120 kW per rack: half of all AI data centres age out, the other half double in value
Compute industry notes · GeekWaedon @GeekWaedon
GeekWaedon @GeekWaedon · November 2026
The point: the valuation formula for AI data centres is being rewritten. What sets an asset's worth is no longer floor area but kilowatts per square metre. A 120 kW rack is a hard threshold: cross it and you own a compute asset; fail to cross it and you own a warehouse.
Two jumps in power density
Traditional enterprise racks run at 5–10 kW, the previous AI generation at 20–40 kW, MGX is designed for more than 120 kW, and DSX Pod integrates seven NVL72 racks in one container (roughly 1,000 kW per unit). In four years rack density has risen by an order of magnitude — far faster than the design margins of existing power and cooling systems.
The result is a brutal split: high-density halls (new AI data centres and retrofits with water loops and high-voltage supply) take the incremental GB300/Rubin orders, while low-density legacy halls are left with inference at the edge, edge nodes and conventional enterprise loads, with rents and valuations drifting down. The examples are already here: Chennai, India is becoming an AI data centre cluster — Vultr announced a 20 MW AMD MI355X deployment, CapitaLand is investing about INR 1.94 billion in a facility with 34 MW of IT load and 53 MW total power, and STT GDC, CtrlS and Sify are expanding in the same city. All of these new builds are designed for high power density.
The power constraint reshapes the supply chain
Once 120 kW per rack and 1 MW per pod become standard, grid connection capacity is scarcer than land. That explains three things: hyperscalers signing long-term power purchase agreements directly with generators; NVIDIA offering container-scale pre-integration (DSX Pod) to compress the time from power-ready to compute-lit; and European vendors — for example 2CRSi with its ÆTHER project, a planned 20 MW AI Factory scheduled for Q3 2027 — packaging power plus compute as a single product. Suppliers of power conversion, liquid cooling and high-voltage DC equipment consequently hold some of the most certain order books in this infrastructure cycle.
Bottom line
Judge any AI data centre asset or project by three numbers first: kW per rack, expandable grid capacity, and liquid-cooling readiness. Assets that can answer enjoy a scarcity premium during a compute shortage; those that cannot are structurally losing value regardless of location or rack count. 120 kW is the dividing line.
Frequently asked questions
Why does rack power density matter so much?
AI server power has been doubling generation by generation and MGX racks are designed for 120 kW plus. Halls without the power and cooling headroom cannot host the next generation of equipment, which directly determines whether the asset is usable and what it is worth.
Is upgrading an older hall realistic?
It needs water loops, high-capacity power and floor loading at the same time, and retrofits priced per watt are expensive. In some cases a new high-density facility — or a pre-integrated design like DSX Pod — is more economical.
What matters most in AI data centre siting?
Grid connection capacity and cost come first, then network and climate (which affects cooling efficiency). Emerging clusters in Chennai, Ohio and elsewhere are all laid out around power availability.