01Why AI breaks the data hall
For twenty years, a data hall was a power and cooling problem solved at around 5 to 10 kW per rack. AI training rewrote that.
GPUs are power-dense by design: thousands of them, packed tight, running flat out for weeks, and they need to talk to each other at enormous bandwidth. The result is a facility where power density, liquid cooling and the internal network are the design, and everything else follows. A hall built for yesterday's densities cannot simply be upgraded for AI: the power, cooling and weight assumptions are different from the floor up.
The jump from virtualisation to AI training is not incremental. It is a step change of five to ten times that goes straight past what air can carry and into liquid cooling, reinforced floors and a different power and network design.
02The density problem, and what it demands
Everything starts with the power budget per rack, because it sets the cooling, the electrical distribution and even the structural design.
Drag the density below to see what a rack at that power needs: the cooling that can carry it, the heat it rejects and what it costs to run.
Interactive / Rack density and cooling calculator
Set the power per rack
Cooling methods light up where they can carry the load, and the readouts show the heat to reject and the annual energy.
30kW / rack
High density / AI inference
Cooling that can carry this density
At 30 kW, well-managed air can still cope.
03The cooling ladder: air to immersion
As density climbs, cooling moves from blowing cold air to bringing liquid ever closer to the silicon, because water carries roughly 3,500 times the heat of the same volume of air.
Each rung handles more heat, at more cost and complexity. Most AI halls today land on direct-to-chip liquid, often hybrid with air for the parts that do not need it.
| Method | Density ceiling | How it works | Trade-off |
|---|---|---|---|
| Air and containment | ~30 kW | Hot and cold-aisle containment with CRAC or CRAH units | Cheap and familiar, but runs out fast for AI |
| Rear-door heat exchanger | ~70 kW | A water-cooled coil on the back of the rack | A big step up that keeps the air model, but needs water to the rack |
| Direct-to-chip (DLC) | ~150 kW | Cold plates on the GPUs and CPUs, on a liquid loop | The AI workhorse: plumbing, CDUs and a leak-management regime |
| Immersion | ~200 kW+ | Servers submerged in dielectric fluid | The highest density and efficiency, and the biggest operational change |
Ceilings are indicative and depend on coolant temperatures, rack layout and the specific equipment. Two numbers matter alongside density: PUE (liquid pushes it towards ~1.1, against ~1.5 for air) and water. Liquid cooling needs a facility water loop, coolant distribution units (CDUs) and a maintenance discipline that air never required. Designing the white space, the loop and the containment together is the difference between a hall that works and one that throttles. The liquid cooling guide covers the methods in depth.
04The 800G network that feeds the GPUs
In a traditional hall most traffic is north-south, in and out of the building. In an AI cluster it is overwhelmingly east-west: GPUs exchanging gradients with each other, constantly, at line rate.
A slow or congested fabric leaves very expensive GPUs sitting idle, so the network is built as a non-blocking, low-latency leaf-spine: today on 400/800GbE, heading to 1.6T.
Every leaf reaches every spine, so any GPU can talk to any other at full bandwidth: non-blocking, east-west.
800G spine to leaf, ultra-low-loss fibreLeaf to GPU, in rackEast-west traffic
East-west is the workload
GPU-to-GPU traffic dominates. The fabric has to be non-blocking, with full bandwidth between any two endpoints, or expensive accelerators stall waiting on the network.
Rail-optimised
AI fabrics are often wired so each GPU's NIC connects to its own “rail” of the network, spreading collective traffic evenly and keeping latency predictable across the cluster.
Low loss is non-negotiable
At 800G the optical loss budget is tiny. Ultra-low-loss connectors, clean MPO trunks and certified links are what let the fabric run at rate. The structured cabling guide explains why.
Pre-terminated and dense
Base-8 and base-16 MPO trunks, pre-tested and ready to plug, handle the sheer fibre count of an AI fabric: faster to deploy and far cleaner to certify at scale.
