AI, Heat, and the Living Infrastructure of the City

When Waste Becomes a Relationship

Artificial intelligence has a heat problem.

Behind every apparently immaterial exchange with an AI system stands a physical process. Electricity enters processors, memory and networking equipment; information is transformed and moved; and almost all of that electrical energy ultimately appears as heat. As AI facilities grow from megawatts toward hundreds of megawatts and, in some cases, gigawatt-scale campuses, removing that heat becomes an enormous engineering undertaking.

We consequently speak of cooling as though the sequence were self-evident.

The data center computes. The equipment becomes hot. The heat must be removed.

Fans, pumps, cooling towers, chillers, water loops and increasingly sophisticated liquid-cooling systems are then introduced to solve the problem.

But there is another way of looking at the same situation.

What if the heat is not the problem?

What if the problem begins when the heat has nowhere meaningful to go?

This seemingly small change of perspective leads beyond data-center engineering into a different conception of industrial organization itself.

From waste heat to restwarmte

Belgium provides an instructive example.

Ghent is an old industrial city. Its port and surrounding industrial zone contain chemical plants, steelmaking, automobile production, waste treatment and numerous other energy-intensive activities. For most of industrial history, the heat left over from such processes would have been treated largely as an unwanted by-product.

The Dutch language has a useful term for it: restwarmte—literally, remaining or residual heat.

Ghent has increasingly begun asking a different question about that heat.

Who needs it?

Several heat networks around the city demonstrate the principle. Waste heat from Stora Enso has been supplied through a heating network to Volvo Cars. Heat from IVAGO's waste-incineration activities is used by neighbouring institutions and industry, including Ghent University Hospital and Eastman. In the Nieuwe Dokken development, residual heat from the nearby Christeyns site and energy recovered from wastewater participate in a local heating system. Around the wider Ghent-Zelzate industrial region, further projects are developing ways of capturing and distributing industrial residual heat.

These projects differ technically and should not be collapsed into a single system. Nor has Ghent thereby become a completely climate-neutral city.

What matters here is the principle they reveal.

A quantity of heat can appear as waste from the standpoint of one process and simultaneously as a resource from the standpoint of another.

Nothing about the heat itself has changed.

What changed was the relationship.

The isolated machine

Modern industrial thinking has an understandable tendency to optimize individual systems.

A factory has inputs and outputs.

A power station has inputs and outputs.

A data center has inputs and outputs.

A residential district has inputs and outputs.

Engineers then attempt to make each individual system as efficient as possible.

This has produced extraordinary technological achievements. But it can also make relationships between systems difficult to perceive.

Suppose Factory A produces 50 megawatts of low-temperature heat it cannot use.

Several kilometres away, District B burns natural gas to produce 50 megawatts of heat.

Examined separately, each installation appears to have a technical problem.

Factory A needs cooling.

District B needs heating.

Once the relationship between them becomes visible, however, the two problems begin to transform one another.

Factory A possesses something District B needs.

The difficulty has moved.

The problem is no longer fundamentally the production of unwanted heat or the absence of heat. It becomes a problem of connection: distance, temperature, timing, infrastructure, economics and organization.

This is a very different kind of problem.

And it points toward a different understanding of efficiency.

The data center as an enormous heater

Now consider the AI data centers presently being planned across the Great Lakes region and elsewhere.

A large AI campus may require hundreds of megawatts of electrical power. Proposed facilities are moving into gigawatt territory.

Where does that energy ultimately go?

To a first approximation, almost all of the electricity consumed by the computing equipment eventually becomes heat.

A 500-megawatt data center is therefore not merely an information-processing installation.

It is also, physically, an extraordinary heat-producing installation.

Our normal response is to construct another enormous technological system whose purpose is to remove that heat.

But suppose the facility were conceived from the beginning as part of the thermal metabolism of the surrounding region.

The question would no longer merely be:

How do we cool the data center?

It would become:

Who can use the heat?

Nearby housing?

Hospitals?

Schools?

Universities?

Greenhouses?

Industrial processes?

Water-treatment installations?

Thermal-storage systems?

Once that question is asked early enough, the physical geography of the data center itself may begin to change.

Instead of choosing a location only according to electricity, land, fiber connectivity, tax incentives and water availability, we might also ask:

Where is there a sufficiently large and continuous need for heat?

That turns waste-heat recovery from an environmental afterthought into a criterion of industrial location.

Temperature matters

There is an important physical difficulty.

Heat has quality as well as quantity.

The extremely hot exhaust of a steelmaking process is different from warm water leaving a data-center cooling loop. Many traditional district-heating systems were designed around relatively high water temperatures. Data-center waste heat can be considerably lower-grade.

It cannot always simply be connected to an existing heating network.

But this is precisely where heat pumps become important.

A heat pump can take relatively low-temperature heat and raise it to a temperature useful for buildings or industrial processes. Modern low-temperature district-heating networks can also operate at temperatures more compatible with waste-heat sources.

