-Fi has become a strategic access layer rather than just a convenience technology.”

But, he says, there are still significant differences between wired and WiFi networks. And improvements in WiFi connection speeds are pushing bottlenecks further into enterprises’ core networks. This, in turn, is driving investments in the wired LAN too.

“Newer Wi-Fi 6E and Wi-Fi 7 access points are capable of exceeding 1Gbps throughput, meaning a standard Gigabit uplink can quickly become the bottleneck,” says Huang.

As a result, businesses are moving to both 10Gbps connections and 2.5GbE. This is moving closer to “standard” 1Gbps connections in cost. “The premium over traditional Gigabit Ethernet has reduced dramatically, making 2.5GbE an easy investment for organizations refreshing their infrastructure,” he says.

Installing 2.5GbE, and even 10GbE, is simpler than fibre-optic or even higher-speed copper LANs. But it still works well for most office deployments, at distances under 100 metres.

Campus to cloud

The bigger challenges lie both in connectivity from businesses to data centers and the cloud, and within the data center itself.

Here, growing data volumes and the need to reduce latency are prompting CIOs to look at how their networks are designed, as well as the hardware they run on.

“As SaaS and AI adoption grows, hybrid work reshapes access patterns and digital security follows users beyond traditional perimeters,” Nauman Raja, a director analyst at Gartner, told ITPro.

“Enterprise networks are no longer built around the data centre. Instead, they now gravitate around the internet.”

AI too is playing its part: expensive GPUs are a waste of money if the network cannot keep up. Network architects need to meet the high-bandwidth demands of uploading data for model training, usually to the cloud, and the low-latency needs of inference.

Organizations also need to consider how they connect their existing enterprise applications to AI models. Conventional networks are rarely built with these demands in mind.

“I don’t think we’re having a bandwidth problem. We’re mostly having a design or infrastructure problem,” Wim Remes, a technology consultant who has also run operations for a managed security service provider (MSSP), told ITPro.

“Compute and data are not always in the same place. Either we bring data to the compute or we bring the compute to the data. The paths we create on the network need to be the shortest.”

“The top-level issue, to me, is that we’ve ignored ‘the network’ for about a decade and a half,” he adds.

“Does a full-stack developer know the basics of networking? The clear answer is no. The stack in ‘full stack’ stops at the web server at best. At the same time, many a CIO also doesn’t really care about the network anymore.”

The challenge is to bring together new and emerging hardware technologies, such as 400 or 800Gb Ethernet and RoCE, with a network architecture giving the performance data-heavy, latency-intolerant applications need.

“Bad network design, such as routing internet-bound traffic through centralised data centres for software as a service (SaaS) application, or for end-user collaboration, such as video conferencing, will result in poor user experience and inconsistent access for both remote and on-site users,” Gartner’s Raja warns.

“As demand fluctuates with remote working patterns, network architectures must allow dynamic bandwidth allocations,” Raja adds.

“In many cases, smart traffic management and automated policies can reserve sufficient capacity for business-critical applications even during peak usage.”

That, though, means thinking less about the network as hardware, but as an operational issue across the business. After all, no one wants colleagues, or worse still, customers, locked out of applications because someone “tokenmaxxing” is blocking the LAN.