A data centre can take years to move from an early concept to an operational facility.

AI hardware does not wait for the building.

By the time a project moves through planning, engineering, procurement, construction, and commissioning, the technology selected at the beginning may already have a successor. Furthermore, the original equipment may still work perfectly well, but the assumptions surrounding it may no longer represent the best way to deploy AI.

That creates a problem that goes deeper than equipment upgrades.

A building becomes difficult to change once major design decisions have become embedded in its structure and infrastructure. AI, meanwhile, is moving through hardware generations, workloads, and architectures at a much faster pace.

The industry has started to recognize this mismatch. A 2026 Microsoft Research paper argues that traditional data-centre lifecycle management is struggling to keep pace with AI’s rapidly changing models, resource requirements and hardware profiles.

This raises an uncomfortable possibility for data centre developers and engineers:

A facility could be brand new and already be making yesterday’s assumptions.

The project timeline and the AI timeline are moving at different speeds

The difficulty begins with the basic economics of construction.

A large data centre cannot be designed from scratch every time a new accelerator arrives. Developers need enough certainty to secure land, plan the facility, obtain approvals, appoint design teams, procure long-lead equipment, and mobilize contractors.

Those decisions create momentum.

AI infrastructure operates differently.

Accelerator platforms continue to evolve. Workloads change. Operators may alter the balance between training and inference. Moreover, new approaches to computing can also change the requirements placed on servers, networking, and the supporting infrastructure.

ASHRAE’s 2026 AI Data Center Energy Performance Framework describes the situation directly: Furthermore, AI facilities are being deployed at record speed while compute, power, cooling and supporting technologies continue to develop. In some cases, construction is already underway while technical information from equipment manufacturers is still changing.

That creates a fundamental timing problem.

The building needs decisions before the technology has finished making them.

The most important decisions may be the ones that are hardest to reverse

Not every part of a data centre has the same degree of permanence.

A server can be replaced. Software can be updated. A network component can be upgraded.

Changing the basic physical arrangement of a facility is a different matter.

Consider a project that chooses a particular configuration because it is highly efficient for its expected workload. So, the decision may reduce unnecessary space, simplify infrastructure, and improve the economics of the initial deployment.

That optimization can be perfectly rational.

The problem appears several years later if the next generation of equipment needs a materially different arrangement.

A design that was highly optimized for one configuration may have fewer options for another.

For that reason, flexibility needs to be considered at the level of the original engineering decisions. It is not enough to leave some empty space and call the facility future-ready.

Design teams need to understand which assumptions are likely to change and which physical decisions will make those changes expensive.

That requires closer coordination between the people defining the IT environment and the people designing the building around it.

“Design freeze” becomes more complicated when the technology is still moving

There is nothing unusual about freezing a design before construction.

Without that point, costs become difficult to control, and schedules become almost impossible to manage. Contractors cannot build efficiently if the underlying requirements keep moving.

AI does not eliminate the need for design discipline.

It changes what happens around it.

ASHRAE’s commissioning guidance notes that traditional project delivery models are straining because new technical information can emerge while teams are already designing and constructing AI data centres. The organization also recommends tighter feedback between design, construction, commissioning and operations.

As a result, this suggests a different way of looking at design freeze.

The question is not whether a project should freeze its design.

It is which decisions should be frozen and which should retain a path for change.

For example, a project may need to lock major structural and civil decisions early. It may have more flexibility in how certain equipment zones, service routes, or future deployment areas are configured.

That distinction can reduce the impact of technology changes without forcing the entire project back to the drawing board.

The result is a more deliberate separation between fixed infrastructure and changeable infrastructure.

That could become one of the defining principles of AI data centre design.

An AI-ready building is not necessarily an AI-adaptable building

The phrase “AI-ready” sounds definitive.

In practice, it can describe very different levels of preparation.

A facility may have engineers design it to accommodate a particular high-density AI deployment. It may have the required infrastructure and successfully support that workload from day one.

That makes it AI-ready.

It does not automatically tell an operator how difficult the next transition will be.

AI-ready

A facility can support the intended AI hardware, workload, and operating conditions for which its designers planned it.

