Key takeaways

  • AI GPU nodes draw 10–15 kW each. A 10U cluster hits 100–120 kW per rack — 10× the legacy data centre average.
  • Traditional air cooling cannot remove heat at that density. Liquid cooling is no longer optional for serious AI deployments.
  • Power distribution, structural loading, and smart hands skills all need to be re-evaluated before you rack a single GPU node.

For the better part of two decades, a 10 kW rack was considered dense. The industry built its cooling standards, its power distribution architecture, and its SLA expectations around that number. Then Nvidia shipped the H100 — a single GPU drawing over 700W — and everything changed.

Today, a rack of eight dual-GPU servers can draw more power than a typical house uses in a month. The 120kW rack is not a theoretical edge case; it is the near-term baseline for serious AI training and inference infrastructure. And the data centres built during the cloud era — raised floor, CRAC unit, 32A circuit — were not designed for it.

Why 120kW matters so much

The number is significant for three independent reasons that compound on each other:

  • Cooling physics. Air has a low heat capacity. At 120kW per rack, moving enough air to absorb that heat becomes mechanically impractical — the volume and velocity required would create noise and airflow pressures that damage equipment. You need liquid, whether direct liquid cooling (DLC) through cold plates, rear-door heat exchangers, or full immersion.
  • Power infrastructure. A standard data centre cabinet gets a 32A single-phase circuit. 120kW at 230V three-phase is around 174A. You need new PDUs, new busways, and typically a dedicated sub-distribution board for an AI cluster — work that happens long before a GPU is racked.
  • Structural loading. A fully populated H100 rack weighs over 1,500kg including liquid cooling manifolds. Most raised-floor systems are rated for 700–1,000kg per rack. You may be installing on slab — and you need to know before the kit arrives.
The real surprise for most IT teams

It is not the cost of liquid cooling that shocks people — it is discovering that the colocation facility's power circuit, floor loading, and cooling capacity are all inadequate after the hardware has already been ordered. Pre-deployment site surveys are not optional at this density.

What the leading colocation facilities are doing

Facilities that have invested ahead of the curve — several London and Manchester colocations included — are deploying purpose-built high-density zones with rear-door heat exchangers or direct liquid cooling manifolds pre-plumbed to each rack position. These zones carry dedicated power circuits of 100A+ per rack and reinforced slab floors. The smart hands work in these environments looks completely different from a standard deployment.

For facilities not yet retrofitted, AI deployments are often contained to specific zones with bespoke cooling plant, or are staged across multiple standard racks to spread the power and thermal load — accepting a performance trade-off to fit within existing infrastructure limits.

The smart hands implications

This density shift creates real challenges for on-site engineers:

  • Rail kits and cable management. GPU servers have vendor-specific rail kits, and InfiniBand or NVLink cables — thicker and stiffer than standard Ethernet patch cables — need to be dressed without compromising airflow or bending radius. Getting this wrong can throttle GPU interconnect bandwidth.
  • Liquid cooling connections. Connecting and pressure-testing liquid cooling circuits, bleeding air from manifolds, and verifying flow rates before power-on requires specific training. A failed coupling in a live DC is a serious incident.
  • High-voltage awareness. Three-phase PDUs at 60A+ per circuit require electrically aware engineers. Not all smart hands providers have the certification or insurance to work within an arm's reach of live high-voltage distribution.
  • Documentation. Cable labelling and topology documentation for an AI fabric — where every GPU needs to reach every other GPU at line rate — is significantly more complex than a standard server deployment. Mistakes show up as training job failures days later.

What to ask before any AI deployment

Whether you are deploying into a colocation facility or an on-premises data hall, these are the questions to answer before the kit ships:

  1. What is the available power per rack position, and what is the circuit type and amperage?
  2. What cooling infrastructure is available — and has it been load-tested at the density you need?
  3. What is the floor loading capacity, and do you need a structural survey?
  4. Does the facility have pre-plumbed liquid cooling, or do you need to bring your own rear-door heat exchangers?
  5. Are your smart hands engineers certified and trained for high-density GPU deployments specifically?
The data centre industry built itself for 10kW racks. AI wants 120kW. That is not an incremental upgrade — it is a fundamental rethink of what "ready for compute" actually means. — DACPROS Senior Infrastructure Engineer

Where this goes next

Rack densities are not plateauing. The next generation of GPU nodes — already announced — will push individual server draw past 1kW per GPU, and full-rack densities toward 200kW are realistic within two years. Immersion cooling, once a niche solution, is becoming a mainstream consideration for any facility building new capacity today.

For IT teams managing the transition: the engineering skills, the infrastructure checklist, and the smart hands partners you choose for your first AI deployment will determine how fast and how cleanly you can scale from there. See how DACPROS handles high-density deployments, or talk to our team about a site survey before your GPU kit arrives.