One of the most confusing things about the current data centre boom is that two apparently contradictory statements are both true.

Data centres need to be close to their users.

And: Data centres can be built almost anywhere there is enough power and connectivity.

“Data centre” describes a building and infrastructure category, not a single workload. A facility hosting real-time gaming traffic has very different geographic requirements from a facility training a frontier AI model. A bank processing transactions has different requirements from an archive storing backups. Even within AI, training and inference create different location requirements.

Understanding those differences matters increasingly for governments, utilities and communities being asked to allocate hundreds of megawatts of electricity, land, water and infrastructure to proposed AI data centres. The first question should not be, “Where is the data centre? but What is the data centre going to do?

Where Location Matters Most: Latency-Sensitive Computing

Every digital interaction takes time to travel through a network. The farther information has to travel, and the more network infrastructure it passes through, the greater the latency. For many ordinary internet applications, a difference of a few tens of milliseconds is barely noticeable. For others, it can materially affect performance.

That is why cloud companies deliberately put certain infrastructure close to large population centres. AWS Local Zones, for example, extend cloud infrastructure into metropolitan areas specifically so latency-sensitive applications can operate closer to users. AWS says these deployments can support single-digit-millisecond latency for local workloads. Microsoft uses a similar model with Azure Extended Zones, which places computing infrastructure in metropolitan areas and industry centres for low-latency applications. (Amazon Web Services, Inc.)

Some of the workloads where location matters most include:

Online gaming. A delay between a player’s action and the server’s response can directly affect gameplay.

Real-time financial applications. In some trading and transaction environments, tiny latency differences can have economic consequences.

Industrial control systems. Robotics, automated manufacturing and other machine-controlled environments may require extremely rapid communication between devices and computing systems.

Augmented and virtual reality. Sending information to a distant data centre and waiting for the response can create perceptible lag.

Interactive cloud applications. Applications that constantly exchange small amounts of information with users generally benefit from shorter network paths.

Cloud providers have built their infrastructure around this reality. Google says the location of a cloud region can be selected partly on the basis of latency, while Microsoft explicitly publishes round-trip latency measurements between Azure regions to help customers decide where to place workloads. (Google Cloud) For these kinds of applications, a data centre in or near Toronto is not economically equivalent to one in northern Manitoba simply because both have electricity and fibre.

AI Inference: Location Matters, But Not Always in the Same Way

Generative AI complicates the picture.When someone asks a chatbot a question, generates an image or interacts with an AI agent, the model is performing inference: taking an already-trained model and calculating a response.

Latency matters because users are waiting for the result. That is one reason cloud providers are increasingly distributing AI infrastructure geographically. AWS explicitly includes AI and machine-learning inference among the workloads for which its Local Zones can provide low-latency access. (Amazon Web Services, Inc.)

But AI inference is not necessarily as geographically constrained as high-frequency trading or real-time industrial control. If a model takes several seconds to generate a complicated response, an additional 20 or 30 milliseconds of network travel may make little perceptible difference. Conversely, voice AI, autonomous systems and highly interactive agents may be much more sensitive to latency.

AI inference therefore tends to create regional demand for compute, rather than requiring every major model to run inside every city. This also helps explain why hyperscalers operate cloud regions throughout the world. Google, for example, describes its regional infrastructure as serving several functions simultaneously: reducing latency, improving availability and meeting local data-residency requirements. (Google Cloud)

Where User Location Matters Much Less: AI Training

AI training presents an almost opposite geographic problem. Training a large model can involve thousands or tens of thousands of GPUs repeatedly exchanging information with one another for weeks or months. The people who will eventually use the model might be thousands of kilometres away.

The computers doing the training, however, need to be extremely close to each other. That distinction is critical. For large-scale AI training, the important latency is frequently not:

data centre → user but GPU → GPU

NVIDIA describes modern AI infrastructure as requiring extremely high-bandwidth, ultra-low-latency networking between accelerators. Its AI-factory architecture is designed around keeping GPUs efficiently connected and utilized during training and inference. (NVIDIA) NVIDIA notes that spreading synchronous AI training across geographically separate data centres introduces significant inter-region latency and can create performance bottlenecks when models must synchronize their work. (NVIDIA Developer)

So you could train a model in Manitoba, Quebec, Abu Dhabi or Iceland and ultimately serve that model to users in New York, London or Singapore. The training facility does not need to sit beside its eventual customers. It does need enormous amounts of reliable electricity, GPUs, cooling, high-capacity fibre and exceptionally fast internal networking. This changes the economics of location.

The Rise of the Power-Centred Data Centre

Traditional internet infrastructure often followed users. Large-scale AI infrastructure increasingly follows power. A training cluster consuming hundreds of megawatts can be difficult to build in a large metropolitan area where electricity demand is already high, land is expensive and transmission systems are constrained.