AI-adaptable

A facility can accommodate significant changes in hardware, configuration or workload without requiring disproportionate reconstruction, extended downtime or a complete redesign.

That second capability is becoming more relevant because AI facilities are likely to go through several technology transitions during their operating lives.

ASHRAE’s framework now covers not only planning and design but also commissioning, operations, retrofit, and modernization. That lifecycle approach reflects a simple reality: the facility’s useful life extends far beyond the technology installed when it first opens.

A building therefore needs to be judged at more than one point in time.

Opening day is one test.

The first major technology refresh is another.

The first hardware replacement may reveal what the original design got wrong

A new data centre is easiest to judge when everything is new.

The equipment matches the specification. The infrastructure has undergone testing. The commissioning process has validated the systems. The operator can measure performance against the original design objectives.

The real test comes later.

Suppose the operator wants to replace a large portion of its compute platform. The new equipment may have different physical characteristics, different thermal behaviour or different infrastructure requirements.

The question then becomes practical.

• How much of the original facility can continue to be used?
• Can the new equipment enter the facility without disrupting neighbouring systems?
• Or, can the required changes happen in a defined zone, or does the work spread across the facility?
• Can the operator complete the upgrade while other workloads continue operating?
• How much engineering work does the team need before the new hardware can actually enter service?
  

These questions have financial consequences.

A facility that requires extensive physical intervention for every major technology transition can accumulate costs that were invisible in its original construction budget.

Microsoft Research’s 2026 lifecycle study makes this broader point. It considers building decisions alongside IT provisioning, hardware refresh strategies and operational changes, rather than treating each phase as a separate optimization problem. Its proposed framework reported a 40% TCO reduction against the traditional approach in its evaluation.

The lesson is important for new projects.

The cost of a data centre is not limited to the cost of getting it operational.

The economics of flexibility are easy to underestimate

Flexibility often looks expensive when viewed from the initial construction budget.

Additional space, alternative infrastructure arrangements, more accessible service routes, or provisions for future modifications can appear to be unnecessary costs when the project is trying to deliver a specific capacity as efficiently as possible.

The calculation changes when the facility begins to evolve.

Imagine two data centres commissioned at roughly the same capacity.

	Facility A: Optimized for today	Facility B: Designed for change
Initial construction	Lower	Higher
Original deployment	Highly optimized	Highly optimized
Hardware replacement	Potentially disruptive	Designed around easier transitions
Reconfiguration	May require significant engineering work	More manageable within defined zones
Downtime risk	Higher during major changes	Lower if changes can be isolated
Stranded infrastructure	Greater risk if assumptions change	Lower where systems can be reused
Long-term value	Depends heavily on original design assumptions	More resilient to technology changes

Facility A may win the initial capital comparison.

Facility B may win over multiple technology cycles.

That does not mean every project should spend more on flexibility. Some flexibility will never generate enough value to justify its cost.

The important change is to put that decision into the financial model.

Instead of asking only what a feature costs to build, developers can ask what future intervention it could avoid.

That is particularly relevant for AI infrastructure because the replacement cycle of compute equipment can be much shorter than the physical life of the facility.

The next generation of design will need to accommodate uncertainty

Trying to predict exactly what AI hardware will look like years from now is unlikely to produce reliable design decisions.

The better approach is to identify the parts of the facility most exposed to change.

Those might include the physical arrangement of compute environments, infrastructure interfaces, equipment access, service distribution, expansion routes and areas where major systems may eventually need replacement.

The goal is not to create infinite capacity.

That would be economically wasteful.

Instead, designers can create options.

An expansion route is an option.

A service arrangement that allows equipment to be replaced without dismantling surrounding infrastructure is an option.

A layout that can accommodate more than one deployment configuration is an option.

A system that can be upgraded in stages rather than replaced as a whole is an option.

These options have value because they reduce the consequences of uncertainty.

ASHRAE’s integrated design guidance now explicitly recommends adaptive planning and says designers should revisit assumptions as AI technologies and densities evolve. It also emphasizes treating architectural, electrical, and mechanical systems as interconnected rather than designing them independently.

In practice, that is a significant shift from designing a facility around one fixed picture of the future.

Construction speed makes the problem even more important

There is another complication.

