The hidden costs of AI data transfer: why localised data layers matter

Posted on in

Most modern AI systems operate beyond static models. They require a continuous stream of real-time information, including video feeds, sensor readings, user queries, and relevant information generated via Retrieval-Augmented Generation (RAG). While a model file may only occupy a few gigabytes, the ingress data required for inference can rapidly scale into terabytes.

It is at this stage that operational expenditures can rapidly escalate.

Many systems are designed with data residing in a public cloud whilst the high-density computing is performed elsewhere. While this seems adaptable during the planning phase, it frequently leads to significant financial friction. Costs are attached to every byte of data moved, and as user bases expand, these expenditures grow vastly.

These challenges are often invisible during the initial deployment phase, only becoming apparent as scaling increases and monthly invoices reflect the true cost of data movement.

Data egress charges: why AI movement is expensive

Data egress refers to the fees incurred when transferring data from a cloud provider’s infrastructure to an external destination. While ingesting data is typically free of charge, providers apply significant levies when that data leaves their ecosystem.

For standard web applications, these fees may be negligible. AI systems, however, constantly relocate vast datasets to facilitate real-time inference and continuous model optimisation.

For most applications, this might not matter all that much. AI systems, however, are different. They’re always moving a lot of data, especially for things like using it in real time (real-time inference) or for them to continue improving (ongoing learning).

This is a regulatory concern. The UK’s Competition and Markets Authority (CMA) identified in its July 2025 Cloud Investigation that egress charges can act as a commercial barrier because these fees can actively stifle business growth and limit geographical flexibility.

To illustrate the scale of these costs, consider the following use case:

An AI-driven visual analytics system monitoring 100 cameras, typical for a large warehouse or retail environment. Each camera transmits 1080p video at approximately 4 Mbps, resulting in ~1.3 terabytes of data per month. For 100 cameras, the total is approximately 130 terabytes.

If this data is stored in a cloud environment but processed externally, the entire volume must be exported for analysis.

At current UK cloud egress rates (typically between £0.05 and £0.09 per gigabyte), this results in a monthly expenditure exceeding over £9,000. Annually, this totals more than £100,000 solely on the movement of data, excluding storage or compute costs.

This reality makes co-locating the data layer with the processing environment highly advantageous.

At Datum, our fixed-port, carrier-neutral approach eliminates this variable expenditure. By transitioning from a per-gigabyte billing model to predictable, fixed-cost data transfer, the egress tax is removed from the equation. For enterprises scaling AI systems, cost predictability is a strategic asset.

Data gravity and the intelligent hybrid

Data architecture is governed by the concept of data gravity, which dictates that as a data set grows in size, it becomes increasingly difficult to relocate. Eventually, applications and compute power are naturally drawn to where the data resides.

In AI, this happens quickly. Moving massive datasets between environments is computationally expensive, operationally complex, and introduces significant latency. Consequently, it is more efficient to run processing workloads adjacent to the data source.

This does not necessitate the abandonment of cloud services. Instead, the most effective solution is often a blend of environments.

This is the intelligent hybrid model.

Front-end services, user interfaces, and APIs can remain in the cloud to take advantage of on-demand scaling. However, the primary data layer, AI models, and high-intensity inference workloads should operate within a more self-contained, high-density data centre.

By consolidating data and processing within a single environment, organisations reduce the need for constant data relocation, resulting in a system that is more transparent, resilient, and scalable.

Performance costs: the physical limits of AI data

While fiscal costs are paramount, operational performance is equally critical.

Transferring data over significant distances incurs both financial costs and unavoidable latency. These physical constraints cannot always be mitigated by software optimisation: eventually, you encounter the fundamental limits of data transmission speeds.

As AI becomes more sophisticated, particularly with the rise of multimodal models utilising video and audio, the demand on supporting systems increases. Frequent long-distance data transmission is often incompatible with the requirements of these advanced models.

For mission-critical applications requiring immediate action, such as autonomous systems or industrial automation, even minimal latency represents a total system failure. A delay of even a few hundred milliseconds can lead to missed anomalies or incorrect automated responses.

When data must travel to a remote facility and back before a decision is reached, these delays are inevitable.

By situating compute power adjacent to the data layer, the processing occurs where the data is generated or stored. This ensures consistent, low-latency performance that meets the rigorous demands of modern AI. At this scale, minimising latency is essential for the fundamental viability of the system.

Sovereignty: the safe haven approach

There is an additional cost that does not appear on a standard invoice: operational risk.

When AI datasets are transferred across regional borders, they trigger complex legal and compliance requirements. GDPR and similar frameworks regulate not only where data is stored but how and where it is transferred.

Each transfer introduces new regulatory obligations, including data residency rules and audit trails. The administrative complexity of managing these legal and practical requirements scales over time.

Localising the data layer significantly simplifies this process.

A data layer based in the UK allows organisations to keep sensitive datasets within a regulatory framework they already navigate. However, geographical location is only one part of the equation: for data to remain truly local, it must be hosted within a secure environment.

This is where a safe haven strategy becomes vital. Datum’s facilities, built to BS 5979 standards and featuring government-grade security, provide the necessary infrastructure for sensitive workloads. Coupled with NSI Gold Approved standards, this provides a foundation where data is both geographically resident and physically protected.

This approach reduces administrative overhead, mitigates insurance risks, and streamlines regulatory compliance.

The key to scalable AI

As AI systems expand, the physical distance data must travel becomes a defining factor in both cost and performance.

Egress fees, processing delays, and compliance risks all indicate a singular conclusion: transferring vast datasets over long distances is neither commercially viable nor operationally sustainable.

A localised AI infrastructure provides the solution. By aligning data and processing, organisations can eliminate unnecessary expenditure, enhance system responsiveness, and simplify governance.

If you are currently scaling an AI initiative, it is vital to audit your current data movement patterns. Datum’s technical experts can evaluate your data workflows and demonstrate where a localised architecture will provide a tangible commercial advantage.

Back to Insights