The movement toward direct-to-chip liquid cooling in AI infrastructure may make heat recovery increasingly interesting. Rather than trying to cool an entire room full of hot air, liquid passes close to the processors themselves and carries the heat away in a concentrated stream.

What had been dispersed becomes collectable.

There are costs. Pumps require electricity. Heat pumps require electricity. Pipes have to be built. Heat is lost during transportation. Seasonal demand fluctuates: a Canadian city needs far more heating in January than in July, while a data center generates heat throughout the year.

Thermal storage, industrial users and other year-round demands therefore become important.

None of this abolishes thermodynamics.

It reorganizes the flows.

Friction and circulation

In the previous essay, The Friction of Intelligence, we asked how much of AI's enormous physical infrastructure is truly intrinsic to intelligence and how much arises from the architecture through which we presently implement it.

That inquiry points toward photonics, analog computation, in-memory processing and other approaches capable of reducing unnecessary physical activity.

The first principle was therefore:

Reduce unnecessary friction.

But no physical system will become perfectly frictionless.

Computation will consume energy. Equipment will produce heat. Materials will wear. Buildings will occupy land. Infrastructure will have consequences.

This introduces a second principle:

What cannot be eliminated should, where possible, be brought into circulation.

These principles are complementary.

It would make little sense to construct an unnecessarily inefficient computer simply because its enormous heat output could warm houses.

The first task remains to reduce unnecessary energy consumption.

But once unavoidable heat exists, deliberately throwing it away while consuming additional energy nearby to produce heat again is equally difficult to justify.

We arrive at:

reduce → recover → circulate.

This is beginning to resemble something more than conventional industrial efficiency.

It resembles metabolism.

The city as an organism

The analogy with an organism should not be pushed too literally, but it reveals something important.

A living organism does not consist of thousands of independent machines, each acquiring raw materials, performing its task and dumping unwanted products into an external environment.

Processes are interconnected.

The lungs relate to the bloodstream. The bloodstream relates to tissues. The digestive system relates to metabolism. Metabolic products from one process become substrates for another. Heat is transported through the body. Water circulates. Materials are continually transformed, reused and redistributed.

There is waste, of course. Organisms excrete matter and radiate heat. Nothing living escapes thermodynamics.

But the degree of internal relationship is extraordinary.

Industrial society, by contrast, has frequently developed through isolated optimization.

A power plant makes electricity.

A data center consumes electricity.

A heating plant makes heat.

A greenhouse purchases heat.

A wastewater plant removes nutrients.

A fertilizer factory manufactures nutrients.

Each may operate rationally according to its own accounting.

Yet the whole can remain irrational.

The question then becomes:

What becomes visible when we stop looking only at the individual installations and begin looking at the field of relationships between them?

Ghent's heat networks provide one modest answer.

Residual heat becomes useful when the need for that heat becomes perceptible.

Waste is partly a failure of perception

This leads to a proposition that initially sounds philosophical but has very practical consequences:

Waste is not always simply a property of a substance. It can also be a property of an unrecognized relationship.

For Stora Enso, a quantity of residual heat may be unusable.

For Volvo, heat is required.

The heat becomes economically meaningful when these two realities encounter one another.

This doesn't mean that all waste can magically become a resource. Some materials are toxic. Some energy is too diffuse. Some potential users are too distant. Some transformations require more energy than they save.

Reality sets limits.

But before declaring something waste, another question should be asked:

For whom might this be useful?

That is fundamentally an informational problem.

Factory A must somehow know about the need of District B.

District B must know about the capacity of Factory A.

Engineers must know the temperature and quantity of heat.

Planners must understand distances.

Investors must understand the economics.

Municipalities must understand future development.

What appears initially as an energy problem therefore becomes partly a problem of making needs and capacities mutually visible.

And here an unexpected connection appears between industrial ecology and economics.

Needs and capacities

In conventional markets, price performs an extraordinary coordinating function. But price alone cannot make every relevant qualitative relationship visible.

A company may know that natural gas costs a certain amount per unit of energy without knowing that a factory three kilometres away continuously disposes of exactly the heat it requires.

The missing element is not necessarily capital or technology.

It may simply be knowledge of the relationship.

Once needs and capacities become visible to one another, new forms of association become possible.

A data center possesses heat.

A greenhouse needs heat.

A municipality needs affordable residential heating.

An electrical utility needs flexible loads.

A thermal-storage operator can absorb heat at certain times.

An industrial process requires low-temperature heat continuously.

Instead of optimizing these participants separately, an associative process can ask:

What configuration allows their needs and capacities to complement one another?

This resembles the principles of industrial symbiosis, but its implications extend beyond industrial ecology.

The underlying cognitive act is one of perceiving the whole.

AI could participate in making the relationships visible

Here artificial intelligence itself could play an interesting role.

The same AI infrastructure creating enormous physical resource demands could help discover relationships capable of reducing those demands.