The industry is under pressure to build AI infrastructure faster.

Recent reporting has highlighted construction methods aimed at dramatically shortening traditional data-centre delivery schedules as developers compete to bring computing capacity online.

Speed has obvious commercial value.

However, compressing a project can also reduce the time available to absorb new information.

ASHRAE’s commissioning guidance describes this tension. AI projects are moving quickly, while the technologies being deployed are themselves changing. Additionally, the organization recommends more continuous interaction between design, construction, and commissioning, along with smaller and more granular handover milestones where appropriate.

That changes the role of commissioning.

It is no longer simply the final stage where teams check a completed building against its design.

Instead, commissioning can become part of the feedback loop that helps teams understand whether rapidly evolving technologies are integrating as intended.

The same principle applies across the project.

The faster the delivery cycle becomes, the more important it is to make decisions that can absorb change without creating a chain reaction throughout the project.

The real measure of future-readiness may be the cost of changing the facility

The data-centre industry has plenty of established performance measures.

Capacity matters.

Efficiency matters.

Reliability matters.

Construction speed matters.

But AI introduces another useful measure:

How difficult is it to change the facility?

That question can be broken down into practical tests.

• How much of the infrastructure can teams modify without major reconstruction?
• Can equipment be replaced in isolated areas?
• How much downtime would a major technology transition create?
• Which systems have been optimized so tightly that changing them would be expensive?
• How much infrastructure could become stranded after a hardware transition?
• Can the facility support different deployment configurations over its life?
• How quickly can lessons from one deployment influence the next phase of construction?
  
  

The final question is particularly relevant for large campuses.

ASHRAE notes that many AI data-centre developments now use repeated or modular concepts across campuses and portfolios. That creates an opportunity to feed lessons from commissioning and operations back into subsequent phases rather than repeating the same assumptions.

In that environment, adaptability is not just a feature of one building.

It can become a property of the entire development strategy.

AI may change what “good design” means

For decades, data centre design has been about making complex infrastructure reliable, efficient and predictable.

AI adds a different requirement.

The facility has to remain useful while the technology inside it continues to change.

That does not mean every data centre needs to become infinitely flexible. It means project teams need to distinguish between decisions that create genuine long-term value and decisions that simply optimize today’s deployment.

The strongest projects may therefore be the ones that understand where to commit and where to preserve options.

That requires architects, engineers, developers, contractors, equipment suppliers, and operators to think about the same facility across different time horizons.

The architect may be thinking about a building that lasts decades.

Meanwhile, the operator may be thinking about its next hardware refresh.

The contractor is thinking about what can be built on schedule.

At the same time, the IT team may already be looking at the generation after the equipment currently being installed.

Those perspectives need to meet much earlier in the process.

The building may outlast several generations of AI

A data centre is a long-lived physical asset.

The AI systems inside it are not.

That mismatch is unlikely to disappear. If anything, continued advances in AI hardware and infrastructure will make it more important.

The industry’s challenge is therefore moving beyond designing facilities that can handle today’s requirements.

It is about understanding how those facilities behave when today’s requirements stop being the right ones.

A data centre that performs exceptionally well on opening day has achieved an important milestone.

A data centre that can absorb the next major technology shift without becoming a reconstruction project has achieved something more difficult.

That may ultimately be the more valuable definition of AI-ready.

Because the AI system installed on day one may not be the system the facility needs five years later.

The building will still be there.

Where the next generation of data centre design is being discussed

These challenges are becoming increasingly important for the people responsible for taking AI infrastructure from concept to operational reality.

The 4th Data Centre Design, Engineering & Construction Summit UK brings together professionals across the data centre lifecycle to examine the design, engineering, construction and delivery decisions shaping the next generation of facilities.

4th Data Centre Design, Engineering & Construction Summit UK

6–7 October 2026 | London, UK

The summit is made for professionals involved in the planning, design, engineering, construction & delivery of data centre infrastructure. It provides an opportunity to examine how rapidly changing AI requirements are affecting the physical environments built to support them, while exploring practical approaches to delivering resilient and adaptable facilities.

As AI changes the requirements placed on data centre infrastructure, the conversations taking place during the design stage will have consequences long after construction is complete. Join now!