What it takes to build an AI data centre

The AI build-out has turned data centres into one of the most sought-after infrastructure assets in the world. There is more capital chasing the sector than at any point in my twenty years in it, and a new announcement lands almost every week promising capacity at a scale nobody was discussing three years ago.

I have spent most of my career on the other side of those announcements. I started building hosting infrastructure at 17, grew it into a regional data centre business, and ran a facility that held verified 100% uptime for more than a decade before it became part of Contabo. That experience left me with an unfashionable view of the current moment: the money is the easy part.

Powered land, not land

Financing looks like the barrier to entry until you try to energise a site. Power availability depends on real grid access, and a site without it has limited value no matter how large the cheque behind it. A site that is already powered and permitted can be worth several times more than comparable undeveloped land, because years of regulatory and infrastructure work are already banked in it.

If every investment proposal in this sector became a data centre, there would be nowhere left to stand. Most never leave the slide deck. They die on construction risk, on operational complexity, on permitting timelines, on market conditions that move faster than the build.

The shortage nobody prices correctly is people. The AI boom has not been running long enough to produce a surplus of experts, and the pool of professionals who have actually delivered and operated large-scale facilities is small. It gets smaller once you filter for those who have made the jump to AI-specific density.

There is no perfect location

Location is an amalgamation of competing priorities rather than a checklist. Electricity price matters, ideally with room to negotiate. So do long-term grid capacity, political stability, environmental risk, connectivity, latency requirements and local planning rules. Free cooling can transform operating efficiency, but only where the climate allows it. You will not get in Nigeria what you get in Norway.

These variables are not portable across borders. Even comparing electricity costs between European markets requires genuine local expertise, because pricing structures, regulatory frameworks and subsidy mechanisms differ more than the headline numbers suggest. Every jurisdiction needs its own analysis.

The older cloud footprint clustered around Frankfurt, London, Amsterdam, Paris and Dublin because connectivity drove the decision. AI workloads are more tolerant on latency and far less tolerant on power, so they are moving into secondary markets where capacity is actually available. That shift is the single biggest change in site selection since we started.

Designing for hardware that does not exist yet

A facility being designed today will have to host silicon nobody has shipped. Racks that once drew around 10 kW are now approaching 150 kW in the same footprint, and the next generation goes further. Design only for current demand and you commission a building that is already obsolete.

This is why we plan for maintenance, resilience, serviceability and upgrade paths from the first drawing rather than the first retrofit. A facility built for extended operation produces better outcomes over its life, and the cost of that discipline is always lower at design stage than at year five.

Uptime is a thousand small decisions

Investors want confidence that a facility will perform, particularly where new infrastructure has not yet proven itself under real workloads. A track record of delivering and operating is therefore part of the underwriting, not a nice-to-have.

Everyone wants maximum availability. Very few achieve it. Having operated a data centre that held 100% uptime for over a decade, I know it is not a certification, it is a habit. Redundancy standards differ, and meeting one does not eliminate operational risk. Every additional layer of resilience is a deliberate engineering choice with a cost, and the strongest operators tend to exceed standard requirements across cooling, power distribution and security. Those decisions are unpopular during design because they add complexity. They become obvious the first time they prevent a customer-affecting outage.

The risk people underestimate most is human. Faults happen at 3am, and they become disasters when a tired engineer walks into a switching room and makes one wrong move. That is why we treat process discipline, automation and written operational procedure as engineering, not administration. Aviation solved a version of this problem decades ago with standardised checklists and processes designed around fatigue and pressure. I fly commercially, and I have yet to find a better model for high-consequence operations.

Where this leaves new entrants

Demand is extraordinary and it is drawing in newcomers, which is healthy. But this is not a business you enter on capital alone. Power strategy has to be settled long before anyone breaks ground. Location needs a serious read of energy markets, regulation, connectivity and environmental planning. The design has to anticipate future hardware, and the operating model has to be built for decades of service rather than the first tenancy.

At Gi21 we are applying exactly this playbook to a platform of up to eight high density data centres across Europe, with more than 640 MW of combined IT load capacity in preferred locations and construction starting at the first two sites in H1 2027. We are building it as operators who have run this kind of asset before, because the industry's competitive advantage is no longer access to capital. The money is here and there is plenty of it. Experience is what turns it into infrastructure that still performs long after the ribbon cutting.

A version of this article was first published in Data Centre Review.