Artificial intelligence workloads could consume 30% of global data centre capacity within two years, according to Goldman Sachs Research. The projection signals a capital rotation toward infrastructure as compute demand becomes the binding constraint in AI deployment.
Goldman Sachs said investors are entering a more selective phase, prioritizing companies that own and operate large-scale data centres over those offering narrow AI applications. Hyperscale cloud providers are already committing tens of billions of dollars annually to expand computing capacity, while networking systems scale to support higher throughput requirements.
Is Data Centre Capacity The New AI Bottleneck?
The shift reflects structural changes in how AI systems are built and deployed. Training large models requires thousands of chips running in parallel, while inference workloads demand persistent compute availability. Goldman Sachs estimates global data centre power demand could rise 175% by 2030 compared with 2023 levels, roughly equivalent to adding the electricity consumption of a top-10 country.
Still, infrastructure constraints are shaping strategic decisions across the sector. Data centre construction involves long development cycles, complex supply chains, and dependence on grid connectivity. Energy availability and cooling capacity are increasingly influencing site selection, with some operators moving to regions offering stable power and lower land costs.
Goldman Sachs describes the current phase as a “flight to quality,” where capital concentrates on firms with existing infrastructure and scalable revenue models. The firm’s analysis suggests these operators may capture more stable returns, echoing earlier computing cycles where foundational infrastructure providers outperformed application-layer companies over longer periods.
Yet, rising energy demand introduces new variables for policymakers and investors. Governments are assessing grid capacity, while environmental considerations such as water usage and emissions are gaining attention. The next phase of AI expansion will hinge on whether infrastructure buildout can keep pace with model development, with power availability emerging as the critical constraint to monitor.
