What Are the Hyperscale Data Centres Really For?

The computers may become obsolete. The control over land, water and power may not.

Governments are being told that artificial intelligence requires speed.

Data centres must be approved quickly. Transmission systems must be expanded. New generating plants must be built. Farmland must be rezoned. Water must be secured. Environmental assessment must not become an obstacle. Any community or jurisdiction that hesitates risks being left behind.

This argument depends on more than a belief that digital services will continue growing. It depends on a much narrower and less certain assumption:

That enormous, centralized hyperscale data centres will remain the dominant way of delivering artificial intelligence long enough to justify the land, water, power plants and public infrastructure now being committed to them.

That assumption is rarely tested.

There is little reason to believe data centres will disappear entirely. Digital systems will continue to require storage, backups, communications, scientific computing and some centralized processing. But it does not follow that the hyperscale campuses currently being proposed—some demanding as much electricity as cities—will remain necessary at the scale developers now predict.

The distinction is critical.

A society can continue needing data centres while discovering that it built far too many of the wrong kind, in the wrong places, with far more power and water capacity than future technology requires.

Hyperscale is a technological choice, not a permanent law

The present construction rush is often described as though hyperscale computing were the inevitable physical form of artificial intelligence.

It is not.

Hyperscale campuses are one response to the limitations of current chips, models, networks and business arrangements. Those conditions are already changing.

An influential computer-architecture paper, Cloud No Longer a Silver Bullet, Edge to the Rescue, predicted a shift in which cognitive computing would increasingly occur at the network edge rather than almost exclusively in centralized clouds. The authors identified data-centre limitations, network constraints, privacy, latency and safety as forces pushing computation closer to users and devices.

That shift is no longer theoretical.

Google now provides generative-AI functions that run directly on Android devices without making server calls. It specifically promotes on-device processing as faster, more private, capable of operating offline and free of a separate cloud-computing cost for every inference.

Apple has adopted a similar architecture. Its system first attempts to process requests on the user’s device and sends only workloads requiring larger models to its Private Cloud Compute infrastructure. Its published technical work describes an approximately three-billion-parameter model optimized to run locally using techniques including aggressive quantization.

Microsoft describes what it calls a “cloud training, edge inference” model: centralized facilities may train and distribute models, while routine use occurs on local computers and other devices to reduce costs, latency and privacy risks.

These are not fringe companies predicting a distant possibility. They are three of the companies driving the current hyperscale construction race, simultaneously developing systems intended to prevent every AI request from travelling to a hyperscale facility.

That does not eliminate the need for centralized training or cloud services. It does undermine the assumption that every future AI interaction will require a distant, power-intensive data centre.

Routine inference may not remain centralized

The largest uncertainty is not whether advanced models will sometimes require large computing facilities. It is how much of the day-to-day workload will continue to require them.

Training a frontier model can demand enormous processing capacity. But training is episodic. Inference—the repeated use of a trained model by millions of people and machines—can become the much larger cumulative workload. If routine inference shifts onto phones, laptops, vehicles, industrial equipment and local servers, a substantial portion of the demand now being used to justify hyperscale expansion could move away from centralized facilities.

Smaller models may also replace larger ones for many purposes.

A general-purpose model capable of discussing thousands of subjects may be unnecessary for operating farm equipment, monitoring a pipeline, proofreading a paragraph, controlling a warehouse robot or interpreting a medical device. A smaller specialized model can be cheaper, faster and sometimes more accurate within its field.

Microsoft describes small language models as practical alternatives for devices and environments with limited computing capacity. Its Phi-3 technical report demonstrates a compact model designed to run locally on a phone. Research on edge deployment similarly finds that compressed or specialized models can provide sufficient performance for many applications without relying continuously on remote cloud systems.

This creates a plausible future in which hyperscale facilities still exist but occupy a narrower role:

• training unusually large models;
• performing scientific and industrial simulations;
• handling temporary peak workloads;
• providing centralized storage and backups;
• coordinating distributed devices; and
• serving tasks too complex for local systems.