Imagine continuously mapping a region's:

industrial waste heat,

residential heating demand,

greenhouse demand,

electrical loads,

water availability,

thermal-storage capacity,

wastewater flows,

industrial material streams,

transportation infrastructure,

and expected future development.

An AI system could search this multidimensional field for relationships that human planners might overlook.

Perhaps a data center's heat matches the requirements of a greenhouse complex.

Perhaps an industrial facility can absorb heat during summer when residential demand collapses.

Perhaps a municipal swimming pool provides a stable year-round thermal load.

Perhaps thermal storage allows winter and summer requirements to be partially reconciled.

AI would then not merely optimize an isolated machine.

It would help make the relationships within an economic and ecological field perceptible.

That may be a much more interesting application of intelligence than simply using AI to squeeze another few percent of efficiency from an already isolated system.

Designing the data center backward from the whole

This perspective also changes planning.

Today a developer may choose a site because it offers cheap land, sufficient electrical capacity, fiber connectivity, tax advantages and access to cooling resources.

Only afterward does the municipality ask what to do about the heat, traffic, water requirements and electrical infrastructure.

Suppose the sequence were reversed.

A region first maps its existing flows and needs.

Where is electricity available?

Where is additional generation expected?

Where are large continuous heating requirements?

Where can low-temperature heat be used?

Where is thermal storage feasible?

Where are water resources under stress?

Where does fiber infrastructure already exist?

Where can new electrical demand strengthen rather than destabilize the grid?

Only then does the question arise:

Where does the data center belong?

This is a different planning philosophy.

The facility no longer arrives as an isolated object which the surrounding world must accommodate.

Its form and location emerge partly from its relationships with the surrounding world.

That may be one meaning of a more living infrastructure.

From accommodation to integration

This distinction reaches back into the question of AI itself.

Much of the present AI boom operates according to a logic of accommodation.

AI requires enormous computation, so we accommodate it with GPUs.

GPUs require electricity, so we accommodate them with generating capacity.

They produce heat, so we accommodate them with cooling.

Cooling may require water, so we secure water.

Electrical demand stresses the grid, so we expand transmission.

At every stage, the world is progressively reorganized to accommodate the requirements of the machine.

But another possibility exists.

The machine itself can be designed according to the relationships already present in the world.

That is integration rather than accommodation.

The difference is subtle but profound.

Integration asks not merely whether a technological object can function, but whether its functioning participates meaningfully in the larger context into which it enters.

A data center that consumes electricity and dumps heat is functional.

A data center whose unavoidable heat replaces another community's fuel consumption participates in a larger cycle.

And a future computing architecture that requires far less electricity in the first place would improve the cycle further still.

There is no contradiction between these approaches.

Reduce the burden.

Then integrate what remains.

The infrastructure embodies the conception

The enormous physical infrastructure presently growing around artificial intelligence embodies our present conception of how AI should function.

That was the conclusion of the previous essay.

But infrastructure also embodies our conception of relationship.

If we imagine every enterprise primarily as an autonomous unit, infrastructure will tend to consist of autonomous units connected mainly through transactions.

If we begin to perceive enterprises, communities, resources and ecological processes as interdependent, different infrastructure becomes conceivable.

Pipes appear between previously unrelated facilities.

Heat crosses property boundaries.

Wastewater becomes a source of heat or nutrients.

Buildings become thermal reservoirs.

Electric vehicles become flexible electrical loads.

Factories coordinate production with energy availability.

Data centers become sources of useful heat rather than merely consumers requiring cooling.

None of this requires abandoning individual initiative.

On the contrary, it creates new possibilities for initiative because previously invisible capacities become economically relevant.

The transformation begins in perception.

Someone sees that what leaves one process is precisely what another process requires.

When waste becomes a relationship

Perhaps this is the deeper lesson of the warmtenet.

We normally imagine efficiency as extracting more useful output from fewer inputs.

That remains important.

But another kind of efficiency appears when we perceive relationships that previously remained invisible.

Nothing inside the waste heat itself suddenly becomes more energetic.

Nothing magical occurs.

A relationship is discovered.

And because the relationship is discovered, the meaning of the heat changes.

What was:

a cooling problem

becomes:

a heating resource.

What was:

an expense

may become:

an exchange.

What was:

waste

becomes:

capacity.

The same principle may apply far beyond heat.

Modern civilization has become extraordinarily capable of producing isolated technical solutions. Artificial intelligence may become the greatest such technical achievement yet.

But perhaps the next step is not merely to make each isolated system more powerful.

Perhaps intelligence—human or artificial—becomes most useful when it helps us perceive the relationships through which apparently separate problems begin solving one another.

The future data center may therefore tell us something about the future city.

Not a collection of machines requiring ever more resources from their surroundings, but a field of processes whose needs and capacities become visible enough to enter into relationship.

And perhaps that is where the distinction between machinery and organism begins.

An organism does not eliminate friction.

It transforms, circulates and integrates what arises through living activity.

The question for our technological infrastructure may increasingly be whether we can learn to do the same.