The growth logic behind AI infrastructure demand is shifting: the market's focus on the pace of frontier model iteration is gradually giving way to AI's continued spread into enterprise and real-world business scenarios. Even if model development slows, the broad adoption of AI applications could still generate sustained data center demand.

HSBC believes the key factor determining traditional data center demand is not the speed of model iteration, but the commercial diffusion of AI technology. As AI moves from training toward inference and actual deployment, demand will expand further into enterprise IT upgrades, cloud computing, and data interconnection scenarios, leaving traditional data centers with substantial capacity to absorb.

At the same time, supply-side constraints on power, grid connection, and regulation are intensifying. New data center capacity is difficult to bring online quickly, with many projects delayed or canceled, and a tight supply-demand balance is expected to persist into 2027 or even 2028.

Against this backdrop, industry earnings growth remains resilient. The report estimates that relevant data center REITs will see AFFO per share compound growth of roughly 11% to 12% between 2025 and 2028, above the historical average of the past five years; even if financing costs rise further, the drag on AFFO growth would be relatively limited.

Rise of Inference Demand Opens Traditional Compute Space as AI Spreads

As AI applications move from training into the inference stage, the shape of compute demand changes accordingly. Training is typically concentrated in a small number of large-scale projects, while inference must continuously respond to end-user and enterprise business requests, with more dispersed workloads and greater reliance on networking, storage, and general-purpose computing infrastructure.

Microsoft expects that by 2032, less than 50% of new data center capacity will be driven by AI chips. This means that even if AI-related demand continues to grow, not all new data center capacity will convert into high-density AI compute, and traditional computing workloads will still account for a considerable share.

Capital spending by cloud service providers shows a similar trend. HSBC's technology team expects the six major cloud service providers to still increase capital expenditure by 37% and 9% in 2027 and 2028, respectively. Demand from AI application expansion is not only reflected in GPU purchases, but also transmitted to traditional data centers through cloud computing, data processing, and enterprise infrastructure upgrades.

Therefore, a more rational approach to frontier model development does not mean data center demand is cooling in tandem. As AI enters more enterprises and business processes, infrastructure demand may instead expand from a small number of hyperscale training projects to broader, more sustained application scenarios.

Power and Regulatory Tightening Make It Hard for New Supply to Keep Up

Data center expansion is increasingly constrained by power conditions. Multiple U.S. regions have raised electricity, environmental, and water requirements for data center projects, and are requiring large power users to bear more of the cost of power plants, transmission lines, and grid upgrades. Power access has become an important factor affecting whether projects can proceed.

Tighter regulatory requirements further reduce supply elasticity. California has removed the environmental review exemption for data centers and requires operators to disclose water usage; Texas has required a pause on related permit issuance until ERCOT completes an audit. For capital-intensive projects, longer approval and grid connection cycles mean new capacity is released even more slowly.

Project delays and cancellations are already reflecting this change. According to Data Center Watch, in the second quarter of 2026, at least 45 data center projects worth about $68 billion were delayed or canceled; in the first quarter, at least 75 projects involving about $130 billion were affected.

The significance of limited supply is that even if demand growth fluctuates somewhat, new server rooms cannot be added quickly, and the scarcity of existing data center resources may instead be maintained. This supports rent growth and asset utilization and is an important reason the industry can sustain relatively strong profitability.

Limited Impact from Rising Rates, AFFO Growth Still Supported

Rising interest rates remain the main pressure facing data center REITs. Because industry expansion requires continuous capital investment and increased debt, higher financing costs directly compress AFFO. But the report's sensitivity analysis shows that the impact of rate changes on earnings is mainly at the margin of growth, rather than changing the industry's growth trend.

Assuming new debt rates are 50 basis points above the base case, 2027 to 2029 AFFO per share forecasts would fall by about 0.3% to 0.8%. Correspondingly, AFFO per share growth would decline by only about 10 to 40 basis points, leaving rate pressure relatively limited compared with demand and supply factors.

At the same time, HSBC expects relevant data center REITs to see AFFO per share compound growth of about 11% to 12% between 2025 and 2028, clearly above the historical average of the past five years. A higher growth base means that rising financing costs are not yet enough to reverse the earnings growth trend.

In terms of industry logic, new demand from AI application diffusion, supply constraints caused by power and regulation, and strong AFFO growth together form the three structural tailwinds for traditional data centers.