That is very different from a future in which every business process, search, vehicle, appliance and personal assistant constantly draws power from a remote hyperscale campus.

Yet it is the second future that appears to underpin many of today’s electricity projections.

The distributed alternative is already being built

The movement of AI inference away from hyperscale campuses is no longer merely a prediction.

In 2026, energy-technology company SPAN announced XFRA, a distributed data-centre system designed to place enterprise-grade computing nodes at homes and small commercial properties. The liquid-cooled units use Nvidia GPUs and are intended to perform AI inference, cloud gaming and other workloads that do not necessarily require tightly concentrated clusters.

SPAN is working with major U.S. homebuilder PulteGroup to introduce the system in communities under construction. Instead of constructing one enormous facility and then building the generation and transmission needed to serve it, the XFRA model attempts to use electrical capacity already distributed across existing neighbourhoods.

The company says the system can identify unused capacity in residential electrical services and coordinate computing loads around household and grid demand. Homeowners could receive equipment such as smart electrical panels and battery storage, along with discounted electricity or internet, in exchange for hosting a node.

SPAN does not claim that this model will eliminate centralized data centres. It describes distributed computing as a supplement capable of handling inference close to users while large facilities continue performing workloads such as model training.

That limitation is important. It is also the reason the project matters.

The question is not whether every centralized data centre can be replaced by computers attached to houses. It is whether enough routine inference can be distributed to materially reduce the number and scale of hyperscale campuses required.

SPAN claims that a distributed network equivalent to a 100-megawatt centralized data centre could be deployed roughly six times faster and at one-fifth the cost. Those projections have not yet been demonstrated at scale. The model still faces significant questions involving local distribution capacity, equipment security, maintenance, noise, insurance, data governance and homeowner consent.

The planned deployment demonstrates that companies with direct access to advanced AI hardware are already pursuing a fundamentally different physical architecture. Whether that architecture works economically and technically at the scale SPAN predicts remains unproven.

A hyperscale campus requires developers to predict demand years in advance, assemble enormous parcels of land and secure generation, water and transmission before the first server operates. A distributed network can theoretically add computing one neighbourhood or one node at a time as demand appears.

That difference matters enormously.

The centralized model requires governments and communities to accept the risk that projected demand will arrive. The distributed model could allow capacity to follow demand more closely.

If residential and commercial nodes prove viable, they would not make every hyperscale facility unnecessary. They could, however, weaken the case for building speculative, multi-gigawatt campuses years before anyone knows how much centralized inference will actually be required.

The same companies asking governments to treat hyperscale development as an urgent necessity are participating in the development of alternatives that may reduce that necessity. That contradiction should be examined before permanent control over land, water and generation is granted on the strength of temporary technological assumptions.

Efficiency is advancing at extraordinary speed

The International Energy Agency continues to forecast rapid data-centre growth. Its central projection sees global data-centre electricity consumption rising from approximately 485 terawatt-hours in 2025 to roughly 950 terawatt-hours in 2030. Electricity use by AI-focused centres is projected to grow even faster.

Those figures deserve attention. They do not deserve to be treated as certainties.

The IEA also reports that the amount of electricity required for each AI task is declining at a rate it describes as unprecedented in energy history. Its forecast assumes that the growth in AI use will more than offset those efficiency gains.

That may happen. Computing has repeatedly demonstrated a rebound effect: when each calculation becomes cheaper, society performs many more calculations. But the rebound is an assumption about behaviour and markets, not a physical requirement.

Demand may grow quickly enough to absorb every improvement in chip efficiency. It may grow even faster. Or customers may decide that many AI products are not worth their cost. Businesses may deploy smaller models. Consumers may reject constant AI mediation. Regulations may limit some uses. Local processing may replace cloud inference. New architectures may perform the same work with a fraction of today’s electricity.

Any one of these developments could materially change the demand curve. Several occurring together could make current hyperscale forecasts look less like prudent infrastructure planning and more like a speculative construction cycle.

Even the buildings may become obsolete

The risk is not limited to demand. A hyperscale data centre designed around one generation of chips can become unsuitable for the next.

