To accurately gauge the scale of this cycle, an AI datacenter must be viewed as an integrated electrical and thermal system built around extremely dense servers. AI accelerators, particularly GPUs and ASICs, generate significant heat, consume vast amounts of electricity, and must be connected to liquid cooling systems, high-speed cabling, transformers, switchgear, and backup power devices. Goldman Sachs highlights that next-generation AI datacenters often cost between $15m and $20m per megawatt, compared to approximately $10m per megawatt for more traditional facilities.

The most telling example is that of an hyperscaler that may have already secured its chips, signed its construction contracts, and ordered its servers, yet remains waiting for the electrical equipment necessary to commission the site. In such cases, capital is committed and expenditures appear on budgets, but the computing capacity is not yet generating revenue. McKinsey indicates that lead times in North America can reach 80 weeks for certain medium-voltage switchgear and 50 weeks for transformers, making electricity a primary bottleneck in the deployment of AI capacity.

This constraint is also shifting the market's interpretation of the theme, as potential winners are not found solely in semiconductors or major cloud providers. Industrial players capable of rapidly delivering reliable, tested electrical, thermal, and mechanical systems compatible with future chip architectures are positioned to capture a significant share of datacenter-related infrastructure spending. McKinsey further notes that hyperscalers are increasingly prioritizing suppliers able to align early with the roadmaps of GPU manufacturers and proprietary chip designers.

According to reports from Goldman Sachs and McKinsey, the next phase of the cycle should therefore be monitored through highly tangible indicators: 1) lead times for transformers, 2) the adoption of liquid cooling as rack density increases, 3) the cost per megawatt of new datacenters, 4) the actual utilization rate of chips, and 5) the ability of industrial suppliers to remain synchronized with the technical roadmaps of hyperscalers and semiconductor manufacturers.