05What holds it all up
Density does not just stress power and cooling. It changes the room.
A fully populated AI rack can weigh as much as a small car, draw three-phase power through busway, and demand containment and cable pathways planned around the liquid loop. The physical design is where an AI hall is won or lost.
Power distribution
High-density rack PDUs fed from overhead busway, three-phase at the rack, and a power chain sized for sustained near-100% load: AI runs flat out, not in bursts.
Structure and weight
A loaded liquid-cooled rack can exceed 1,500 kg. Floor loading, plinth design and access routes have to be engineered for it; you cannot retrofit a slab.
The liquid loop
CDUs, manifolds, leak detection and a facility water loop, designed alongside the white space rather than bolted on. Serviceability and redundancy of the loop become uptime-critical.
Containment and airflow
Even liquid-cooled halls have an air component. Hot and cold-aisle containment, blanking and airflow management still matter for the heat that stays in air.
Structured cabling at scale
Ultra-low-loss OM4 and OS2 and pre-terminated MPO trunks, designed for the 800G fabric and its breakouts, with the counts taken straight off the design model.
White-space planning
Rack layout, pathways, the loop and the fabric planned as one. Get the white-space design right and the hall scales; get it wrong and you strand power, cooling or ports.
An AI hall is where the whole lifecycle converges. The 800G fabric rests on the fundamentals in the structured cabling guide, it has to be run to the disciplines in data centre operations, and the facility itself is chosen against the criteria in how to choose a data centre.
06How Optronix helps
AI and GPU halls are exactly the work we are built for: high-density, 800G, liquid-ready, designed and delivered end to end.
We design, supply, install, certify and migrate across the full data centre lifecycle, for hyperscale, colocation and enterprise clients across the UK, EMEA and worldwide. Power distribution, containment, the liquid loop and white-space fit-out are our everyday work.
Model-derived design, fast
We take a project from a single client drawing to a costed, model-derived design (densities, cooling, fabric and cabling counts) in under 48 hours. The counts come straight off the model, so the design you sign off is the one we build.
Certified, warranty-backed cabling
The 800G fabric is installed as ultra-low-loss, pre-terminated, fully tested and certified plant, backed by a 25-year manufacturer system warranty through our accredited installer status with the major cabling manufacturers.
- 01
Design and model
Density, power, cooling strategy, fabric and cabling counts, modelled and costed from your drawing, fast, so decisions are made on real numbers.
- 02
Supply and install
Power distribution, containment, liquid-ready white space and the 800G cabling plant, supplied and installed by our own engineers, not subcontracted.
- 03
Test and certify
Every link tested and certified to standard, documented and warranty-backed, so the fabric runs at rate from day one.
- 04
Migrate and operate
Staged migration of live estates and ongoing operational support (smart hands, moves and changes) across the lifecycle.
AI and high-density builds
Send us the drawing and the target densities.
We will come back with a costed, model-derived design (power, cooling, 800G fabric and cabling) built to scale.
Sources and notes
- Rack densities and accelerator and system power from public GPU and AI-system vendor specifications (for example 8-GPU servers at ~10 kW; GB200 NVL72-class racks at ~120 to 132 kW).
- ASHRAE TC 9.9 thermal guidelines and common industry practice for air, rear-door, direct-to-chip and immersion cooling density ceilings and PUE.
- IEEE 802.3 (400/800GbE) and leaf-spine and rail-optimised fabric design as used in AI clusters; cabling per TIA-942 and ISO 11801 (see our structured cabling guide).
- Uptime Institute on facility power, water and operations for high-density compute.
General guidance, not a design specification. All densities, cooling ceilings, PUE and energy figures are indicative and depend on the specific equipment, coolant temperatures, layout and site. Calculator figures are rounded for clarity; energy uses a representative PUE for the simplest cooling method that can carry the load and a £0.20/kWh illustration. Confirm against vendor specifications and an engineered design.