THE DATA CENTRE INTERVIEW
Demand almost always ramps more slowly than anticipated before accelerating well beyond original projections – a pattern with direct consequences for networking, infrastructure sizing and cost management. The cloud-to-on-premises journey is also becoming a familiar trajectory. Pilots frequently begin in public cloud environments, where speed and accessibility make early experimentation straightforward. As usage scales and value proves itself, however, the cost profile changes.“ Costs often spiral up faster than expected,” Tim notes.“ At that point, many organisations look to shift more of their AI estate on-premises to regain control over performance, cost and data location.”
Use case clarity and business focus drive smarter AI deployments Three structural factors help explain why organisations continue to misjudge their AI infrastructure requirements. The first is a lack of use case clarity. Without precise objectives, engineering teams default to over-building.
“ A red flag for me is when an organisation is trying to tackle too many projects or use cases at once,” Tim says. His rule of thumb is to begin with fewer than five, prioritising those that are both feasible and high in business value. Early wins generate confidence and provide the financial justification for subsequent investment.
The second factor is starting with technology rather than the business problem. The convergence of data, compute, storage and networking in AI workloads creates genuine complexity – and that complexity can push procurement teams towards acquiring broad capability on a speculative basis. The more disciplined approach is to deploy precisely what current workloads require, while retaining the architectural agility to scale as AI maturity develops.
The third factor is legacy infrastructure. When existing environments are not designed for AI workloads, the instinct is often to add hardware rather than revisit architecture.“ The biggest issue I see is data silos,” Tim says.
“ The work required to build a usable knowledge graph across multiple data sources is almost always massively underestimated.” Bolting AI capability onto unsuitable foundations creates bottlenecks; the more effective path is adopting modern, validated architectures that are purpose-built for AI demands.
24 June 2026