Performance bottlenecks aren't uncommon when it comes to rolling out new technology, regardless of how capable or game-changing that technology might be. Every generation of new tech has encountered roadblocks that had to be overcome before it was truly able to shine. Virtualization forced organizations to rethink resource allocation, cloud transformation had us shift our focus toward scalability and elasticity, and microservices introduced entirely new challenges around observability and distributed systems.

There's something different about AI, however, in that it's exposing something far less obvious that not many in the industry are paying attention to. The models themselves are becoming astonishingly capable, but every interaction facilitated by AI depends on an invisible chain of infrastructure stretching across users, applications, APIs, cloud platforms, and increasingly, multiple geographic regions. The intelligence may live inside the model, but despite a rise in edge computing, the user experience still happens miles away – sometimes hundreds of miles away.

Perhaps that's why Cisco's recent AI Readiness Index Report found that only 13% of enterprises are fully prepared to capture AI's potential, despite widespread surface-level investment in AI technologies. The gap between ambition and execution is widening with AI, and as the novelty begins to fade and the technology becomes normalized, that gap is looking wider and wider.

A recent Censuswide survey commissioned by DE-CIX, which covered more than 400 IT and infrastructure decision makers across the US and the UK, exposes that gap further. While 96% of respondents believed their networks were ready to support future cloud and AI initiatives, those same organizations reported spending an average of 11.5 hours every week resolving cloud connectivity issues. That's more than a full working day lost not to building AI, but to keeping the connectivity infrastructure beneath it functioning.

AI Applications Don't Live in One Place Anymore

The architecture of enterprise applications has changed almost beyond recognition in recent years. Not long ago, an application request typically remained within a relatively self-contained environment. Today, a single AI prompt might authenticate a user, retrieve context from a vector database, query multiple APIs, invoke a large language model hosted in another region, validate permissions, and return a response – all within a matter of seconds. Every additional "hop" creates another opportunity for delay, congestion, or network failure, to send productivity spiraling.

That's why application performance is now inseparable from network performance. We often talk about AI models becoming faster or more accurate, but model capability doesn't matter to users. They experience the cumulative effect of every network hop required to deliver a response, and even the most advanced model cannot compensate for the inefficient trucking of data back and forth over public Internet routes.

According to the survey, latency between cloud environments (28%) emerged as the most significant cloud connectivity challenge organizations expect to face over the next year, ahead of security vulnerabilities including DDoS attacks (27%), and downtime or reliability concerns (26%). What's interesting is that these aren't isolated problems. They're all symptoms of the same underlying issue – modern applications have become highly distributed systems whose performance depends on every link in the chain.

Observability Is Expanding Beyond the Application

When applications span multiple cloud providers, external AI services, edge environments, and distributed data stores, performance can no longer be understood purely by observing the application itself. Teams also need visibility into how requests move between environments, where latency accumulates, and whether traffic consistently follows the paths they expect. The network, in other words, has become a more active participant in determining application behavior.

This explains why network security and application security are becoming increasingly connected. The survey found that 23% of respondents identified data sovereignty as a major cloud connectivity concern, alongside broader worries around visibility into data flows. As AI deployments expand across jurisdictions and providers, organizations are no longer interested solely in whether data reaches its destination quickly, but whether it reaches the right destination, through the right path, with the right level of governance.

Reducing Friction Instead of Adding Capacity

When performance deteriorates, the go-to solution for most industries has been to add more – more compute, more storage, more bandwidth. But while these added resources can alleviate symptoms, they don't tackle the underlying cause.

Enterprises are beginning to take note of this. The survey reveals that more than 60% already use private interconnection, while a further 31% are actively considering it. Rather than relying exclusively on the public Internet to carry business-critical traffic, these enterprises are rethinking the architecture underpinning their distributed environments. Will this make applications faster? Yes. But it will also make them more predictable, resilient, and easier to operate, while maintaining full visibility and control over what data goes where.

This is prompting a change in how we measure application performance, and it's long overdue. Success is no longer defined solely by how quickly a server responds or how much compute sits behind an AI model. Increasingly, it depends on how consistently organizations can move information between every component participating in a transaction. Removing unnecessary friction from those journeys will ultimately deliver greater value than adding yet another layer of complex infrastructure that risks piling on to the 11.5 hours IT teams already spend troubleshooting connectivity.

Performance Is the Real Measure of AI Maturity

AI has raised expectations for application performance in ways few technologies have before. Users increasingly expect intelligent systems to respond instantly, regardless of where data originates or which services contribute to the final result. Meeting those expectations will require far more than powerful models or larger budgets. It's time to rethink how we control our networks, ensuring that every interaction can travel through heavily distributed environments without unnecessary delay, disruption, or uncertainty. The performance bottleneck has been exposed – now we need to fix it.