JLL has warned that new data centres may reach obsolescence sooner than expected as rack power densities increase and cooling requirements change. A building designed for conventional air-cooled servers may not readily accommodate high-density liquid-cooled systems. Electrical distribution, floor loading, ceiling height, water systems and backup power can all restrict future conversion.

The industry is therefore facing two different forms of uncertainty at once. The first is that future computing may become more distributed and efficient, reducing the need for some hyperscale capacity. The second is that the hyperscale capacity that remains necessary may require buildings substantially different from those being constructed today. The result could be a facility that is not obsolete because computing disappeared, but because computing changed.

This is particularly important when public officials describe a data centre as a permanent economic asset. The building may stand for decades, but the technological design that gives it value may have a much shorter life.

The possibility of overbuilding is no longer speculative

Warnings about excess capacity are now coming from within the technology and financial sectors.

The Institute for Energy Economics and Financial Analysis reported that Microsoft had moved from warning about insufficient AI infrastructure to acknowledging that data-centre capacity could be overbuilt. Alibaba chair Joe Tsai has warned that companies building facilities speculatively may be creating the beginnings of a bubble.

Energy consultancy E3 has also cautioned utilities against treating every proposed data-centre load as certain. Its research found that relatively few operating facilities consistently used their reported peak capacity, and fewer than half of the centres examined maintained load factors above 80 per cent.

This matters because proposed demand can become infrastructure before it becomes actual consumption.

Utilities may build generation and transmission around developer forecasts. Municipalities may expand water and wastewater systems. Governments may alter land-use rules, provide tax concessions or accelerate approvals. Communities may reserve electrical capacity that is then unavailable to housing, agriculture or other industries.

If the projected facilities are delayed, built at a smaller scale, operated below capacity or never completed, the company may walk away from part of its plan.

The wires, pipes, roads, substations and public debt do not walk away with it.

The companies building hyperscale centres understand these risks

It would be naive to assume that the world’s largest technology companies do not see these changes coming. They employ many of the researchers developing on-device models, specialized chips, edge networks, optical systems, model compression and distributed computing. They know how quickly each generation of hardware improves. They possess internal demand and utilization data that governments and communities cannot see.

They are developing the technologies that may reduce dependence on centralized facilities while simultaneously moving with extraordinary speed to secure the resources needed to build those facilities.

That contradiction does not prove a concealed plan. There are several ordinary explanations.

The companies may believe total demand will grow faster than efficiency improves. They may fear losing the AI race if a competitor secures capacity first. They may be building defensively, preferring excess infrastructure to the risk of insufficient supply. They may expect enormous models to remain commercially dominant.

But another possibility deserves much closer examination:

The hyperscale data centre may be the immediate justification for acquiring assets whose value will survive even if the original computing forecast does not.

The servers may become obsolete.

The land, water rights, generating capacity, grid connections and regulatory permissions may not.

The more durable asset may be the resource position

A modern hyperscale proposal is not merely a warehouse filled with computers.

It can include:

• hundreds or thousands of acres of assembled land;
• industrial zoning;
• water licences or long-term supply agreements;
• wastewater and treatment infrastructure;
• dedicated gas pipelines;
• private or contractually dedicated generating plants;
• substations and high-capacity transmission connections;
• fibre corridors;
• road and rail access;
• tax concessions;
• rights-of-way; and
• agreements with governments and utilities extending for decades.

The computing equipment may be replaced every few years.

The power plant may operate for generations.

The water allocation may become more valuable as scarcity increases.

The industrial land may retain privileged access to transmission, pipelines and public infrastructure long after the original servers have been removed.

A company that miscalculates future AI demand may therefore still emerge holding an extremely valuable collection of strategic assets.

The public may think it approved a data centre. What it may actually have approved is a privately controlled land, water, energy and communications platform whose first tenant happens to be artificial intelligence.

Water rights can outlive the data centre

Water is often presented as an operating input: a quantity withdrawn for cooling, treated and perhaps partly returned. That framing is incomplete.

The enduring question is not simply how much water a facility uses while operating. It is what legal and commercial control the company acquires over the water source.

