The Invisible Infrastructure

AI is no longer just a tool. It is becoming a layer of economic and social infrastructure.

Something important has happened to artificial intelligence over the past two years, but I am not sure we have named it correctly.

We still talk about AI as though it were a product.

A tool. An assistant. A chatbot. Something people use to write emails, generate images, summarise documents or produce code.

That description is becoming inadequate.

AI is beginning to look less like a new category of software and more like a new layer of infrastructure.

The distinction matters.

Infrastructure is not defined by whether we can see it. It is defined by dependence.

Electricity became infrastructure when the economy began reorganising itself around the assumption that electricity would always be available. The internet became infrastructure when businesses, governments and individuals began designing their behaviour around permanent connectivity.

Something similar is now happening with artificial intelligence.

Increasing amounts of human activity are being routed through models.

People use them to interpret information, write software, conduct research, prepare legal arguments, analyse financial documents, formulate business strategies, produce marketing material, study unfamiliar subjects and make decisions.

None of this requires some hypothetical future artificial general intelligence.

It is happening with systems that already exist.

And that may be the more important story.

The most consequential technologies rarely arrive as a single dramatic event. They become consequential because thousands of small decisions begin accumulating around them.

A company discovers that a task that once took three hours now takes twenty minutes.

A programmer stops searching documentation manually.

A university student stops beginning research with Google.

A manager asks an AI system to analyse a spreadsheet before reading it herself.

A financial adviser uses a model to test an argument.

A journalist uses one to investigate a subject before making a phone call.

Each decision looks trivial.

Collectively, they represent a change in the architecture of economic life.

The question is no longer simply how many people are using artificial intelligence.

The more interesting question is how much human activity is now passing through it.

We do not yet have a particularly good way of measuring this.

We measure model users. We measure tokens. We measure subscriptions. We measure investment. We measure semiconductor sales and data-centre construction.

But these are indirect measures.

What we really want to know is something closer to AI throughput.

How much reasoning, production, administration and communication is being mediated by machine intelligence?

That number will eventually matter enormously.

During industrialisation, electricity consumption became a rough proxy for economic development because electricity was embedded in productive activity.

In the AI economy, computational inference may become a similar measure.

Not because tokens have some magical economic significance, but because they record the movement of human work through machines.

A billion additional model queries are not merely a billion conversations with a chatbot.

They may represent millions of documents analysed, thousands of software functions written, legal arguments tested, scientific papers interpreted, business decisions examined and administrative processes completed.

Once viewed this way, several apparently separate developments begin to converge.

The extraordinary demand for GPUs.

The construction of enormous data centres.

The sudden importance of electricity generation.

The scramble for transmission infrastructure.

The competition for semiconductor manufacturing.

The debate over water consumption.

The willingness of the largest technology companies in the world to spend hundreds of billions of dollars building computing capacity.

These are usually reported as stories about the technology industry.

They may eventually be understood as something much larger.

The physical construction of an intelligence infrastructure.

This also explains why the economic effects of AI are difficult to observe cleanly.

Technological transitions rarely produce an immediate, uniform increase in productivity.

Before electricity transformed factories, businesses had to redesign factories around electricity.

Simply replacing a steam engine with an electric motor did not produce the full productivity gain. The larger gains came when factories were reorganised around smaller motors, flexible machinery and different production layouts.

The technology changed the architecture of production.

Artificial intelligence may be following the same pattern.

Giving every employee access to a chatbot is the equivalent of installing an electric motor in a steam-era factory.

Useful, certainly.

Transformative, perhaps not.

The larger effects will come when organisations begin rebuilding workflows around the assumption that machine intelligence is continuously available.

That process has barely begun.

And it raises a more interesting social question than whether AI will “take our jobs”.

What happens when access to reasoning itself becomes cheap?

For most of history, competent analysis was expensive.

It required education, experience, labour and time.

Expertise remains valuable. Judgment remains scarce. Knowledge of a field remains indispensable.

But the cost of producing a first approximation of intellectual work has collapsed.

A person can now ask a machine to explain a technical subject, interrogate an argument, write functioning software, examine a contract or construct a financial model almost instantly.

Sometimes the answer will be wrong.

Sometimes dangerously wrong.

But the relevant comparison is not between AI and perfection.

It is between AI and what the person would otherwise have had.

Often that alternative was nothing.

No analyst.

No tutor.

No programmer.

No researcher.

No second opinion.

That is why the social consequences may extend well beyond labour productivity.

Artificial intelligence changes the distribution of cognitive capability.

A small business gains access to analytical capacity that previously belonged to a larger company.

An individual gains access to capabilities once requiring a team.

A competent generalist can move into specialised domains much faster.

And experts themselves can dramatically increase the scope of what they can examine.

The advantage may therefore flow less to people who possess information than to people who know how to interrogate information, test claims and exercise judgment.

This is already creating a strange inversion.

As machine intelligence becomes abundant, human discernment becomes more valuable.

The scarce resource shifts.

Producing an answer becomes cheap.

Knowing whether the answer is good becomes expensive.

That may turn out to be one of the defining characteristics of the AI economy.

We are building machines capable of generating extraordinary amounts of analysis, argument, code, imagery and information.

The bottleneck increasingly moves downstream.

Verification.

Judgment.

Selection.

Responsibility.

In other words, the rise of artificial intelligence does not eliminate the need for human intelligence.

It changes where human intelligence is most valuable.

There is another consequence.

Infrastructure creates dependence long before society formally acknowledges that dependence.

Nobody announces the precise moment a technology becomes essential.

It happens gradually.

Then one day the system fails and everyone discovers how much had quietly come to depend upon it.

That moment will eventually come for artificial intelligence.

A major model outage, computing shortage, cyberattack, energy constraint or regulatory disruption will reveal that activities we still describe as “AI-assisted” have become AI-dependent.

By then the language will probably have changed.

We will stop talking about artificial intelligence as something people occasionally use.

We will talk about compute capacity, inference availability and model access in much the same way we talk about bandwidth, electricity and telecommunications.

That transition may already be underway.

The chatbot was simply the interface through which most people first encountered it.

The larger phenomenon is underneath.

We are beginning to build an economy in which intelligence itself is becoming infrastructure.