The conversation around AI data centers is changing quickly. According to Divyabhanu (Bhanu) Singh, Senior Vice President - Tech Industries, HCLTech, the questions customers are asking today are markedly different from those they were asking a year ago.

Earlier discussions centered on whether organizations could build AI infrastructure and secure enough GPU capacity. Now, the question has moved from “can we build it?” to “can we operate it at a cost structure that produces a return?” That is both an economics and an engineering problem.

“This AI data center issue is not a technical problem any longer; it’s a capital allocation problem,” said Singh. 

Regulation has also become a more important part of the infrastructure equation. In the US, AI governance requirements continue to evolve at federal and state levels, while the EU AI Act became generally applicable in August. For Singh, this reinforces a wider shift: where AI runs, where data resides and who controls the infrastructure are increasingly strategic business decisions.

From GPU capacity to business economics

The shift toward ROI changes how infrastructure decisions are made.

AI data centers require major investment and the conversation is moving beyond the technology team. Infrastructure leaders need to understand how capacity translates into business outcomes, how long new infrastructure will remain relevant and where different workloads can run most efficiently.

At the same time, the bottlenecks are becoming more complex.

“GPUs were a constraint last year. Today, land, power, cooling, the electrical infrastructure, the grid and the labor—all those elements have come together and have become a multi-layered constraint setup today,” he said. 

Power availability, grid capacity, land, supply chains, labor and regulation all shape where and how quickly AI infrastructure can be built, while conversion economics influence whether existing facilities can be repurposed. These constraints are tightly connected, so resolving one can quickly expose another.

That makes AI data center strategy an exercise in coordinating dependencies rather than solving for a single scarce resource.

Designing for inference

A second change is the movement from training-heavy environments toward greater inference demand.

Bhanu said this shift is forcing organizations to reconsider how their data centers are designed. Training and inference place different demands on infrastructure and architectures built for one phase of AI adoption may not be suited to the next.

The question is how to design for a moving target while technology, workloads and consumption patterns continue to evolve. 

Power, land, cooling, electrical infrastructure and labor all influence how quickly that capacity can come online. Singh also emphasized that infrastructure availability cannot be viewed separately from the surrounding physical environment, including access to power, water and communities willing to accommodate data center development. 

A hybrid future

For Singh, enterprises will need public cloud, private AI infrastructure and sovereign environments as different workloads create different requirements.

He compared the current discussion with the earlier cloud transition, when enterprises had to decide what belonged in their own data centers and what should move to public cloud.

AI will follow a similar path.

“It’s going to be a horses for courses policy that all enterprises are eventually going to take,” said Singh. 

Public cloud can be attractive when speed, innovation and access to complex capabilities matter most. Private AI infrastructure becomes important when enterprises have sensitive data, regulatory requirements or workloads that need to stay close to the people and systems using them.

Sovereign environments add another consideration when jurisdictional control and data residency shape where workloads can run.

“The answer isn’t cloud or private or sovereign. It’s all three. Cloud for flexibility. Private for production economics. Sovereign where jurisdictional control matters,” said Singh.

That makes the future inherently hybrid. The challenge for enterprises is managing those environments consistently, with common approaches to governance, cost control and workload portability.

“The hard part isn’t picking one. It’s running all three as one,” he added. 

From cost per token to outcomes per watt

The economics of AI also require new measures of success.

Singh argued that organizations need to think beyond raw infrastructure consumption and focus on the business value created by the energy and compute they use.

“The currency or the metric of success is no longer pure. It’s now moved to token per watt, or the watts per token, and really, it’s moved even further to the outcomes per watt,” he said. 

That reframes infrastructure efficiency around what the workload actually achieves.

That makes orchestration increasingly important. Enterprises need the ability to determine at runtime which environment is best suited to a workload based on cost, latency, data sensitivity and regulatory requirements, while maintaining consistent governance and cost control across the estate.

For infrastructure leaders, the challenge is no longer simply to build capacity. It is to create flexible architectures that can adapt as workloads, business priorities, deployment models and technologies evolve.

The inference era makes that adaptability essential. AI data centers are a moving target, and designing for change from the outset will give enterprises greater freedom to respond as demand and economics continue to shift.