Alberta’s Water Act transfer system permits the licensed right to divert an allocated volume of water to be transferred permanently or temporarily where transfers have been authorized through an applicable water-management plan or Cabinet order and subject to provincial review. This does not physically transfer water from one parcel to another, but it can transfer the legal right to divert a defined volume under an established priority. Household rights and registrations for traditional agricultural use are not transferable under these provisions and remain attached to the land.

That makes the terms of a data-centre water arrangement crucial.

Does the company receive a temporary right tied specifically to cooling the approved facility?

Does it purchase an existing allocation?

Can that allocation later be transferred to another parcel, affiliate or industrial use?

Can the water right be sold separately if the data centre closes?

Does the right revert to the Crown, municipality, irrigation district or previous use?

Can agricultural land be acquired primarily for its water position and then taken out of production?

Can the company amend the licence after approval to accommodate a different use?

A promise that a data centre will use efficient or closed-loop cooling does not answer any of these questions.

A closed loop can still require initial filling, makeup water, flushing, treatment, discharge and emergency supply. More importantly, reduced consumption does not necessarily eliminate corporate control over the underlying allocation.

The public interest is not protected merely because the company uses less of a resource it continues to control.

Power generation can outlive the computing demand

Alberta promotes scalable and off-grid generation as central advantages in attracting AI data centres. Under Alberta’s 2026 levy rules, electricity not drawn from the broader grid can qualify for a zero-per-cent levy rate, creating a direct incentive for proponents to pair hyperscale computing with their own generation.

The resulting project may include a gas-fired power plant, substation, fuel supply, transmission connection and computing campus marketed as one investment. But the legal approvals and ownership may be separated. That raises several questions.

If the data centre closes, can the generating facility remain in operation?

Can it sell electricity into Alberta’s grid?

Can it be sold independently?

Can ownership of the power plant be transferred to an affiliate or third party without reopening the original public-interest assessment?

Can the computing company retain contractual priority over the electricity while reducing its local operations?

Can the generation asset be expanded after the data centre has supplied the initial justification for approving it?

Would the plant have been approved in the same location, at the same scale and on the same timetable if it had been proposed openly as a standalone thermal power project?

A generating facility may have a useful life of several decades. A transmission corridor may remain longer. The original computing equipment may be commercially obsolete before either has recovered its capital cost.

The approval process must therefore examine the power asset not only as an accessory serving a data centre, but as an enduring industrial facility with possible future uses of its own.

Land control can also be separated from the promise

Land presents a similar problem.

Large technology campuses may require the assembly of numerous parcels, the conversion of agricultural land, new industrial zoning and access to public infrastructure. The economic-development case is usually tied to a particular company, investment figure and computing proposal.

But land can be held through subsidiaries, partnerships, leases, options and layered corporate structures.

The company publicly associated with the project may not be the registered owner. The landowner may not own the servers. The company holding the water licence may differ from the company operating the generating plant. Each entity may be incorporated in Canada while its effective control, financing or parent ownership is foreign.

Alberta restricts some forms of foreign ownership of rural controlled land, but the Foreign Ownership of Land Regulations contain important exemptions relevant to integrated data-centre developments. The rules do not prohibit acquisitions for power plants, pipelines or electric distribution systems; certain industrial, commercial and transportation facilities may also qualify for exemptions, and registered leases of up to 20 years are exempt. Governments assessing a hyperscale project should therefore identify the ultimate beneficial owners, controlling interests, contractual rights and affiliated entities across the complete development rather than stopping at the name appearing on an individual land title.

The central question is therefore not merely whether one parcel technically complies with the Foreign Ownership of Land Regulations. It is whether a foreign-controlled corporate group can obtain effective long-term control over a much larger integrated project by dividing land, generation, water rights and computing operations among multiple entities.

A fragmented legal structure should not produce a fragmented public-interest assessment.

This question is especially timely because Alberta has considered reforms that would exempt certain commercially or industrially zoned land from the regulations. The province currently describes that policy work as paused.