Suddenly, locations that might once have appeared remote can become extremely valuable if they have:

• abundant electricity;
• available transmission capacity;
• inexpensive land;
• fibre connectivity;
• political willingness to approve large facilities;
• suitable cooling conditions;
• and enough infrastructure to support enormous computing clusters.

This is why whether a proposed AI data centre is “too far away” from major population centres may not matter. If it is primarily a training facility, distance from people may matter very little. Distance from adequate electricity matters enormously.

Storage and Batch Computing Can Be Even More Flexible

Some data centre workloads are even less location-sensitive. Data backups, archival storage, offline analytics, rendering and other batch workloads can often tolerate delays that would be unacceptable for interactive applications.

If a company launches a computation that takes six hours to complete, an additional 40 milliseconds of network latency is meaningless. That allows companies to optimize these workloads around other factors such as electricity prices, land costs, tax treatment, climate or available capacity.

It also allows cloud architectures to separate workloads geographically. A company might keep customer-facing applications near users while sending large analytics jobs or AI training workloads to a distant compute centre. There is no requirement that everything happen in the same building.

There Is Another Kind of Location Requirement: Jurisdiction

Physical distance is also not the only reason data centre location matters. Sometimes the important issues are sovereignty and law: which country the computers are in. Governments, banks, healthcare organizations and regulated industries may face requirements governing where data is stored or processed.

Cloud companies therefore build regions in specific countries even when a nearby foreign region could technically provide adequate performance.

Google explicitly lists data residency alongside latency and cost as a consideration when selecting cloud regions. Microsoft similarly describes Azure Extended Zones as infrastructure that can satisfy both low-latency and data-residency requirements. (Google Cloud)

That helps explain data centre expansion in places such as the Gulf, Southeast Asia and other markets that may look redundant if data centres are viewed only as machines serving the nearest population.Countries increasingly want domestic cloud and AI infrastructure for economic, regulatory and sovereignty reasons.

In that case, the national border can matter more than the additional network milliseconds.

One Data Centre Can Also Do Several Things

The categories are not absolute. NVIDIA’s concept of an “AI factory,” for example, includes training, fine-tuning and high-volume inference within the broader infrastructure. (NVIDIA)

A hyperscale campus could therefore perform training workloads for a global customer while simultaneously supporting inference, storage or enterprise computing. Operators can also distribute different pieces of the same service across several locations.

A company might train a model in a giant power-rich campus, host the master model there, deploy inference capacity in several regional cloud facilities and then use smaller edge locations for applications requiring extremely low latency. The architecture becomes layered:

Large centralized compute → regional compute → local/edge compute → user.

That is one reason treating “AI data centre” as one homogeneous category is increasingly misleading.

What Communities Should Ask

This distinction becomes especially important when a developer approaches a government or community proposing a very large AI data centre. Before assessing whether the location makes sense, governments should know exactly what is being proposed.

Is it:

• an AI training campus?
• an inference facility?
• a conventional cloud region?
• colocation space?
• storage?
• an enterprise data centre?
• a mixed hyperscale campus?
• or simply a powered site being developed in the hope that a future customer will decide what to do with it?

Those answers determine whether proximity to population centres is an asset, irrelevant or potentially a disadvantage. They also determine the likely economic impact. A giant training facility could consume hundreds of megawatts while employing relatively few people after construction. An edge facility might be much smaller but serve an important local economic function. A sovereign cloud region may have value because domestic companies and governments require Canadian processing capacity even if the same computing could technically be performed elsewhere.

So when a developer says a community is an ideal location for an “AI data centre,” the appropriate response is not simply to ask whether the community has land and electricity, but to ask: Ideal for what?

The New Geography of Computing

For decades, one of the central rules of digital infrastructure was that computing should move closer to users. That rule still applies. But AI has introduced another extremely powerful geographic force.

Computing is also moving toward energy. The result is not one universal map of where data centres should be built but several overlapping maps. Low-latency applications cluster around users. Cloud regions cluster around markets and jurisdictions. Edge infrastructure moves even closer to the applications it serves. AI training clusters gravitate toward enormous pools of electricity and suitable infrastructure. Storage and batch workloads can move toward the cheapest and most efficient locations. And sovereign infrastructure follows national borders.

That is why a giant AI facility in the Middle East, the Canadian Prairies or Atlantic Canada is not inherently badly located simply because it is far from the world’s largest technology customers.

If the facility is training models, geographic distance from those customers may be almost beside the point. But if the same developer claims it is building infrastructure to deliver ultra-low-latency services to distant users, location becomes considerably harder to ignore.

The geography only makes sense once you know what the computers are supposed to be doing.