Digital Content Editor, Eve Goode speaks exclusively with Dev Tyagi, Co-CEO of Radian Arc building effective infrastructure.
As AI continues to accelerate rapidly, what are the most common gaps that organisations are leaving when it comes to infrastructure support?
The most pervasive gap is building AI infrastructure on top of what already exists rather than designing for what AI actually needs.
Most data centres reflect the physics and economics of yesterday’s workloads.
They are built for web, storage and training-style tasks, prioritising throughput over the real-time latency that inference demands.
Networks are oversubscribed and GPU resources concentrated in a handful of large regions, creating congestion and distance from users.
A second gap is unmanaged fragmentation.
Even purpose-built AI infrastructure is often assembled across multiple stakeholders, with no single entity owning or controlling the full stack.
This creates operational blind spots, resilience risks and sovereignty vulnerabilities.
A third is thermal planning.
Cooling has historically been treated as a facilities afterthought, but as power per rack can exceed 100kW, it becomes a core system constraint that directly limits where and how compute can be deployed.
Organisations that haven’t integrated thermal architecture into their AI infrastructure strategy from the outset find themselves constrained both physically and commercially as workloads scale.
Many organisations rely on fragmented infrastructure across multiple providers, why is this approach becoming more difficult for AI workloads?
AI workloads, particularly inference, demand continuously available, low-latency, high-density compute, but this requirement pulls in opposite directions across a fragmented estate.
When your infrastructure is split across multiple providers, each controlling a different piece of the puzzle, it becomes very hard to meet all three of those needs at the same time.
When multiple stakeholders own different layers of the platform, optimising the full system becomes structurally difficult: compute orchestration, workload placement, thermal management and network efficiency all need to work in concert.
There is also a sovereignty dimension.
Relying on infrastructure spread across multiple foreign-owned providers exposes organisations to conflicting legal regimes.
The US Cloud Act, for instance, can compel US-based companies to surrender data regardless of where it resides.
Fragmented multi-provider architectures make it harder to guarantee where data is processed, stored or at risk, which is an increasingly critical concern as AI becomes embedded in sensitive government and enterprise operations.
As AI continues to evolve rapidly and data centre projects can take years, how will the industry bridge this gap?
The traditional 24 to 36-month datacentre build cycle is incompatible with the pace of AI deployment.
Modular, reference-design-based builds helps bridge this gap by compressing this to nine to twelve months for core data centres, while also enabling scalability from core to edge without starting from scratch each time.
A second way to bridge the gap is to leverage existing infrastructure, particularly telco Points of Presence (POPs).
Telcos already hold thousands of geographically dispersed real estate locations, with power and fibre connectivity already installed, that with the right GPU and cooling infrastructure added, can deliver low-latency AI inference close to where customers need it, without new builds.
This turns dormant edge capacity into a near-term asset.
The third approach is to work with end-to-end partners capable of orchestrating the entire supply chain – land, power, cooling, compute and cloud services – rather than assembling it piecemeal.
This reduces coordination delays that compound the timeline problem when multiple vendors are involved.
4. As AI models become more powerful, how are issues such as power availability, cooling and data sovereignty changing the way infrastructure is designed?
Power availability, cooling and data sovereignty are increasingly design inputs rather than downstream considerations.
On power and cooling, AI datacentre electricity consumption has grown roughly 12% per year since 2017 – more than four times the rate of overall global demand.
As a result, cooling is no longer a facilities decision; it is a core design constraint that shapes system efficiency and resilience.
Thermal architecture directly determines rack density, deployment location and inference scalability.
Liquid cooling, including immersion, is gaining traction not as a preference but because air cooling cannot keep pace with the thermal density of modern AI hardware.
On data sovereignty, conflicting international regulations – with the EU’s GDPR and the US Cloud Act as clear examples, and the EU Cloud and AI Development Act (CADA) as a recent development – mean foreign-owned infrastructure can create legal and operational risks.
As AI resilience becomes increasingly tied to national security, infrastructure is being designed and localised within national or regional borders, with joint ventures and sovereign deployments becoming a structural requirement rather than a differentiator.
5. What does “full stack” or “end-to-end AI infrastructure” look like in practice and why is it becoming the future of AI-ready data centres?
“Full stack” or “end-to-end AI infrastructure” means owning and integrating every layer needed to run AI workloads, from strategy and site planning through to deployment, operations and scaling.
In practice, it includes consultancy, land and power acquisition, core data centre design, mechanical and electrical systems, high-density server hardware, modular data centre builds, edge compute nodes, high-speed connectivity, cloud services and ongoing management of compute workloads via effective orchestration and an AI control plane.
It covers both centralised AI training environments, such as large language model infrastructure and distributed edge environments for low-latency inference close to where data is generated.
This approach is becoming the future of AI-ready data centres because AI demand is growing too quickly for traditional fragmented models and 24 to 36-month build cycles.
Enterprises and governments need faster deployment, predictable scalability and infrastructure designed specifically for dense, power-hungry AI workloads.
Modular builds and reference designs can compress timelines while making expansion easier.
Just as importantly, full-stack infrastructure supports resilience, data sovereignty and operational control.
Organisations can reduce reliance on external or foreign providers, keep sensitive data within national or regulatory boundaries and avoid conflicts between privacy laws such as GDPR and overseas legislation.
As AI becomes business-critical and strategically important, controlling the whole stack is increasingly essential.
For governments and enterprises treating AI as critical infrastructure, the priority now is to move full-stack strategies: decide where you need sovereign control, where shared capacity is acceptable and which partners can deliver an integrated, AI-ready estate at the pace the market demands.