This is the first article in ‘Beyond Bespoke: Rethinking Data Centre Delivery’, a new six-part series from Bryden Wood led by Stefanos Gkougkoustamos, Architecture Director and Data Centre Lead. The series asks why the sector still delivers each facility as a one-off when the facilities themselves are diverging, and what a productised alternative looks like in practice. This opening piece sets out the argument: how power, workload and location are pulling data centre design in different directions, and why delivery has to change to keep up.
How AI infrastructure is diverging, and why delivery must change to keep up.
For most of the last decade, a data centre was recognisably one thing: a low-rise building designed from the inside out, its form dictated by the power and mechanical plant it existed to serve. Capacities varied, resilience levels varied, cooling strategies varied at the margins, but a designer moving between projects was moving between variations of a single building type. The industry worked from one basic model, adapted to fit the site, and the way those buildings were delivered reflected that.
That coherence has since come apart, and delivery has not kept pace.
Three conditions, now diverging
First is power. Site selection used to begin with connectivity, land and proximity to the markets a facility served, and electricity was a constraint to be resolved rather than the question that decided the answer. That has now inverted. On most large schemes we see, the governing criterion is where a grid connection of the required scale can be secured within a timeframe an operator can plan around, and that points steadily away from established hubs and onto land with no digital infrastructure history at all. The result is a category of facility whose location is determined almost entirely by the electrical network, and whose design follows from that.
Second is workload. The balance of what these buildings do has been moving from training artificial intelligence (AI) models towards running them, and that changes the demand profile inside them: different utilisation patterns through the day, different redundancy expectations, different tolerance for interruption. It arrives alongside a continuing climb in the power drawn by each rack of equipment, and a cooling landscape that is separating rather than converging. Conventional air cooling, liquid delivered directly to the processor, and servers immersed entirely in fluid now coexist as legitimate answers, sometimes within the same estate. A building designed around one of them is not readily adapted to another.
Third is the building itself. Alongside the remote campus sits a smaller and more constrained category: facilities within cities, often inside existing structures, sized by what the fabric and the local supply can carry rather than by what the operator would choose. The logic there is different again. It rests on proximity to the demand being served, on floor area that already has power in it, and on rules that specify the jurisdiction a facility must sit in. These are not scaled-down versions of the large campus. They are a different problem with a different set of constraints.
All three conditions are still moving, and each pulls facility design in a different direction.
Delivery has not moved with them
What has not changed is how these buildings reach site. In most of the market, each facility is still approached as a discrete project: designed largely from first principles, procured against a specification written for that scheme alone, assembled on site with much of the work carried out in place, and commissioned at the end of a programme that has usually absorbed every preceding delay.
The consequences will be familiar to anyone who has delivered one. Long-lead equipment, ordered many months ahead of need, is committed against a design that is still moving. A late change to an occupancy agreement triggers the redesign of a facility that is already half built. The resulting rework consumes programme, budget and a considerable quantity of material that should never have been bought. We are, in effect, building prototypes at scale and then discovering their defects during commissioning, at the point in the programme where there is least room to absorb them. In any comparable manufacturing sector, that would be understood as a failure of method rather than an unavoidable feature of complex work.
One product, multiple configurations
The answer is neither to standardise the market into a single building, which the three conditions above make impossible, nor to accept that every facility must be designed from scratch. It is to separate the parts of a data centre that genuinely respond to local conditions from the parts that do not.
In our work we treat that as roughly a 25/75 division. Around a quarter of the facility is irreducibly local: ground conditions, civil works, foundations, the connection to the grid, and planning and permitting. The remaining three quarters, covering the structure, the envelope, the internal finishes and the mechanical and electrical systems that make up the working machine, can be designed as a product. That portion can be manufactured off site, tested as a functioning system before it leaves the factory, and assembled on arrival rather than built in place.
The distinction that matters is between the product and its configurations. The product is the part that is designed, tested and resolved once: a defined set of components, interfaces and rules. A configuration is how that resolved solution meets a particular set of site conditions. A remote power-led campus and a constrained urban conversion are different configurations of the same product, not different products. Designing those configurations deliberately, in advance, is a fundamentally different exercise from discovering them one project at a time.
That distinction has a commercial consequence. Programme certainty is now worth more than specification. An operator choosing between two schemes is rarely choosing on technical merit alone. They are choosing on when capacity becomes available and how confident they can be in that date. Manufacturing off site and testing before delivery take work out of the sequence that determines the completion date, and move risk into a factory with repeatable quality rather than leaving it on a site subject to weather, labour availability and sequencing constraints.
Designing for divergence
Both categories will grow, and for different reasons. The remote campus follows the electrical network. The constrained urban facility follows demand, existing building stock and the rules governing where data may sit. Neither is a temporary condition, and neither is a variant of the other. Treating them as one market, with a single specification flexed until it fits, is precisely what produces the bespoke, one-off delivery the sector can no longer afford.
So the next generation of facilities should not all be designed as variations of the same building. Different power and workload conditions warrant different reference architectures, and those in turn warrant different delivery models: what is manufactured off site, what is assembled, what is commissioned and when, and what is genuinely left to local resolution.
That is a decision to take upstream, in defining the product and resolving its configurations against the conditions they will actually meet, rather than one to discover project by project.
The industry’s assumption that data centres are fundamentally the same building, adjusted slightly from one site to the next, is becoming harder to defend. Power availability, workload requirements and location constraints are driving genuinely different facility types. The question now is not whether delivery models should change, but how quickly the sector can adapt.
Coming next in Beyond Bespoke
In part two, ‘Time to Capacity’, Stefanos looks at why a facility is only finished when it is energised, cooled and commissioned, and why commissioning, as the last step, absorbs every delay that comes before it.
Later in the series:
‘What Can Actually Be Productised’ – where standardisation holds, and where the site takes over
‘Retrofit as a Product’ – turning existing buildings into a repeatable category
‘The Cost of Building Prototypes’ – what one-off delivery really costs, and who pays for it
‘Designing for Change’ – building for a technology that will change several times over
Follow Bryden Wood on LinkedIn to catch each part as it is published, or explore our approach to data centre design.
