AI’s strain on data centre connectivity

As the adoption of AI continues to surge globally, the data centre interconnect (DCI) networks underpinning this transformation are undergoing rapid change. A recent global survey has highlighted that data centre bandwidth demands are set to increase by at least sixfold over the next five years, driven largely by AI workloads. 

Total Telecom sat down with Chief Technology Officer International, to discuss the survey’s findings and how network operators are moving quickly to facilitate surging demand.

“Everybody got the headlines for building all the compute and the data centres[…]  but if there’s no network connecting all the good innovations that are happening there, then it would be a car without gas,” said he. 

Ubiquitous AI usage could present numerous challenges to network operators, not only in terms of sheer data traffic but also in complexity.  

“Historically, network traffic has grown at a rate of 20–30% per year,” said he. “AI is set to accelerate this growth significantly, meaning operators are rethinking their architectures and planning for how they can meet this demand sustainably.” 

But what is it about generative AI that puts such strain on the network? The answer lies in the technology’s core architecture, where training AI models, inference, and feedback loops generate constant movement of data between compute clusters. This creates far more intense networking demands than traditional cloud workloads, which tend to be more transactional or batch-based.  

“It is simply the nature of AI,” he explained. “How it trains, how it interacts with other machines, how it is so hungry for data, and how it generates new data […] That makes a big difference on the expected impact on the network.”  

AI traffic also differs significantly from traditional workloads. Hatheier described AI traffic as “unpredictable and exciting,” requiring networks to be flexible and dynamic, often handling large bursts of data with low latency. For example, an AI system may be required to automatically change a traffic light from red to green, or analyse a video for real-time security purposes,  both of which will require a network that can guarantee a high quality of service.  

“You do not know when you will require that data,” he emphasised. “Networks need to be built with these type of use cases in mind.” 

Another key feature emerging from the rise of AI is the growing shift toward distributed computing. As requirements for AI compute continue to grow, the training of large language models (LLMs) is becoming increasingly distributed across different data centres around the world, instead of being conducted in a single location. 

As AI drives up compute demand, sustainability has become a core concern for data centre operators. AI workloads are highly energy-intensive, and with growing scrutiny on emissions and energy use, operators are under pressure to expand capacity without expanding their environmental footprint.

Data centres consume significant amounts of energy, to both power servers and to keep them cool. As AI workloads increase, so does the demand for compute, and with it, the strain on power and space. This makes sustainability more important than ever, with operators looking for smarter, more energy-efficient ways to scale their infrastructure without compromising performance. But, as the bandwidth demands of AI intensify, so do the physical limitations within data centres, namely power consumption and space efficiency.