AI Infrastructure, Energy and the Path Forward.

In the early stages of AI development, data availability and compute capacity determined how quickly new models could advance. That equation has changed. As Radium CTO Vijay Gadepally explains, “Ten years ago the constraint was data collection. Five years ago it was compute. Today it is power.”

This shift is visible across the world. In the United States, most new data centre permits are capped at 99 MW, just below the 100 MW regulatory threshold, with gigawatt-scale campuses constructed from clusters of these modular units. Notably, as of early 2025, not a single data centre in California exceeds the 100 MW mark, a clear indicator of how infrastructure strategy is now shaped by regulation as much as technology. Regions facing grid stress have paused new interconnections, while jurisdictions with an abundance of hydroelectric power are inviting large-scale infrastructure projects.

The implications extend far beyond siting. Power availability now shapes system architecture. Small modular reactors, once considered promising, remain years from commercial deployment. As a result, many operators rely on natural gas with solar supplementation, which introduces emissions and significant water requirements for cooling.

One near-term opportunity lies in adaptive intelligence, dynamically aligning model performance with real-time energy conditions. Organisations that merge energy management with compute optimisation will have an advantage as infrastructure demands escalate.

AI workloads fall into two categories: training, which builds models and inference, which applies them at scale. Historically, training consumed the majority of compute. That balance has now reversed.

As Gadepally observes, “Models have become so large that inference is now the dominant compute cost.” Trillions of daily queries across consumer and enterprise applications have created a distributed, always-on demand profile that requires infrastructure close to end users.

This shift has profound economic implications. Training remains centralised and resource-intensive, dominated by a few major players. Inference, however, is driving the need for global low-latency architecture and sustained power availability. Companies scaling for inference are establishing the foundations of a decentralised AI economy.

The once-prevailing belief that AI would be served exclusively from massive centralised data centres is becoming outdated. Regulatory, trust and latency considerations are pushing workloads toward distributed, localised deployments.

As Gadepally notes, large training hubs will remain essential, “but inference should be close to users, especially for medical or safety-critical workloads.”

Frameworks like GDPR reinforce this direction. The “right to be forgotten” is extremely difficult to implement in generative models that cannot isolate individual contributions. This is driving jurisdictions to mandate that certain classes of data remain within national boundaries.

One emerging solution is data embassies, where infrastructure is hosted abroad yet legally governed by the home nation. This model preserves sovereignty while leveraging global capacity and represents a pragmatic balance between control, resilience and scale.

While infrastructure and energy shape the boundaries of what AI systems can do, talent and proximity determine where innovation truly happens. Regions building meaningful AI capability are those that invest in people, create environments where ideas circulate and enable regular interaction between founders, engineers and domain experts.

Innovation ecosystems grow when there is clarity in regulation, access to technical expertise and opportunities for collaboration. These conditions allow experimentation to turn into execution and research to translate into commercial impact. For emerging hubs, success depends on deciding what requires local oversight, such as sensitive data or essential services, and what can be distributed to where it is most efficient.

Ultimately, ecosystems thrive when they attract builders, encourage cross-disciplinary exchange and support the progression from concept to company.

AI infrastructure is no longer conceptual. It is being built now, shaped by constraints of energy, sovereignty and economics. Organisations that understand AI as a systems-level challenge rather than a software problem will define the next wave of innovation.

As Radium CEO Adam Hendin notes, “Digital and AI-related services now outpace traditional financial services in revenue for major consulting firms. The shift is systemic. AI is becoming the foundation of all professional services.”

According to the International Energy Agency, global investment in data centres in 2025 is projected to reach $580 billion, surpassing the $540 billion spent on new oil supply, underscoring how infrastructure has become a cornerstone of the global economy. For leaders and jurisdictions alike, the opportunity is clear: shape how these systems evolve, cultivate the talent capable of advancing them and create an environment where innovation compounds.

The infrastructure era of AI has begun. The question is no longer when, but how it will be built and who will shape its design.

Vijay Gadepally is Co-Founder and CTO of Radium, a unified, secure AI cloud platform for building and deploying private AI across the full model lifecycle. He has more than a decade of experience in AI and high-performance computing, with a distinguished career at MIT. As a senior member of the technical staff at the MIT Lincoln Laboratory Supercomputing Center, he has collaborated closely with MIT’s CSAIL. His research in AI, supercomputing and sustainable computing has earned recognition across government, academia and industry.

Adam Hendin, Co-Founder and CEO of Radium, is a serial entrepreneur with a record of building and scaling global technology operations. He has established multinational supply chains, deployed digital infrastructure across Europe and built Canada’s most advanced wireless remanufacturing facility. Hendin has completed multiple successful exits and negotiated commercial agreements with Fortune 100 and 500 companies.