# Density Tiering Principle Don't try to cool 160kW racks in the same thermal envelope as 20kW colo racks. It doesn't work — the cooling architecture, airflow management, and failure modes are fundamentally different at each density tier. **The tiers:** - **≤20kW/rack (colocation):** Standard MDC with DX cooling. Air-cooled condensers, nothing exotic. This is the bread-and-butter modular deployment. - **30-40kW/rack (inference):** Delta AI CDC container. Bolt-on unit with rear door heat exchangers. Self-contained cooling loop that handles the higher density without redesigning the base MDC. - **~160kW/rack (training — GB300):** Entirely separate conversation. Liquid cooling, different physical form factor, different operational model. This is not an upgrade from the MDC — it's a different product category. **Why separation matters:** Mixing densities in a single enclosure forces the cooling system to serve two masters. You either over-provision for the low-density racks (wasting capital) or under-provision for the high-density ones (thermal throttling). The physics don't compromise — you shouldn't either. **Decision rule:** Match the physical enclosure to the workload tier. Don't design a universal box. Design three boxes, each optimised for its density range. --- See also: [[Delta AI CDC Container]] | [[Modular Data Centers MoC]]