The next phase of AI growth starts with infrastructure

A data center can go up in a year. A new grid connection, in some markets, takes five to seven. That gap is the real constraint on AI growth right now, not any single material or technology.

I heard that same idea echoed in two very different rooms this year: at the World Materials Forum in Paris, and on stage at the AI Infra Summit with Leo Aspauza, who leads cloud supply chain and sustainability at Microsoft.

Understanding the real capacity constraints

In Paris, I sat with mining executives, energy and grid-infrastructure leaders, materials scientists, policymakers and researchers from a dozen-plus countries. The sessions ranged from mining and refining efficiency to recycling, data-center design and hardware architecture — proof this isn’t a narrow problem with a simple fix.

At the AI Infra Summit, Leo put it well when he said organizations can grow capacity by both building more and getting more out of the infrastructure already in place. What struck me is that the industry needs both approaches. New infrastructure remains essential, but so does improving the efficiency of the infrastructure already in operation.

I’ve been saying some version of this to our own teams for a while now but hearing it independently from someone heavily involved with infrastructure at Microsoft's scale told me it’s now a shared reality across the AI industry. The discussion is about how growth and efficiency work together, not just one or the other.

Take global data-center electricity demand, for example. It could more than double by 2030, according to the International Energy Agency, from around 415 terawatt-hours to nearly 945 terawatt-hours. And every gigawatt of new AI capacity puts added demand on transformers, cooling systems, high-power batteries and materials like copper, rare earths and specialty steel.

That pressure is one reason why the industry is becoming more focused on preserving and reusing infrastructure value wherever possible. In AI systems, it’s often more efficient to retain valuable data and context than to repeatedly recreate it from scratch.

The materials challenge is really about refining capacity, supply-chain control and whether we can recover and reuse what's already been mined and built, not just scarcity in the ground. In storage, we’ve been working on this issue for years under a different name: circularity.

Component scarcity is a big part of why longer equipment lifecycles and tiered storage architectures have become important complements to ongoing infrastructure expansion.

Storage is no longer a sideline conversation

These same discussions are also changing how people think about storage. Like infrastructure growth more broadly, storage depends on making better use of power, materials and deployed assets.

For a long time, infrastructure conversations were mostly about chips and power, while storage was in the background. I’ve watched that change over the past couple of years, and my conversation with Leo confirmed just how far it’s shifted. We spent as much time discussing fleet circularity and lifecycle planning as we did on raw capacity.

Microsoft, for example, reported a 92% reuse and recycling rate for decommissioned servers and components moving through its Circular Centers in 2024 — proof this isn’t just a talking point for either of us. Leo and I spent a good part of our conversation on how the industry is extending drive life, recovering components and making fleets more predictable at scale. The fact that the AI Infra Summit exists at all tells you hardware makers, cloud operators and materials people now belong in the same room.

One number captures why I keep making this point: in a one-exabyte deployment, our Mozaic 4+ technology improves infrastructure efficiency by about 47% compared to a standard 30TB configuration. In practical terms, that allows organizations to store more data and support more AI workloads within the infrastructure they already operate. Efficiencies like these help unlock additional capacity from existing investments rather than requiring every increase in demand to be matched by new power, facilities and materials.

Storage decisions are now GPU decisions

Outside of engineering circles, you don’t often hear much about how every time a system has to recompute context it already generated, it’s burning GPU cycles that could be generating something new.

I continue to emphasize this point because it’s almost always cheaper to store and retrieve valuable context than to recreate it. As context windows, KV cache and vector environments keep growing, where the data lives — memory, SSD, or hard drive — has a direct line to how efficiently you are using your most expensive compute.

This is what I would ask infrastructure leaders to consider, because we talk a lot about GPUs and less about how smarter data placement is one of the most effective and immediate ways to get more out of the GPUs already in use.

These conversations highlight the reality of AI infrastructure today. Scaling AI now depends on using energy, materials and storage more thoughtfully. That’s a systems challenge, not a single-technology one and it’s one we can solve together, not alone.

Learn more: See how Seagate’s Mozaic platform is redefining the economics of AI-scale storage.