The project must be assessed as a non-fragmented whole

Alberta’s own guidance for AI data-centre developers identifies separate municipal permits, Alberta Utilities Commission power-plant approvals, Alberta Electric System Operator access, Indigenous consultation and project-specific water, pipeline and broadband requirements. Although Alberta describes its Environmental Protection and Enhancement Act system as providing a coordinated “single window” for air, land and water effects, a hyperscale development may still require separate municipal, utility, grid, water, pipeline, land-title and other decisions.

The EPEA approval process does not necessarily examine the combined ownership, economic and intergenerational consequences of the complete development. A municipality may assess rezoning without deciding whether the associated power plant requires a full environmental impact assessment.

A utility regulator may assess generation or grid connection without examining the long-term transferability of water rights.

A water decision may consider a licensed withdrawal without addressing cumulative land conversion, foreign control or the economic consequences of the computing campus failing.

Each regulator can conclude that the component before it is acceptable while no single process assesses the combined environmental, ownership, economic and intergenerational consequences of the complete development.

Alberta’s Environmental Assessment (Mandatory and Exempted Activities) Regulation identifies particular classes and sizes of projects for which environmental assessment is mandatory.

The critical policy question is whether a development divided into computing, generation, water and related components is evaluated against those thresholds as one undertaking or as several separate applications—allowing the environmental and economic significance of the complete development to escape assessment.

An integrated, non-fragmented whole-project assessment should begin with a simple determination:

Would the land acquisition, water supply, generating facilities, pipelines, substations, transmission, wastewater systems and computing buildings proceed in substantially the same form if the other components did not exist?

Where the answer is no, they are one integrated project for assessment purposes, regardless of ownership structure, application sequence or municipal boundaries.

The complete project should be aggregated when determining whether mandatory environmental-assessment thresholds have been reached.

No subsidiary, lease, joint venture, power-purchase agreement, separate application or later construction phase should be permitted to conceal the total scale of the undertaking.

Foreign control must be measured across the complete project

The same integrated approach should apply to ownership.

Government should disclose:

• the ultimate beneficial owner of every project company;
• the citizenship or jurisdiction of control of each parent entity;
• all options, leases, rights of first refusal and purchase agreements;
• lenders holding security over land, water or generation;
• agreements allowing another party to direct the use of an asset;
• cross-default provisions connecting separate companies;
• guarantees provided by foreign parent corporations;
• contractual rights to electricity, water and site access;
• expected ownership following construction; and
• which assets may be sold independently after approval.

No exemption, exclusion, Cabinet decision or classification applying to one component should prevent consideration of the total land interests and effective foreign control associated with the integrated project.

A corporation does not need to hold every title directly to control the development.

Control can arise through long-term leases, financing, supply contracts, exclusive access agreements, operating covenants and the ability to appoint directors or veto changes in use.

The public record should reveal economic control, not merely registered ownership.

Approval must follow the assets beyond their original use

The most important conditions should govern what happens after the original proposal changes. A hyperscale approval should not operate as a permanent gateway through which land, water and power assets can later be detached from the computing use that justified them.

Before approval, governments should require enforceable answers to the following questions.

Water

• Is the water licence or allocation legally tied to the approved data-centre use?
• Does it expire or revert automatically if computing operations cease or remain below a defined level?
• Can it be transferred to another parcel, affiliate, buyer or industrial activity?
• Can it be sold separately from the data-centre land?
• Can the licence be amended without a new whole-project review?
• Who owns the treatment, storage and wastewater infrastructure?
• What happens to unused or reserved allocation if projected buildout never occurs?
• Does agricultural or municipal priority return when the facility closes?
• Are decommissioning, flushing and contaminated-water treatment financially secured?

Power and generation

• Who owns each generating facility, substation, transmission asset and pipeline?
• Is generation legally dedicated to the approved data-centre operation?
• Can the plant sell power to the grid if data-centre demand declines?
• Can it be expanded, repowered or converted to another fuel without reassessing the complete project?
• Can the generation asset be sold independently?
• Does the original data-centre company retain priority access after selling the facility?
• Are ratepayers exposed if the computing load never reaches its forecast level?
• Who pays for stranded transmission or generation?
• Would a material change in computing demand trigger a new public-interest review?

