Building AI data centres: what it actually takes in 2026.
A technical brief for operators, investors and end-users planning AI-ready capacity. Power, cooling, cost, timeline, and the operational risks that quietly decide whether a programme lands on schedule.
What an AI data centre actually is — and whether you can get there from here.
The gap between a traditional enterprise hall and an AI campus is not incremental. Density, electrical topology and cooling have all moved together, and the design decisions you make at concept stage outlive the first three GPU generations.
AI vs traditional data centre
A traditional facility is designed for 5–10 kW racks of mixed enterprise compute. An AI data centre is engineered around dense GPU pods: 100–300 kW racks today, 600 kW on the roadmap, liquid loops to the chip, low-latency east-west fabric (InfiniBand or Spectrum-X), and power topology that treats a training cluster as one synchronous load rather than thousands of independent servers.
Retrofit vs greenfield
Most enterprise halls cannot be retrofitted economically beyond ~30–50 kW per rack. The constraints are slab loading for liquid CDUs, riser space for secondary fluid loops, and incoming utility capacity. Retrofit makes sense when the shell, substation and water permits already exist; greenfield wins when you need >100 kW per rack at scale or behind-the-meter generation.
Avoiding obsolescence on day one
Specify to the next GPU generation, not the current one. Design the white space for 130–200 kW per rack with headroom to 400 kW, oversize busways and manifolds, and use modular CDUs that can be uprated. Lock the mechanical and electrical topology; keep the IT layer swappable.
The hardest problem is no longer compute. It is electrons.
Grid interconnection queues, behind-the-meter generation, and the move from 300 W to 2,000+ W per square foot are reshaping where AI capacity can physically exist.
Where the megawatts actually come from
Tier-1 grid interconnection queues in the US, UK and Ireland now run 4–7 years. Practical paths to power: acquiring sites with existing capacity allocations, co-locating with retiring thermal generation, behind-the-meter gas or hydrogen-ready turbines, and on-site renewables with battery firming. Power strategy is now a site-selection question, not an engineering one.
Behind-the-meter generation
BTM gas turbines or reciprocating engines can deliver 50–500 MW in 12–24 months versus 4–7 years for a grid upgrade. Trade-offs: fuel-price exposure, emissions reporting, and the cost of redundancy. Most operators pair BTM generation with a smaller grid tap for black-start and resilience.
From 300 W to 2,000+ W per square foot
The jump is not a linear upgrade — it changes the entire electrical topology. Medium-voltage distribution moves closer to the racks, busways replace whip cabling, and UPS strategy shifts toward lithium and rack-level batteries. Expect to spend more on copper and switchgear than on the IT load itself.
Liquid is now the default, not the upgrade.
Rack densities have crossed the threshold where air cooling is a niche choice. Designing the secondary fluid loop is now on the critical path.
Density per rack, today and projected
Today: 100–300 kW per rack is in production for H100/H200 deployments. 2026–2027: 600 kW per rack for Blackwell Ultra and Rubin-class systems. Designing for less than 130 kW per rack is a short-lived facility decision.
Liquid cooling: mandatory or optional
Air cooling remains viable up to ~40 kW per rack with rear-door heat exchangers. Above that, direct-to-chip liquid is required, and immersion is a credible option for the densest pods. Most 2026 builds are hybrid: liquid to the GPU/CPU cold plates and air for memory, networking and ancillary.
CDUs, manifolds and piping without slipping the schedule
Treat the secondary fluid loop as a long-lead item: order CDUs and manifolds before the building is closed in, pre-fabricate piping skids off-site, and standardise on one quick-disconnect family. Commissioning the loop in parallel with the white-space fit-out is what compresses the last six months of the programme.
The numbers behind a 2026 build.
Capex per megawatt has roughly doubled since 2022. Lead times for the long-lead items have stretched into multi-year territory.
| Metric | Value | Note |
|---|---|---|
| Cost per MW (2022) | $5M–$12M | Air-cooled, ~10 kW per rack baseline |
| Cost per MW (2026) | $9M–$25M | Liquid-cooled, 130 kW+ per rack, dense electrical |
| Transformer lead time | 128+ weeks | Medium-voltage units; switchgear similar |
| Generator lead time | 60–90 weeks | 2 MW+ standby diesel and gas |
| Build window (greenfield) | 24–48 months | Site to first power, dependent on grid |
| Liquid CDU lead time | 30–52 weeks | MW-class units from tier-1 vendors |
The risks that quietly decide the programme.
Supply chain, talent and sustainability are not soft topics on an AI build — they are the constraints that move first-power dates by quarters.
Supply chain when backlogs are years long
Order long-lead items at concept-design stage, not at construction start. That means transformers, switchgear, CDUs and chillers ordered before drawings are issued for construction. Hold framework agreements with two vendors per critical item, and treat allocation slots as a strategic asset.
Finding commissioning engineers for liquid-cooled AI
The talent pool for >100 kW liquid-cooled commissioning is small and concentrated around a handful of hyperscaler programmes. Practical answers: partner with specialist commissioning agents, invest in graduate programmes 12–18 months ahead of energisation, and standardise on documented procedures so the work is repeatable across sites.
Sustainability targets against AI consumption
A 100 MW AI campus consumes as much as a small city. Credible decarbonisation routes: 24/7 matched PPAs (not annual offsets), waste-heat reuse into district heating, closed-loop cooling to remove water draw, and locating compute next to renewable generation rather than load centres. Reporting needs to track Scope 2 hourly, not annually.
Questions we hear most from operators and investors.
Planning capacity, evaluating a site, or reviewing a build?
setloop.io works with operators, investors and end-users on AI data centre feasibility, technical due diligence and architecture — from site evaluation through to commissioning.