For decades, cooling was the department you called after the important decisions were made. The IT team specified the compute, the real-estate team found the building, and somewhere downstream a mechanical engineer was asked to make the heat go away. At 5–10 kW per rack, that sequence worked. At AI-era densities, it is a recipe for a stranded asset.
The density break
Modern AI training racks draw an order of magnitude more power than the enterprise racks most existing facilities were designed around. Air simply stops being a viable primary coolant well before these densities — the physics of moving enough air through a rack becomes absurd long before the economics do. Direct-to-chip liquid cooling and, for some workloads, immersion are no longer exotic options; they are the default assumption for serious AI capacity.
This has a consequence that many operators have not fully internalised: cooling now shapes the building, not the other way around. Structural loading for liquid distribution, plant space for coolant distribution units, floor layouts organised around manifold runs, commissioning regimes for liquid loops — these are architectural decisions. Retrofitting them into a facility designed for air is possible, but it is expensive, disruptive and rarely elegant.
“At AI densities, the thermal architecture is the facility. Everything else is fit-out.”
The AI Thermal Factory
This is why we increasingly talk about the AI Thermal Factory: treating the thermal system as a productive asset to be designed, optimised and even monetised — rather than a cost to be minimised. Three shifts define the approach.
- Design from the chip outward. Start with the workload's thermal profile and let it drive rack architecture, room layout and plant sizing — not the reverse.
- Treat heat as a product. High-grade waste heat from liquid-cooled AI facilities can serve district cooling and heating, desalination pre-heat and industrial processes. In the right location, heat reuse turns an operating cost into a revenue line and a planning advantage.
- Measure what matters. PUE remains useful, but water usage effectiveness and cost-per-token-cooled tell you more about an AI facility's real efficiency. Chasing a headline PUE while consuming unsustainable volumes of water is not optimisation — it is accounting.
For the Gulf in particular
In hot climates, these questions carry extra weight — and, counterintuitively, extra opportunity. Liquid cooling's efficiency advantage over air grows as ambient temperatures rise, and the region's investment appetite means new facilities can leapfrog directly to thermally-native designs rather than dragging air-era assumptions forward. The operators who treat cooling as a design parameter from day one will build the region's most efficient AI capacity. The ones who don't will spend the next decade paying for the retrofit.
About the author
Saandeep V. Dandekar
Executive Director — Infrastructure Strategy & Delivery, MetaDecrypt AI Atlas