From grid-to-chip, power availability, supply chains, and regional variability are reshaping the long-term outlook for data center infrastructure.

Today, much of the industry’s attention is fixed on the immediate challenge of bringing more GPUs online and securing the capacity needed to meet surging AI demand.

But are we becoming too focused on solving today's GPU constraints? And what should operators be doing now to prepare for the realities of the physical grid-to-chip pipeline over the coming decade?

In a recent DCD>Broadcast, Schneider Electric took on these questions as part of its wider Management & Operations Innovation Day, exploring managing facilities at scale.

For Aparna Prabhakar, chief strategy and sustainability officer for energy management at Schneider Electric, these questions point to a much broader shift. Rather than focusing solely on the next 12 months, she argues that the industry needs to rethink what the infrastructure model should look like over the next five to 10 years – and recognize what has fundamentally changed in the last 18 to 24 months.

“The constraint is no longer about compute; it’s driven and constrained by energy – energy is the bottleneck to scale AI.”

In this way, as GPUs and CPUs become available faster than the grid connections and permits required to power them, the industry’s critical path has shifted from time-to-chip to time-to-power. That change demands a different approach to planning, where adding more capacity comes second to making better use of the capacity that already exists through smarter, more integrated operations, as Prabhakar highlights:

“The winners will be those who orchestrate power, cooling, and compute as one system, not independently.”

Centralized or decentralized? That’s not the question

As operators grapple with increasingly uneven access to power, it’s tempting to question whether AI infrastructure will become more decentralized, following available grid capacity, or if it will instead rely on on-site energy solutions such as small modular reactors (SMRs) and microgrids.

For Prabhakar, framing the future as a choice between centralized and decentralized infrastructure misses the point:

“I don’t think we are moving toward a simple binary model where everything is either centralized or decentralized. The way I see it, we are moving toward a more dynamic portfolio – a layered model across training, regional inference, and Edge.”

This is because AI workloads themselves are becoming increasingly diverse, and each demands a different infrastructure model. Large-scale training clusters, regional inference, and Edge deployments all have distinct requirements for performance, latency, resilience, and location.

What connects them is not where they sit, but how they are powered. Regardless of the deployment model, access to reliable energy remains the defining constraint, and the factor that will shape where AI infrastructure can scale next.

Software and orchestration become the AI control plane

Against this backdrop of more distributed AI infrastructure, software is emerging as the critical layer that enables operators to scale efficiently. Executives are looking beyond independent hardware choices, investing in orchestration and intelligent services that can optimize infrastructure across complex AI environments. That starts with breaking down the traditional divide between operational technology (OT) and IT.

“Historically, IT and OT were optimized separately. The compute team focused on utilization, workload performance, and availability, and the OT team focused on power, cooling, redundancy, and efficiency,” explains Prabhakar.

That separation is becoming difficult to sustain. AI workloads are dynamic, with demand shifting rapidly across clusters, racks, and facilities. The infrastructure supporting those workloads must be just as adaptive.

This is where orchestration becomes essential. Rather than managing compute and physical infrastructure independently, operators need a unified view that connects workload placement with the realities of power distribution, cooling, UPS systems, and thermal management.

“If workloads shift across racks or zones, your cooling strategy should be able to respond. If power availability changes, your workload placement should be informed by that. If there are thermal hotspots, the system should be able to detect and respond before it becomes an operational issue.”

In this model, the value of software becomes more pronounced – it becomes the intelligence layer that continuously aligns compute with the physical infrastructure supporting it. The result is data center infrastructure that is more connected, more responsive, and ultimately more efficient as AI deployments continue to scale.

Standardization as a strategy for an unpredictable market

The traditional data center delivery model – building projects one site at a time with highly customized designs – worked well when deployment cycles were measured in years. But the pace and scale of AI infrastructure have fundamentally changed that equation.

Operators cannot afford these lengthy, sequential design processes or bespoke engineering for every deployment. Instead, they are moving toward greater standardization, earlier collaboration across the value chain, and deeper partnerships between technology providers.

As such, Prabhakar emphasizes that accelerating deployment requires organizations to move together rather than in isolation:

“No one scales AI alone. If speed matters, the entire ecosystem must move together. Schneider, for example, is driving a lot of reference designs for future architectures with our ecosystem partners – that’s cutting the boundary between chip, server, power, and infrastructure providers.”

Today, reference architectures are becoming a pivotal tool in reducing infrastructure complexity, shortening deployment timelines, and building greater resilience into an increasingly volatile market. By aligning chip manufacturers, server vendors, power specialists, and infrastructure providers earlier in the design process, operators can respond more quickly to shifting technologies – all the while reducing the risk of costly redesigns as AI infrastructure continues to evolve.

Designing for change, not for certainty

One major challenge with standardized strategies lies in ensuring facilities can adapt to whatever comes next. We’ve seen chip architectures evolving at an unprecedented pace. Therefore, it is imperative to understand that infrastructure decisions made today will determine how easily operators can accommodate future workloads.

“In a market defined by change, optionality is the ultimate advantage,” Prabhakar says.

That philosophy extends across every layer of the physical infrastructure – from flexible power distribution and scalable cooling architectures to software management systems that can adapt as workloads evolve.

Rather than facing wholesale equipment replacement every five years, operators should be designing for incremental evolution. A modular approach reduces the risk of stranded capital by avoiding infrastructure that is too tightly coupled to a single processor generation, power architecture, or cooling strategy. Prabhakar contextualizes:

“For example, the market talks a lot about 800VDC. Three years ago there was no conversation around it, but we are moving pretty fluidly. Continuing five years ahead, the winning model is about modularity and flexibility to evolve.”

She explains that it is a mindset the software industry has embraced for decades. Software is built to evolve continuously through modular components and iterative upgrades, rather than periodic reinvention. Increasingly, the same principles are shaping physical infrastructure, with modularity and adaptability becoming foundational for data centers designed to support whatever the next generation of AI demands.

AI’s next frontier is energy intelligence

The key takeaway for Prabhakar is that the conversation around AI infrastructure can no longer be confined to the data center itself. As power becomes the defining constraint on AI growth, infrastructure strategy must increasingly become an energy strategy.

“The next phase will be defined by those who can connect all the dots – cooling, power, services – into one intelligent integrated ecosystem.”

That shift demands a more holistic view of infrastructure. Rather than optimizing compute, power, cooling, and software in isolation, operators will need to orchestrate them as a single system – one that can adapt to changing workloads, evolving chip architectures, and increasingly complex energy landscapes.

“The winners will be those who really think more end-to-end, grid-to-chip, and scale more intelligently around it,” Prabhakar concludes.

In the end, the next generation of AI infrastructure must make the most of the power available, respond intelligently to change, and evolve without constant reinvention. It’s the shift toward energy intelligence.

For more information, please visit Schneider Electric's AI Factory website.