01 / The problemBuilt for then, asked to do now
There is nothing wrong with your data centre. It was built well, for the workloads of its day. The problem is that AI changed what a rack does.
A single 8-GPU server draws what a whole rack used to, and an AI rack can draw sixteen times a traditional one. Put that into a hall designed for 5 to 10 kW and you hit four walls at once: not enough power, not enough cooling, not enough floor strength, and a network that cannot feed the GPUs.
Power runs out
Feeds, busway and PDUs sized for 5 to 10 kW racks cannot deliver 40 kW or more to a cabinet, often before you even reach the cooling limit.
Air cannot cope
Past about 30 kW a rack you cannot move enough air to keep the chips in specification. AI density needs liquid, which most legacy halls have no provision for.
The floor is not rated
A loaded, liquid-cooled AI rack can weigh as much as a small car. Raised floors and slabs built for lighter kit may not take it without reinforcement.
The network throttles
AI traffic is east-west and heavy. A 1/10G legacy fabric leaves expensive GPUs idle; AI needs a non-blocking 400G or 800G fabric on low-loss fibre.
02 / The barWhat “AI-ready” actually means
AI-ready is not a label. It is six things being true at once, and falling short on any one makes it the bottleneck. This is the bar your hall is measured against.
- 01
High-density power
Tens of kilowatts to a single rack: high-capacity busway, three-phase rack PDUs and a power chain sized for sustained, near-full load.
- 02
Liquid cooling, or readiness for it
Rear-door, direct-to-chip or immersion, or at least the floor loading, pathways and water to add it. See the liquid cooling guide.
- 03
Capacity and resilience
Enough total capacity, with the right redundancy, for far more load per square metre than the hall was planned around.
- 04
Structural capacity
Floor loading and access routes that take heavy, dense racks, plus the plumbing for a liquid loop where needed.
- 05
A high-speed fabric
Non-blocking 400G or 800G leaf-spine on ultra-low-loss, pre-terminated fibre. See the AI/GPU data hall guide.
- 06
Space and containment
Room to reconfigure into AI-density rows with proper containment; you often cannot simply swap kit into the existing layout.
03 / The checkCheck how AI-ready your hall is
Answer six questions about your hall as it is today. You get a readiness score, a shape of where the gaps are and a prioritised list of what to upgrade: the starting point of any AI retrofit.
Readiness score 6 / 12
Partially AI-ready
A focused retrofit will get you there: targeted upgrades, not a rebuild.
Indicative only. A real assessment looks at far more, from redundancy and electrical phases to water and planning, but this shows the honest shape of the gap and where to start.
04 / The pathClosing the gap without starting over
You almost never need a new building. AI-readiness is overwhelmingly a retrofit: upgrading power, adding liquid cooling, re-cabling and reconfiguring space, usually in phases while the rest of the hall keeps running.
- 01
Start hybrid
Carve out a high-density zone, liquid-cooled AI rows beside the existing air-cooled estate, rather than converting the whole hall at once.
- 02
Upgrade the power path
Higher-capacity busway, three-phase rack PDUs and distribution to the AI zone, with the capacity and redundancy the load needs.
- 03
Add liquid cooling
Rear-door heat exchangers, or direct-to-chip with CDUs and a loop, designed in with the white space rather than bolted on.
- 04
Re-cable for the fabric
Ultra-low-loss OM4 or OS2 and pre-terminated MPO trunks for the 400G or 800G fabric, tested, certified and warranty-backed.
- 05
Reconfigure and contain
Re-lay the rows for density with proper hot or cold-aisle containment, and reinforce floor loading where the heavy racks land.
- 06
Phase it, stay live
Sequence the works zone by zone, so the hall keeps serving while AI capacity comes online.
Minimal disruption
05 / How Optronix helpsFrom “can our hall do AI” to a working high-density hall
Making an existing hall AI-ready is exactly the work we do: assess, design and deliver the retrofit with our own engineers, across Europe, North America, the Middle East and APAC. We turn the question into a costed plan, and the plan into a working high-density hall.
AI-readiness audit
We survey your hall against the real bar, power, cooling, floor loading, fabric and space, and give you a clear, costed gap analysis: what is ready, what needs upgrading and what it takes.
Design the retrofit
A coordinated, costed design for the upgrade, with power, liquid cooling, cabling and white-space changes modelled together, so the build runs cleanly and in phases around your live estate.
Deliver it with the hall live
Our own engineers install the power, cooling, containment and cabling zone by zone, tested and commissioned, bringing AI capacity online with minimal disruption to what is already running.
Or build new
If a retrofit truly cannot get there, we design and build AI-ready halls from the ground up too. See our AI/GPU data hall guide.
Related services
AI readiness
Start with an AI-readiness audit.
We will tell you plainly where the hall stands, what it will take and what it costs, then design and deliver the retrofit.
Sources and notes
- Rack densities and AI-system power from public accelerator and AI-system vendor specifications (8-GPU servers about 10 kW; GB200-class racks about 120 to 132 kW).
- Cooling thresholds and methods per ASHRAE TC 9.9 and common industry practice; the fabric per IEEE 802.3 (400 and 800GbE).
- The readiness checker is an indicative self-assessment to show the gap; it is not a substitute for an engineered survey.
General guidance, not an engineered assessment. Densities, thresholds and the checker score are indicative; real AI-readiness depends on your specific power, cooling, structural, electrical and planning constraints. Confirm with a site survey and an engineered design.