In this chart

Over the past decade, the share of new global data center construction starts classified as AI-ready or liquid-cooling-capable has gone from a rounding error to near-total market dominance. In 2017, almost all new capacity was built to conventional specifications: retail colocation, wholesale colocation and hyperscale facilities cooled by air. By 2026, the picture has inverted. Most construction starts are either AI-ready or equipped with direct liquid-cooling capabilities designed to support the rack densities that modern AI workloads demand.

The driver is straightforward. NVIDIA’s Blackwell superchips, the backbone of the most intensive AI model training workloads, require densities of more than 100 kilowatts per rack. That is simply beyond what traditional air-cooled data centers can support. Hyperscalers have responded by commissioning facilities with direct-to-chip or immersion liquid-cooling as standard. Meanwhile, model inference, relevant to a far broader group of companies than model training, requires lower rack densities of around 20 to 50 kilowatts per rack, still well above conventional air-cooled thresholds. Operators are racing to brand new cloud facilities as AI-ready and retrofit older ones to capture this demand.

For private-market investors, the shift carries three distinct implications. First, older data centers without future-proofed designs face a growing risk of obsolescence — not at some distant point, but as hyperscaler leases roll and tenants reassess whether existing shells can support next-generation workloads without full redevelopment. Second, the capital intensity of building to current specifications has risen sharply, with developments now frequently requiring billions of dollars of investment and correspondingly larger equity and debt raises. Third, that capital intensity translates into higher depreciation loads for investors to underwrite.

The parallel with office real estate is uncomfortable but instructive. Investors who balked at the capex required to refurbish offices post-pandemic found themselves holding stranded assets. Whether the same dynamic plays out in data centers, with hyperscaler lease expiries as the trigger rather than hybrid working, may be one of the more consequential questions in private infrastructure over the next decade.