AI factory design – from high density designs to complete AI factories

Starting point for data centre design

A data centre reference design is a pre‑validated, fully engineered blueprint that defines how power, cooling, and IT infrastructure should be organised to build a reliable, scalable, and AI‑ready data centre.

Uses proven, standardised layouts.
Ensures infrastructure compatibility.
Reduces risk during deployment.
Supports AI workloads.
Enables faster data centre site build-outs.

Working closely with all the leading GPU, TPU, server manufactures, and integrators, we continue to expand our library of fully engineered reference designs to address the fast pace and change the industry is experience. AI demands more 'reference designs' we deliver.

If your existing infrastructure is AC-based, the near-term question is not whether to rip and replace, it is where to introduce DC conversion and how close to the rack you can place it. The practical tradeoff is between preserving upstream AC distribution, adding rack-adjacent power sidecars, and reserving enough physical and electrical space to support future 800 VDC loads. Sidecar or power-rack approaches allow operators to host 800 VDC equipment without converting the whole data centre to DC, which matters because early 800 VDC adoption will apply to specific high-density AI clusters, not every load in the building. Larger future expansions may justify more centralised 800 VDC support, but that becomes a data hall or cluster-scale optimisation rather than a near-term requirement for the entire site.

The key tradeoff here is flexibility versus isolation. A shared power chain can give operators more room on days when mechanical load drops (e.g., when chillers are not working as hard and more of the available capacity can support IT load). A separated architecture gives the IT side a cleaner boundary for future changes, including 800 VDC upgrades, but it also makes mechanical power harder to reallocate because that capacity sits inside its own train. Before committing to either structure, operators should pressure-test seasonal operating modes, maintenance scenarios, generator loading, CDU power paths, and the expected split between air-cooled, liquid-cooled, and future DC-powered IT.

Two paths exist, and the right one depends on whether the workload can tolerate interruption. For high-availability inference or workloads that cannot move cleanly to another site, UPS remains the closer-in protection layer because it supports uninterrupted ride-through, short-duration disturbance protection, and AI load smoothing. BESS plays a different role: it supports grid services, power capping, renewable utilisation, longer-duration stored energy, and load smoothing. A useful distinction is that the BESS can typically store one to two hours of energy and acts more like a line-interactive energy source. In brief, BESS can be used to strengthen a site energy strategy, but do not assume it replaces UPS until the workload and ride-through requirements are known.

Prefab can cover more than power skids, but operators should think in terms of which site activities they want to remove from the critical path. Major prefabricated elements can include medium-voltage gear, e-houses, low-voltage power trains, Power Skids, IT pods, integrated racks, packaged chillers with onboard pumps, and selected piping or service-hallway assemblies. The practical value comes from reducing field coordination amongst electrical contractors, mechanical contractors, construction managers, and commissioning teams. Power Skids can integrate UPS, switchgear, batteries, ATS, and management software, with listed configurations including 1,250 kW at 480 V/60 Hz and 2 MW, 2.25 MW, and 2.5 MW 400 V models.

A useful way to think about reference designs is not "standard versus custom," but "validated starting point versus blank-sheet engineering." AI-ready reference designs can shorten the planning cycle by giving project teams a pre-engineered baseline for power, cooling, IT layout, and controls integration. Current Schneider Electric reference designs include 7.392 MW NVIDIA GB200 NVL72 liquid-cooled AI clusters, 7.536 MW NVIDIA GB300 NVL72 designs, 10 MW–12.4 MW Vera Rubin NVL72 designs, and 1 MW 12-rack prefab AI solutions. The trade-off is that modular elements can be prefabricated in major chunks, whilst the complete site architecture still requires integration across utility interconnect, mechanical plant, medium-voltage distribution, IT deployment strategy, and construction sequencing.

You can use redundant capacity for more compute, but operators should treat that as an operating-mode decision rather than free capacity. Traditional concurrently maintainable facilities reserve electrical and mechanical headroom so systems can come offline for maintenance whilst the committed IT load stays online. AI changes the economic pressure because grid capacity, deployment speed, and compute demand make idle redundant capacity harder to justify. For example, a facility designed to support 200 MW of IT load during concurrent maintenance may physically support a higher operating point, such as 240 MW, but the extra compute comes from capacity that previously protected maintenance and failure scenarios. The real decision is whether the business values additional tokens per facility more than it values the original redundancy strategy. Before tightening margins, define the operating state explicitly: normal redundant mode, high-throughput mode, maintenance mode, and failure response mode.

The workload should drive the design once the operator knows whether the site will support general colocation, high-density neocloud customers, training clusters, inference platforms, or a mix of those models. A general colocation facility may need to preserve optionality at the bulk power and cooling level until tenant requirements become clear. A focused AI cloud or single-purpose AI factory can design directly around a known GPU roadmap, cluster topology, rack density, and cooling requirement. Schneider Electric's AI data centre offer frames the same problem around integrated infrastructure for cooling, power, software, pod and rack systems, and lifecycle services for high-performance AI workloads.

Do not design the entire data centre as if every load will become 800 VDC. The practical design question is where 800 VDC enters the architecture and how large the DC conversion block needs to be. Early deployments can use rack-adjacent sidecars, often aligned to one or two GPU racks, inside an existing AC data centre without a major upstream retrofit. Larger future clusters may justify centralised 800 VDC blocks in the 2 MW to 5 MW range, potentially scaling towards larger chunks, whilst longer-term architectures may bring medium voltage into solid-state transformer-based conversion at roughly 10 MW-class scale. 800 VDC architectures address rack densities moving beyond 400 kW by relocating AC-to-DC conversion outside the IT rack and using power racks or sidecars to reduce disruption in existing AC-based facilities.

The key trade-off is isolation versus dynamic capacity sharing. Separate power trains can simplify boundaries between IT and cooling infrastructure, especially when large chillers, CDUs, and data halls do not align cleanly in the same increment. Shared infrastructure can preserve flexibility during free-cooling conditions because unused mechanical capacity may be redirected towards IT load if the transformer and downstream architecture allow it. The risk with full separation is stranded power: if mechanical load drops but the IT transformer already operates at its limit, the operator cannot recover that unused cooling-side capacity for compute. Gap to flag: this design choice needs site-specific modelling around data hall size, chiller block size, transformer allocation, redundancy target, and seasonal operating modes before it becomes a defensible standard.

Electrical, airflow, and liquid flow modelling should happen before finalising the design because each model addresses a different class of failure. Electrical modelling validates fault current, protection, safety, reliability, and what-if scenarios around maintenance or added load. CFD or airflow simulation validates containment, air-side capacity, and redundant cooling behaviour. Piping and fluid-flow modelling validates whether every rack receives the required coolant flow during normal, redundant, and maintenance states. The next step is coupling IT load behaviour to facility behaviour, so AI load fluctuations can propagate through power systems, CDUs, chiller controls, and cooling loops before the site operates under real load.

The model should start in design and mature towards operational what-if analysis as live data becomes available. A static ETAP-style model can already support offline scenarios such as maintenance switching, adding equipment, or testing whether a configuration stresses the electrical infrastructure. The stronger operating model connects real load data to the simulation layer so operators can evaluate what happens when chillers come offline, loads shift, or large AI clusters ramp up before executing changes in the field. EcoStruxure Power Monitoring Expert supports power analysis, voltage disturbance waveform analytics, alarm grouping, system capacity tracking, and power quality analysis, whilst EcoStruxure IT supports monitoring, planning, modelling, power and energy analysis, and cooling optimisation.