Data centre demand shifts; the binding constraint does not

Principal Asset Management's case for data centres, published in September on IPE Real Assets' supplier hub, opens with a number built to end the argument before it starts: global digital data created, consumed and stored is up 19,600% since 2010. From there the asset manager leans on industry forecasts of roughly 14% annual growth for the sector through 2030, with global capacity approaching 200GW, and makes the case that data centres remain the standout structural growth story in real estate. The demand arithmetic is the easy half; the harder question sits in a distinction buried in the piece's second section, where Principal separates model training from AI inference — and in doing so moves the trade.

Training, as Principal describes it, is campus-scale work: large, high-density facilities placed in markets where land and power are more readily available. Inference, which the firm expects to grow behind it, shares locational requirements with cloud computing and in Principal's telling reinforces demand in primary and strong secondary markets. Read those two sentences together and the geography of the sector's next growth layer inverts the geography of its first: the training buildout could chase cheap land and uncontested interconnection, while the inference layer lands where land is expensive, where the queue is long and where the neighbours have already learned to show up at planning meetings. That implication the piece does not dwell on, and it is the one that should govern underwriting.

Principal frames the payoff as a stable cloud base with AI as a second layer of growth on top, together producing a deeper and better diversified demand profile. On the demand side that is fair, but on the delivery side it is close to meaningless: cloud and inference draw on the same grids, wait in the same interconnection queues, and face the same local consent fights. A portfolio long both carries one constraint. As this publication has argued, grid permission is the underwriting asset in digital infrastructure, and queue positions and interconnection contracts price before electrons do; Principal's two-layer framing diversifies the tenant book without touching that.

The sovereignty thread in the piece deserves more weight than it gets. Principal notes that data sovereignty requirements are driving additional demand across European markets, alongside continued hyperscale expansion of availability zones, and places the two observations side by side without joining them; they should be joined. Sovereignty-motivated demand must, by its nature, be served inside the jurisdiction that demands it, which removes the option of chasing cheap land and power elsewhere — likely the sharpest test of whether a 14% growth rate survives contact with permitting regimes and constrained grids.

The arithmetic itself is a demand forecast wearing the clothes of a supply forecast. This publication's tracking puts the delivery constraint in supplier slots and raw materials, with staffing the final bottleneck before energization and two-thirds of halls short of required staffing, so a compound growth rate with no delivery curve attached is not an underwriting input until someone prices the switchgear and the crews who install it. Fourteen percent annually compounds to a doubling in roughly five years; the queue that would have to clear to deliver it has not been priced at that pace anywhere in the visible portion of Principal's argument.

The macro prelude is the least promotional and most useful part of the piece. Principal reminds readers that investors entered 2026 braced for slower growth, cooling inflation and eventual Fed cuts, and instead met rising geopolitical tension, higher oil, rates backing up and a meaningful probability of further tightening priced in — against a US economy that has stayed resilient, with inflation generally improved and technology-driven productivity still supporting growth. It then asks whether markets have pulled forward too much of the AI future too quickly and into too narrow a set of stocks.

That question belongs one level down as well: if the equity market has concentrated the AI trade, there is a physical version of the same concentration — every operator's growth plan queues in the same places and competes for the same delivery inputs. The diversification Principal sells at the demand layer is partly undone by the shared constraint at the delivery layer, and the 200GW figure is where that tension shows up. It is a capacity number, not a delivery schedule.

Watch two things over the next two years. The first is rent pricing for inference-era capacity in primary and secondary markets, which is where the shift Principal describes becomes visible in numbers rather than forecasts, and where returns across operators will separate. The second is whether interconnection queue positions begin trading as assets in their own right, detached from the sites they were filed to serve. If that market develops, the sector will have conceded something Principal's five forces leave unspoken: the scarce asset was never the tenant.