When AI is cleared for take-off, but it’s not reaching cruising altitude

Enterprises everywhere are deploying powerful AI environments on GPU-rich cloud infrastructure – yet many models still underperform in production. On paper, AI should be soaring but in reality, it rarely reaches cruising altitude. The question is: why?

For many organizations, the constraint is no longer just compute. Increasingly, it’s the cost, efficiency, and reliability of moving data to where AI models actually run. Network architects and infrastructure teams find themselves managing ever‑more‑complex environments, only to discover that AI performance is being limited by something far less visible but no less critical.

If we think of AI as an aircraft ready for departure, then take-off – and sustained flight – depends on one thing above all else: a continuous and precisely managed fuel supply. In AI architectures, that fuel is data.

Running on empty: why AI stalls without data flow

For AI to fly, the data must move quickly, securely and continuously. This is especially critical given that data in transit is one of the most vulnerable points in any AI pipeline. The challenge is that modern AI environments are rarely located where the data they depend on resides. Datasets are now distributed across on‑premises systems, edge locations, and multiple cloud platforms, while AI models are typically centralized in specialized compute environments optimized for GPUs rather than proximity.

It’s one reason why more than 80% of AI projects fail – in part due to inadequate data movement to the point of AI processing. And it must be moved securely especially as data in transit is highly susceptible to cyber-attacks.

To continue with our aviation metaphor, this is like expecting an aircraft to fly to a different airport every time it needs fuel – which isn’t impossible but is certainly impractical. The cheaper and more efficient option would be to find a system for bringing the fuel to the plane, and not vice-versa. The same principle applies to AI architectures, where the data must move to the AI-optimised location, not the other way around. However, it’s important to do this in the right way. Fragmented datasets, blind spots, and inconsistent pipelines can easily disrupt the data flow, limiting AI performance even when the underlying infrastructure appears ready for action.

Drifting off course: the disconnect between data and compute

There’s an assumption that data will simply fulfil its supporting function by flowing to where it’s needed. Unfortunately, the reality of AI is not that simple – mainly because “AI workloads, especially those involving large language models and real-time analytics, require terabytes to petabytes of data to be moved daily”.

Even under optimal network conditions, transferring data volumes at that scale can take days or even months. Every delay also means the data that does arrive is often stale – and for models dependent on real-time inputs like fraud detection or predictive maintenance, acting or training on outdated information erodes accuracy and trust.

The financial ramifications of poorly managed data movement environments are equally as compromising, with one government report revealing cloud egress fees for moving “data once from one cloud provider to another” can account for up to 10% of annual spend for some customers. Then there’s the “double bubble” effect, where enterprises are paying for data sitting idle in one environment while expensive GPU resources wait in another.

Taken together, these architectural challenges create an expensive yet barely visible level of operational turbulence that is growing increasingly difficult to stabilise at scale. For decision-makers in the operational cockpit, that means addressing the disconnect between data and compute has become a top priority.

Turbulence ahead: where traditional data movement falls short

AI has accelerated innovation at a pace that has exposed the limitations of traditional data transfer approaches. What was once a simple problem solved by incremental bandwidth upgrades and a few open source tools have been outpaced by the complexity of modern enterprises, which in turn demands more reliable and repeatable data movement than ever before.

Plus, modern dataflows often need to be organised, validated and secured before being moved – normally across multiple environments with different protocols, formats and compliance requirements. Each of these steps increases the likelihood of delays, decreases data freshness and raises concerns around business governance.

This is also why scaling infrastructure by adding more compute or more storage doesn’t resolve the issue. This time, it’s the equivalent of expanding an airline fleet without considering how often all the aircraft can be refuelled across a range of airports. It’s the same problem, multiplied. Having data available somewhere in the system isn’t enough anymore, because channelling it seamlessly and securely at scale is too resource intensive.

Until the data pipelines are fit for purpose, organisations will struggle to support their AI infrastructure in a sustainable way. And with that, the reality gap between AI’s promise and the value it actually delivers in production will only continue to widen.

Adjusting the flight path: making data movement a strategic priority

To bridge this gap, enterprises need to completely rethink how they approach data movement – designing, measuring and organising it in line with the broader AI strategy. Metrics such as data transfer speed, reliability, chain-of-custody and cost-efficiency should sit alongside model accuracy and compute utilisation as key indicators of AI performance.

As systems have to continuously facilitate high-volume data movement without adding unnecessary complexity or cost, an architectural shift is also required. In effect, the fuel supply chain needs to become as agile and durable as the aircraft itself.

This is where investment in be-spoke data movement solutions play a critical role. Rather than relying on gradual networking improvements and improvising with simpler tools that will never catch up with AI demand, organisations can deploy solutions designed to move petabyte datasets across public clouds, regions, and data centres. In some cases, this can compress transfer times from months into days.

Making the structural changes needed to modernise data pipelines in this way builds an AI data fabric that’s quicker, simpler, compliant with enterprise governance mandates and more insightful. With that, organisations can feel the benefit of a service infrastructure where reduced idle compute, lower operational overhead and full visibility over data movement replace the fragmented, costly pipelines that have previously held AI scalability back.

Cruising altitude: optimise AI performance with better data movement

Many enterprises have cleared AI for ascent. But their approach is being obstructed by a lack of interoperability between GPU clusters in a specialised location and enterprise datasets that, ironically, are up in cloud environments or remote data centres.

Without finding ways to consolidate this fragmented landscape into a consistent, scalable and well-governed flow of data, these ventures will drift off their intended flight path – delivering inconsistent and unpredictable performance despite significant investment.

Data is where the next phase of AI maturity will be defined, so the organisations with the most powerful models or largest GPU clusters will be rendered ineffective without efficient and scalable movement. If this prerequisite is not met, AI simply won’t be embeddable into core business operations.

However, the enterprises that do prioritise data movement within their AI strategies – and invest in petabyte data movement solutions and data fabrics in the process – will be able to keep pace with the speed of innovation. Not only will their AI initiatives be consistently fuelled, but they’ll also have the infrastructure to reach new heights.