By: Ivo Ivanov

The meteoric impact of artificial intelligence over the past few years is difficult to overstate, and progress is moving so quickly that it is better measured in months rather than years. Some of the biggest tech events of the year, such as the Consumer Electronics Show (CES) and Mobile World Congress (MWC), were dominated by emerging AI use cases, from LLM-powered humanoid robots to “sight beyond sight” vehicle-to-cloud software capable of giving cars human-like senses. Businesses also have high expectations for the next generation of AI, testing and deploying everything from problem-solving AI agents to advanced data analytics and forecasting tools. We’ve been conditioned over the past decade or more to believe that all we need to apply AI and use it effectively is data. More data typically means broader applications and better results. In 2025, however, that singular approach is being brought into question.   

For years, AI innovation was synonymous with the cloud. Training clusters and centralized hyperscale platforms, the brains behind AI, draw power and data into vast facilities designed to meet the demands of machine learning models. But this is just about the training of AI. With real-time demands and expectations of inference now resting heavily on its shoulders, AI is breaking free from these traditional gravitational centers. In the interests of speed and instant access, AI is now in action almost everywhere – embedded in devices, vehicles, retail environments, and digital agents that respond within milliseconds. So perhaps the question we should be asking isn’t just how much compute power can be deployed or how much data can be gathered – but how intelligently it can all be connected.   

This direction of travel has brought one vital characteristic to the fore: latency. No matter how massive the model or how sophisticated the silicon, high latency is kryptonite to AI. It’s the obstacle in the corridor, the roadworks on the highway – slowing everything to a crawl. It interrupts the flow of information between people, devices, and AI systems, introducing delays that can degrade user experiences, limit real-time decision-making, and even compromise safety in critical applications such as autonomous vehicles or remote health monitoring, where split-second responses can make all the difference. Make no mistake, in this post-AI world, latency is the new currency, and the next frontier of AI will be won or lost on the network layer.

The hidden enabler of AI innovation

Latency has always been a technical consideration in network design, but post-AI, it has become a business-critical variable. Every additional millisecond can distort outcomes in systems that rely on real-time processing, learning, and adaptation. AI agents designed to predict, recommend, or act autonomously are only as good as the data pipelines that feed them. If information arrives too late, decisions are made on stale insights, rendering even the most powerful models ineffective. Consider predictive maintenance in industrial manufacturing. Here, AI agents continuously monitor sensor data from machinery to flag anomalies before they escalate into failures. But if that data is delayed by even a fraction of a second, the insight may arrive too late to prevent damage or downtime. The same logic applies to AI in fraud detection, where instantaneous analysis of transaction patterns can mean the difference between blocking fraud and letting it through. In both cases, latency isn’t just a technical hurdle – it directly affects business continuity and customer trust. 

The rise of AI hubs and engineering near-zero latency

Traditionally, Internet Exchanges (IXs) were designed to facilitate efficient data exchange between networks, improving performance and reducing transit costs. But as AI workloads migrate to the edge, the role of IXs is evolving. Rather than serving solely as aggregation points for global, largely content-based Internet traffic, IXs are