Land

• Who is the ultimate beneficial owner of each parcel?
• What land is owned, leased, optioned or controlled through contractual rights?
• Can the land be sold separately from the data-centre operation?
• Does industrial zoning remain if the approved computing use ends?
• Can the owner substitute another industrial activity without new environmental assessment?
• Do tax concessions, servicing agreements or infrastructure commitments survive a change in ownership or use?
• Must agricultural land be reclaimed or offered back for agricultural use?
• Can a foreign-controlled affiliate acquire the land after the initial approval?
• Are all related parcels aggregated for foreign-ownership and environmental review?

Corporate separation

• Can the data centre, generating plant, water licence and land be divided among separate companies after approval?
• Does a sale of shares avoid conditions that would apply to a sale of the underlying asset?
• Are public-interest obligations binding on successors, lenders and insolvency purchasers?
• Can a parent corporation withdraw while leaving thinly capitalized subsidiaries responsible for remediation?
• Is a new approval required when effective control changes, even if the registered owner does not?
• Can bankruptcy separate profitable water or power assets from environmental liabilities?

These questions should not be buried in confidential agreements.

The answers should be public, enforceable and attached to the assets themselves.

Conditions should survive sale, restructuring and bankruptcy

A corporate promise is useful only while the corporation remains willing and able to honour it. Approval conditions must therefore run with the land, licence, generation facility and transmission rights—not merely with the original applicant.

They should remain binding after:

• a corporate merger;
• the sale of a subsidiary;
• a change in beneficial ownership;
• the assignment of a lease;
• transfer of a water allocation;
• refinancing;
• foreclosure;
• insolvency;
• sale by a receiver; or
• replacement of the computing tenant.

Financial security should be calculated for the full integrated project, including server removal, hazardous materials, cooling systems, pipelines, generation, water infrastructure and restoration of the site.

Governments should also prevent companies from transferring the valuable assets into one entity while leaving cleanup obligations in another.

The water licence should not survive profitably while the remediation liability is abandoned.

The generating plant should not be separated from the security required to close the computing campus.

The land should not retain privileged industrial rights while the public inherits obsolete infrastructure.

Rapid technological change argues for more scrutiny, not less

There is no need to claim that hyperscale computing will disappear. The stronger conclusion is that governments cannot know whether the scale now being proposed will remain necessary long enough to justify permanent changes in resource control.

On-device AI is already operating. Smaller and specialized models are already being deployed. Hybrid cloud-edge systems are already being marketed by the same companies building hyperscale campuses. Energy efficiency per AI task is already improving extraordinarily quickly.

Industry analysts are already warning about technological obsolescence. Corporate leaders are already acknowledging the danger of overcapacity and speculative construction.

These are not hypothetical developments invented by opponents of data centres. They are visible inside the technology industry itself. The rush forward is therefore difficult to explain solely as confidence in one permanent computing architecture.

Perhaps the hyperscalers believe demand will overwhelm every efficiency gain. Perhaps they are making defensive investments because none can risk allowing a competitor to control the available electricity. Perhaps the construction race is simply a bubble produced by abundant capital, optimistic forecasts and fear of being left behind. But it would be naive to assume these companies have failed to notice that computing technology is changing beneath the buildings they are rushing to construct.

Governments should be asking what the companies gain if their own forecasts prove wrong.

If the data-centre demand disappears, do they still control the land?

If processing moves to personal devices, do they still hold the water allocation?

If the servers become obsolete, do they still own or command the generating plant?

If the campus is sold, can the land, power and water be separated and sold for different purposes?

If the original project no longer exists as proposed, do the approvals and privileges remain?

The public is being asked to surrender regulatory time, grid capacity, water security, agricultural land and future flexibility because hyperscale computing is said to be urgently necessary.

Urgency is not evidence of permanence.

Before governments allow temporary technological demand to create enduring private control over essential resources, they must assess the complete project, identify its ultimate owners and bind every approval to the use that justified it.

Otherwise, future generations may discover that the computers were the least important part of the deal.