AI infrastructure is entering a phase most datacentres were never built for. The question operators now face is whether they can scale without their costs compounding alongside it. The scale of the challenge is already visible in the numbers. AI-driven datacentre electricity consumption has climbed at roughly 12% a year since 2017, four times the pace of global electricity demand growth overall. The UAE is no exception, with data centre electricity consumption projected to double from around 3 TWh in 2025 to more than 6 TWh by 2030. For operators, this goes beyond a distant sustainability concern, it’s a near-term capacity planning problem.
Intelligence Per Watt: The New Currency of AI Infrastructure
The environmental case for efficient AI infrastructure is well established, but framing this purely as a sustainability story misses the point. This is about infrastructure efficiency at system scale, and specifically about maximising intelligence per watt (IPW) across the full lifecycle.
AI is entering an agentic phase of inference, where models don’t just respond. They interpret, decide and act continuously in real time. That shift is fundamentally changing what infrastructure needs to deliver, driving up compute demand and putting sustained pressure on energy, latency and system efficiency across the entire stack. NVIDIA’s model of the new AI infrastructure identifies five layers, with energy at the base: a recognition that a system can only generate as much intelligence as the power available to run it.
That makes energy foundational to IPW. When IPW is higher, AI models deliver the same or better performance while drawing less electricity. This reframes the conversation: AI stops being an energy liability and becomes a driver of efficiency at scale, provided the infrastructure underneath it is designed with that outcome in mind. The applications are tangible. Higher IPW AI is better equipped to manage smart grids, reduce industrial waste and optimise resource-intensive systems.
The implications extend beyond operations. In the inference era, infrastructure efficiency shapes capital allocation, how quickly workloads can be deployed, and whether a system can scale without compounding its costs.
The role of edge in AI efficiency
Research indicates that running smaller, specialised AI models locally at the edge can cut energy consumption by 60 to 80% compared to large, general-purpose models operating out of central cloud datacentres. This decentralisation produces AI applications that are leaner, faster and higher in IPW — a strong argument for building datacentres around efficient model architectures and purpose-fit hardware, rather than simply scaling up existing infrastructure.
But the efficiency question is not reducible to a choice between centralisation and edge. Energy is only part of the picture. True infrastructure efficiency also encompasses how materials are sourced, how capacity is planned and how lifecycle decisions are made over time. A genuinely sustainable datacentre is one that compounds operational gains — each improvement in efficiency feeding into lower energy use and, in turn, higher IPW.
Translating the IPW imperative into infrastructure design
Moving into the inference era of AI highlights a fundamental challenge in the design of datacentres: Air-cooled datacentres were designed for an era of batch compute processing, not agentic AI. The more utilization and rack density increases, so do inefficiencies in the form of increased energy consumption, water usage, and accelerated hardware lifecycles creating additional costs and carbon emissions.
Solving this problem requires a holistic approach to the infrastructure stack rather than a series of incremental improvements in an architecture designed for a different use case.
One such solution currently gaining popularity is the implementation of liquid-cooling technologies and modular architecture. By adopting liquid-cooling in the architecture of a data centre, the thermal cap can be overcome, resulting in high compute density at a reduced energy expense. Additionally, by incorporating a modular approach, infrastructure need not go through complete replacements due to hardware updates and thus eliminates unnecessary expenses.
These tangible results can be quantified by looking at the following case study; Submer’s existing infrastructure assets have seen energy savings amounting to 913.68 GWh, water savings of 3,653.95 million litres, and CO₂ equivalent emissions savings totalling 323,110 tones. These numbers are derived from life cycle impact rather than just point efficiency, making them even more relevant when one considers the long-lasting impacts of infrastructure decisions taken now.
The implication for operators planning AI infrastructure is significant: efficiency is not a feature to be added; it is an architectural condition to be established at the outset. As AI workloads become as operationally critical as power or connectivity, the infrastructure supporting them will need to meet the same standard — delivering more intelligence per watt, consistently and at scale.
