The world of data infrastructure is being rebuilt. Traditional data centres were designed for enterprise software, modest CPU loads, and inexpensive air-cooling. AI has overturned that logic. The new generation of GPU clusters demands tens of kilowatts per rack, liquid cooling, ultra-fast fabrics, and fully automated operations. Nations now see AI data centres not as a technical upgrade but as a strategic asset—central to productivity, innovation, and sovereignty.
The original idea of a data centre as a warehouse of servers no longer applies. The shift to AI workloads has rewritten the rules of power, cooling, security, and national capability. As the figures illustrate, this is not evolution. It is a structural break.
READ I Ukraine war: Why a frozen conflict now looks inevitable
AI data centres need infrastructure upgrade
Conventional data centres were built around CPU workloads consuming only a few kilowatts per rack. Air-cooled rooms were sufficient. Networking needs were modest. None of this fits today’s AI systems. GPU clusters now draw tens of kilowatts per rack. Liquid cooling is no longer optional. Fabrics have moved to 400/800 Gb. GPU-direct storage has become essential.
Figure 1 visualises this shift effectively. It shows how AI data centres differ in density, cooling, networking, and workload patterns. Only a handful of operators can handle racks above 100 kW. Most organisations must either retrofit old facilities at high cost or build new ones from scratch.
Google’s Power Usage Effectiveness (PUE) of 1.09 is a reminder of what the frontier looks like. Industry averages remain near 1.5. AI workloads push facilities to the edge of what traditional designs can support.
Structural constraints of existing facilities
Many emerging-market facilities lack the power, cooling, or operational maturity to support AI workloads. The limitations fall into four buckets.
Power is the first barrier. Grids are congested and slow to expand. High-density AI racks require dedicated feeders and stable supply. Thermal ceilings are the next constraint. Air systems cannot dissipate the heat that GPU racks generate.
Networking lag is another problem, since AI fabrics need 400/800 Gb latency paths that most legacy buildings cannot deliver. Operational maturity is the final gap. Zero-trust designs, DPU-based isolation, and automated workload management remain rare.
The visuals in the Word file make these gaps clear. They show why upgrading a traditional facility is not just a matter of adding more servers. It requires rebuilding the entire environment.
A clear taxonomy for AI data centres
To bring order to this complexity, the report proposes a three-tier taxonomy, shown in Figure 2. It divides AI data centres into:
- Edge and entry deployments, designed for local inference and low-latency prototyping.
- Regional training clusters, which balance compute and networking needs for research, enterprise, and mid-scale model training.
- National AI superhubs, multi-megawatt campuses built for sovereign AI capability and hyperscale workloads.
This taxonomy has practical value. It allows policymakers and corporations to plan stepwise investments. Edge clusters prevent stranded assets. Regional clusters create scalable training capacity. National superhubs provide the compute backbone required for defence, large-model training, and economic competitiveness.
The visuals in the Word file strengthen this structure. They show how workload types map cleanly to deployment tiers.
A framework for National AI Data Centre Strategy
Taxonomy answers what to build. The framework in your document answers how to operate and scale it. Figure 3 summarises the three pillars: AI-optimised operations, security and sovereignty, and workload readiness.
Each pillar addresses a specific constraint that traditional facilities fail to meet.
AI-optimised operations rely on AIOps for predictive maintenance and automated scaling. Zero-touch provisioning reduces manual errors. Observability stacks using OpenTelemetry and Prometheus ensure continuous monitoring.
Security and sovereignty require zero-trust architectures, DPU-enabled workload isolation, and firmware protection compliant with standards such as NIST SP 800-193. As AI becomes central to national security and critical infrastructure, this pillar grows more important.
Workload readiness demands next-generation GPU clusters, ultra-fast 400/800 Gb fabrics, and GPU-direct storage. Benchmarking tools such as MLPerf and iperf help tune performance and avoid bottlenecks that can derail large training workloads.
Each pillar aligns with each tier of the taxonomy. Edge deployments emphasise low latency. Regional clusters prioritise operations. National superhubs integrate all three pillars, with sovereignty concerns at the core. This architecture gives governments and corporations a roadmap for AI scale-up without wasteful experimentation.
Why integration matters
The power of your framework lies in its integration. Taxonomy shows where an organisation stands on the maturity curve. The operational framework shows what competencies each stage requires. When combined, they allow decision-makers to sequence investments and avoid stranded capital.
This approach is especially useful for emerging markets. Many of them leapfrog directly to AI use cases—language models, healthcare inference, and digital governance—but lack the infrastructure to support them. Your model enables a staged path, beginning with inference clusters and scaling to national superhubs without financial or operational shocks.
The policy imperatives
AI is no longer a niche workload. It is the foundation of competitiveness for governments and corporations. Nations that fail to build AI-ready data centres will face structural disadvantages in productivity, innovation, and security. The gap between countries that invest early and countries that fall behind will widen rapidly.
Traditional facilities cannot support the power loads, thermal management, or security standards that AI requires. Retrofitting is expensive and slow. Building fresh with the right taxonomy and framework is often the more rational choice.
For governments, the priority should be a national AI data centre plan that aligns with industrial policy, education, digital governance, and defence needs. For corporations, the priority should be avoiding stranded assets and planning capacity growth with clear visibility of operational maturity and sovereignty requirements. The message of your report is simple but critical: AI needs new infrastructure. Without it, competitiveness erodes.
AI has changed the economics and architecture of data centres. Power, cooling, networking, and security must all be redesigned. Your taxonomy and three-pillar framework offer a coherent roadmap for this transition. They allow organisations to scale responsibly, reduce risk, and align infrastructure with national goals.
AI is not another workload. It is the new backbone of competitiveness. Countries and corporations that recognise this will shape the AI economy. Those that cling to legacy designs will struggle to keep pace.
Yugesh Panta, Joydeep Hazra, Badri Narayanan Gopalakrishnan work with Infinite Sum Modeling LLC; Santosh Kumar Nukavarapu and Sai Kiran Poka are with Old Dominion University.