The significance of a data centre extends well beyond the hardware it houses. Yet industry discussion often focuses on GPUs, racks, models and compute, as if the facility was merely a protective shell for whatever technology happens to be fashionable. The servers may be the reason a building was financed, but the facility manager inherits something much slower and more expensive: electrical service, switchgear, cooling plant, water systems, structural capacity, fire protection, security, acoustics, loading access and a site tied into the public grid.
That distinction matters in Ontario because the load is growing. The IESO’s 2026 Annual Planning Outlook expects data centres to account for 8.6 per cent of Ontario electricity demand by 2050. At the same time, ASHRAE’s current AI data-centre framework describes legacy rooms built around roughly 5–10 kW server racks as confronting AI deployments that can exceed 100 kW per rack and require liquid cooling, higher-capacity distribution and very different operating behaviour.
A reasonable conclusion is that no building with a 30-, 40-, or 50-year physical lifespan can be expected to operate within a single, stable computing regime. The design challenge, therefore, extends beyond efficiency.
The load will change before the building does
Facility teams are accustomed to planning for growth. Data centres now need a second planning axis: mutation. The future load may be larger, smaller, denser, more distributed, more liquid-cooled, more inference-heavy, more edge-oriented or simply owned by somebody with a different operating model.
A facility that can only be efficient at one narrow design point is not necessarily a durable facility. A highly optimized mechanical system can become a constraint if it cannot be staged down. Electrical capacity can become stranded if distribution cannot be sectionalized or reassigned. Floor space that looks generous under one rack geometry can become awkward when heavier liquid-cooled systems change pipe routes, service clearances and loading patterns.
That suggests a practical principle for data centre owners: design the long-life layers of the building to tolerate several plausible computing futures, and let the short-life layers be replaced without requiring a new building every time the hardware thesis changes.
Make the electrical backbone divisible
The temptation in a high-demand project is to think of electrical capacity as one giant number: megawatts secured, megawatts delivered, megawatts available. The useful questions are more granular. Can major electrical blocks be isolated without taking the whole site with them? Can a future tenant operate only part of the plant without dragging idle transformers, UPS systems or cooling equipment through an inefficient low-load regime? Are spare conduits, busway paths and switchgear positions documented? Can equipment be removed through a real access path, or has the building been assembled around components that can only leave by demolition?
AI also changes the shape of power demand, not merely the total. ASHRAE’s retrofit guidance warns that synchronous AI workloads can create large step loads and stresses that legacy electrical systems may need substantially different distribution, buffering and harmonic-management strategies. Designing for that future does not mean installing every conceivable technology on day one. It means protecting the spatial, electrical and controls architecture required to add or substitute it later.
Pouring one cooling future into concrete
Cooling is where a building can become obsolete while looking brand new. Air management remains fundamental, but high-density compute is accelerating direct-to-chip liquid cooling and hybrid systems. The U.S. Department of Energy’s 2024 data-centre design guide treats air, liquid cooling, electrical systems and heat recovery as one interacting system; ASHRAE’s current AI framework explicitly recommends hybrid strategies when modernizing facilities that still have useful air-cooling assets.
For a new building, the goal is not to predict the right cooling architecture, but to avoid making an irreversible choice. That can mean reserving routes and service zones for liquid distribution, using plant arrangements that can be staged, providing space for coolant distribution units where the business case may later justify them, designing around useful water-temperature ranges instead of the coldest temperature a current vendor happens to request, and keeping controls, metering and commissioning data granular enough that the next operator can understand what the plant is actually capable of doing.
The facility manager should be able to answer a deceptively simple question five years from now: if the rack changes, what has to change with it? The best answer is not ‘the building.’
Water is a site condition, not a footnote
Water strategy deserves the same treatment. Lawrence Berkeley National Laboratory’s recent work shows how dramatically workload-level water use can vary depending on server efficiency, grid characteristics, utilization, cooling technology, climate and infrastructure. There is no single water solution that is automatically green.
For facility managers, that points to metering and optionality rather than branding. Know the direct water demand. Know the indirect dependence created by the cooling choice. Know what happens during drought restrictions, maintenance, poor water quality or a change in process temperatures. If evaporative cooling is part of the first design, the owner should still understand what conversion to drier or more closed-loop operation would be required later.
A building that can only meet its thermal load by consuming water, under assumptions that may not survive the building’s life, has inherited a hidden lease on local conditions.
Heat reuse needs a customer, not a press release
Waste-heat recovery is another place where a good idea can become decorative infrastructure. Nearly all the electricity entering IT equipment eventually becomes heat, but reusable heat is not valuable merely because it exists.
The U.S. Department of Energy’s design guidance outlines the conditions: a useful heat host should be nearby; the temperature must match the host’s needs, preferably without an expensive lift; and the ownership, incentive and operating arrangements have to survive real life. Newer international guidance makes the same economic point from the other direction. Remote data centres and long heat-transport distances can erase the value quickly.
Design for heat recovery where the site offers a real path and leave connections, plant space and a temperature strategy that can support it later. But do not justify equipment today on the assumption of a future greenhouse, district loop or neighbouring building that no one is contractually prepared to use. Heat reuse becomes infrastructure when there is an off-taker. Before that, it is an option—and options should be designed cheaply enough to remain options.
Commission the exit path
Data-centre commissioning usually asks whether the facility can survive failure while it is operating. The same discipline should be applied to change.
Before first occupancy, the owner should have an asset and state map that a future operator can actually use: one-line diagrams that match reality; equipment capacities and efficient operating ranges; spare and abandoned pathways; isolation boundaries; controls points; water and refrigerant inventories; structural loading information; fire and security zones; major removal routes; and a record of what was deliberately left adaptable.
A question that is usually postponed until a lease ends or a technology program dies is how does this place come apart? Which systems can be mothballed independently? What can be sold, reused or reassigned, remediated? Which utility assets still have value if the server hall changes use? What would a partial conversion look like rather than an all-or-nothing shutdown?
The asset is the capability to change
Ontario is planning for a much more electricity-intensive economy, and data centres are now large enough to matter to provincial system planning. That makes every major facility a long-term infrastructure decision, even if its first customer thinks in three-year hardware cycles.
The building does not need to know what replaces today’s GPU. It needs enough electrical divisibility, thermal flexibility, structural capacity, metering, access and documentation.
The greatest risk is building a facility that works perfectly for one computing generation yet becomes a stranded shell when that generation passes. The industry could focus more on designing the facility’s second life from the outset, before the first server arrives.
Joshua W.J. Brown is a Toronto-based writer, filmmaker and independent systems researcher from Sarnia-Lambton whose work examines infrastructure, technology and the institutions built around them.
