For most of the last decade, the AI conversation was a software conversation. Better architectures, bigger datasets, cleverer training techniques. The hard questions lived in research labs, and infrastructure was something you rented by the hour from a hyperscaler.
That era is over. Today, the binding constraint on AI is not the model — it is the megawatt. Anyone who has tried to secure large-scale GPU capacity, a grid connection for a high-density facility, or a firm power purchase agreement in a major metro over the past two years knows exactly what I mean.
The queue starts at the substation
Over a trillion US dollars is being committed to data centre infrastructure through 2030. But capital is the easy part. The physical inputs — land with fibre and power adjacency, grid connections, transformers, switchgear, cooling plant — move on infrastructure timelines, not software timelines. A substation can take years from application to energisation. High-voltage transformers carry multi-year lead times. No amount of venture capital compresses the physics of a 132kV connection.
This is why the centre of gravity in AI has moved from the model layer to the power layer. The organisations that will deploy AI at scale in 2027 and 2028 are the ones securing their power and facility positions now.
“Compute is bought in minutes. Power is secured in years. That mismatch is the defining tension of the AI buildout.”
What it means for enterprises and governments
If you are an enterprise or public-sector organisation planning serious AI workloads, three practical implications follow.
- Treat power as a strategic input, not a utility bill. Understand your realistic load trajectory — training versus inferencing, peak versus sustained — before you commit to a deployment model.
- Decide build-versus-buy with open eyes. Public cloud, GPU-as-a-service, colocation and on-premise private compute all have different power, latency, data-residency and cost profiles. The right answer is usually a portfolio, not a religion.
- Sequence the long-lead items first. Grid applications, land options and cooling architecture decisions belong at the front of the programme, not the end.
The regional dimension
This shift favours regions that can move on energy. The UAE — with abundant generation, the Middle East's only nuclear fleet and sovereign programmes like the 1 GW Stargate UAE cluster — has understood that AI leadership is, at its core, energy leadership. India, adding over 1,500 MW of new data centre capacity through 2027, is pairing engineering depth with an increasingly sophisticated power procurement market.
The winners of the next phase of AI will not necessarily be the organisations with the best models. They will be the ones that treated megawatts as seriously as they treated mathematics.
About the author
Saandeep V. Dandekar
Executive Director — Infrastructure Strategy & Delivery, MetaDecrypt AI Atlas