Q&A: AI’s infrastructure crunch could actually make it more sustainable

AI demand is growing much faster than the physical infrastructure needed to support it. Companies can deploy new AI applications relatively quickly, but the power, water, chips, cooling equipment, transmission infrastructure and data center capacity behind them can take years to expand. That mismatch is increasingly becoming a constraint on AI growth.

Dan Magestro, Chief Data & AI Officer at Sphera, believes those constraints could ultimately be good for AI’s sustainable development, as he explains to Digital Journal. As Chief Data & AI Officer, Magestro leads Sphera’s data strategy and organization, supporting Sphera’s portfolio of data-driven sustainability and risk solutions. Magestro also leads Sphera’s support of the AI economy, connecting Sphera’s products and expertise to the unique needs of companies navigating the AI boom. He has over 20 years of industry and consulting leadership and holds a Ph.D. in nuclear science.

Magestro’s argument is when power, water and hardware are abundant, companies have less incentive to scrutinize how efficiently they’re being used. Scarcity changes the equation. It pushes companies to get more computing power from every watt, reconsider where workloads run, improve hardware utilization and understand the trade-offs between resources rather than simply building more infrastructure.

Digital Journal: Why do you believe the constraints slowing AI growth could actually make AI more sustainable?

Dan Magestro: It comes down to the interdependencies between the industries driving the AI boom. AI software can scale incredibly quickly, but the physical resources behind it cannot. For example, a company can deploy a new model or expand an AI application relatively quickly, while adding new power generation, transmission infrastructure or data centre capacity can take years. 

AI also depends on finite resources like water, chips, cooling equipment and the supply chains supporting that infrastructure. Such constraints also expose inefficiencies and force companies across the AI ecosystem to make real choices about how they use limited resources

That  friction isn’t necessarily a bad thing. When resources become constrained, companies have a much stronger incentive to understand how efficiently they’re using them, which leads to more sustainable practices

Electricity is a good example. If power is abundant and inexpensive, there is less pressure to scrutinize every watt. When capacity is limited, efficiency suddenly becomes much more important, and so does visibility into new sources of demand like data centers, which leads to better policy and more resilient energy planning. That can drive improvements ranging from more efficient chips to smaller AI models and more strategic decisions about where and when compute-intensive workloads run.

In short, friction forces efficiency. The constraints we’re encountering today could ultimately force the AI ecosystem to become more disciplined about how it uses resources.

DJ: What are companies getting wrong about the sustainability challenge surrounding AI?

Magestro: One mistake is looking at each AI segment or constraint independently. In my role at Sphera, I see every part of the AI ecosystem – data center operators, their suppliers, energy and water utilities, AI users, and the regulators and communities around them – often solving for its own piece.

In my view, power, water, hardware, data center capacity and supply chains are all interconnected. If you solve an electricity constraint without considering water use, for example, you may simply shift environmental pressure somewhere else. Similarly, rapidly expanding data center capacity without understanding the hardware and supply chains behind it can create a new bottleneck downstream.

DJ: Does making AI more sustainable require companies to sacrifice growth or performance?

Magestro: Not necessarily – I’d argue the opposite. A lot of the actions that make AI infrastructure more sustainable are also simply good business decisions.

A supplier that better understands its supply chain risk becomes more resilient. An operator tracking emissions gains better visibility into its operations. A company measuring the resources consumed by its AI systems can make more informed decisions about where those systems deliver enough value to justify their resource requirements. Each of these makes sense in pure self-interest, while benefitting others along the way.

And because these businesses are interconnected, those improvements can extend beyond an individual company. Better supplier data can improve an operator’s decisions, and an operator’s tracked emissions can become a utility’s clearer demand signal. Companies acting in their own interest end up helping the peers and partners around them. 

DJ: What should companies be measuring as AI use grows?

Magestro: We’re going to need metrics that connect AI output to the resources required to produce it.

One example is “tokens per watt,” which NVIDIA has promoted as a way of measuring useful AI output relative to electricity consumption. Measures like that can create incentives to improve efficiency throughout the AI stack, whether that’s designing chips that provide more computing power per watt, developing smaller models or becoming more deliberate about when particular models need to be deployed.

The broader principle is important: Companies shouldn’t measure AI growth solely by how much compute they’re adding. They should also understand how efficiently they’re turning resources into useful outcomes.

DJ: What can business leaders do now to prepare for continued AI infrastructure constraints?

Magestro: First, understand where your business is most vulnerable. Identify which resources and partners your AI growth depends on, whether that’s energy, water, chips, suppliers or data center capacity, and determine what happens if those resources become more difficult or expensive to access.

Second, build a more complete picture of your impact. Required disclosures won’t necessarily tell you where your biggest risks or tradeoffs are. Tracking energy, water, emissions and other operational factors together can help leaders understand the tradeoffs they’re making rather than optimizing one resource at the expense of another.

Finally, don’t wait for regulation or a shortage to force action. Building better supply chain visibility, environmental accounting and operational risk management capabilities now can improve infrastructure decisions today while preparing companies for future requirements.

DJ: Do you expect the gap between AI demand and infrastructure capacity to go away anytime soon?

Magestro: No. AI infrastructure will continue expanding, but physical infrastructure simply moves on a different timeline than software. Power plants, transmission lines, data centers and supply chains cannot scale overnight.

That means constraints are likely to remain part of the AI landscape for some time. The opportunity is to use that friction to get more value from the resources we already have. If companies understand how energy, water, hardware and supply chains interact, they can improve efficiency without simply shifting costs or environmental impacts elsewhere.

AI’s limitations are real, but used well, they can help us build the next phase of AI more sustainably.