The Economy Beneath AI

A few days ago, I published The Sovereign Agentic Economy as an open-source map.

It was deliberately unfinished. I invited readers to take any branch, challenge it, extend it, replace it or turn it into an article of its own.

This is my first fork.

I wanted to follow one part of that map further: if artificial intelligence becomes increasingly capable of acting rather than merely answering, where does economic power actually move?

The answer may have less to do with the model than we currently assume.

For the past few years, the AI story has mostly been told as a race. Which company has the smartest model? Who topped the latest benchmark? Who has the most powerful chips, the largest data centers, the biggest financing round, the shortest path to something resembling general intelligence?

All of that matters. But I increasingly suspect that we are paying too much attention to the most visible part of the system.

A chatbot that drafts an email mostly needs to be intelligent. An AI agent that books freight, pays a supplier, files customs paperwork or rearranges a production schedule needs much more. It needs permission to enter systems, a way to prove who it represents, access to current information, a way to move money and hardware secure enough to trust with all of that. Somewhere underneath, it also needs electricity.

Once AI moves from answering questions to taking actions, the model becomes one component inside a much larger economic machine.

And some of the most valuable parts of that machine may turn out to be the least glamorous.

There may be two AI markets, not one

It is common to hear that AI models will eventually become commodities. That may be partly true, but I doubt they will all commoditize in the same way.

At the top of the market sits what we might call frontier intelligence: the best reasoning, planning and cross-domain capability money can buy. Building these systems is extremely expensive. They require enormous amounts of computing power, specialized infrastructure, energy and capital. There are good reasons to think that this part of the market may remain concentrated.

But most of the economy does not need the smartest model in existence.

A warehouse does not need a philosopher. An accounting department does not need a machine capable of writing a novel. A railway system may care more about reliability than brilliance, while a customs authority may value security, speed and traceability more than benchmark scores.

For much of the economy, the relevant question will be simpler:

Is the intelligence good enough for the task?

Good enough to reconcile a ledger, track inventory, check a regulation, identify a maintenance problem or reroute a shipment. These jobs can often be handled by smaller, cheaper and more specialized systems.

The structure may begin to resemble a hospital. Most patients do not begin with the world’s leading specialist. Nurses and general practitioners handle the large volume of ordinary cases, while the expensive specialist is called when something unusual appears.

AI could develop in much the same way. Cheaper systems handle most routine activity; expensive frontier intelligence is used selectively when its additional capability justifies the cost.

If that happens, the economic question changes. It is no longer only who builds the smartest model?

It becomes:

Who can deliver enough intelligence, at the right price, with the right reliability, for the job that actually needs doing?

From navigating software to expressing intention

For roughly thirty years, software has trained us to navigate.

We open an application, find the right menu, enter information, compare options, click buttons and confirm the result.

Agents could reverse that relationship.

Instead of manually operating an airline website, I might simply say:

Find me the best flight to Paris next Thursday. Arrive before 5 p.m., keep it below €400, use my preferred airline if the difference is reasonable, and do not book until I approve it.

The software still exists somewhere underneath. But I am no longer navigating it. The agent is.

That does not mean interfaces disappear. Their role changes.

The important screen may increasingly become the place where I express an intention, review what the agent proposes, authorize important actions and intervene when something goes wrong.

This matters because a large part of the digital economy was built around human attention. Advertising depends on people looking at screens. Referral businesses depend on people clicking links. Many software products are priced according to the number of humans interacting directly with them.

If agents begin doing some of that navigation for us, part of the value may move away from the visible interface and toward the systems underneath it.

Three layers seem especially important.

The device may become less of a screen and more of a vault

One might assume that powerful cloud-based agents make phones and laptops less important.

I am not sure that follows.

An agent acting for me could eventually need access to my location, calendar, messages, contacts, payment details, photographs, health information and identity. The more useful it becomes, the more sensitive the context it needs.

Something has to guard that context.

The device in my pocket may therefore become less important merely as a screen and more important as a trusted physical boundary between me and the systems acting on my behalf.

Secure chips, biometrics, local processing and permission controls may become more valuable precisely because the agent becomes more capable.

The more powerful the agent, the more valuable the gate around the human may become.

This is why I am not convinced that agentic AI necessarily destroys the premium around companies such as Apple. It may change what that premium is paying for.

Less interface.

More trust.

Intelligence still needs reality

An intelligent model can know an extraordinary amount and still be useless if its information is stale.

An agent needs to know whether the shop is open now, what a flight costs now, whether a product is in stock, where the traffic is, whether a law changed yesterday and whether a shipment actually arrived.

This creates what I think of as the reality layer: the maps, search systems, commercial databases, government registries, inventory feeds, transport networks and financial information that connect digital intelligence to the current world.

That layer could become enormously valuable.

A model can be brilliant and still make a terrible decision if it is acting on yesterday’s reality.

Companies controlling reliable, live information may therefore hold a different kind of advantage from companies building intelligence itself.

And someone still owns the infrastructure

There is also a less visible group of companies that supply the data centers, cloud systems, networks and enterprise infrastructure on which AI depends.

Their position is unusual because they can simultaneously be investors in AI companies, suppliers to them, distributors of their products and competitors building their own models.

That makes the relationship more complicated than a simple contest between one AI laboratory and another.

The labs may build the intelligence, but someone still owns the physical and commercial infrastructure through which that intelligence reaches the economy.

Whichever model wins, the landlord may still collect rent.

The shipping-container problem

If agents from different companies and countries are going to work together, they need common ways to communicate.

A purchasing agent may need to query a supplier. A logistics agent may need to confirm a shipment. A banking agent may need to authorize payment. A government system may need to verify a document.

Without common standards, every agentic system becomes its own island.

The analogy I keep returning to is the shipping container.

The container did not require every ship, truck, port and railway to become identical. It standardized the object that moved between them. Once everyone could handle the same box, very different transportation systems could connect to one another.

Agents may need a digital version of the same idea.

A European agent does not have to be built like a Chinese one. A bank does not have to use the same model as a shipping company. They simply need agreed ways to exchange instructions, identities, permissions, proofs and payments.

Technical standards for this are already beginning to emerge. Their names matter less here than the direction: different AI systems are gradually learning how to connect to the same tools, communicate with one another and transact.

Once intelligence becomes distributed among many agents, interoperability itself becomes infrastructure.

Sovereign AI is really a question of control

Almost every government now talks about AI sovereignty, but the phrase can mean very different things.

The simplest version is that national data remains stored on servers physically located inside the country. That may be useful, but it does not necessarily mean the country ultimately controls the technology or the legal environment around it.

A stronger version requires sensitive information—health records, defense data, civil registries—to remain within national borders for both storage and processing.

The strongest version asks a much wider set of questions.

Who controls the chips? Who owns the cloud? Who operates the models? Where does the data go? Who can inspect the system? Where does the electricity come from? And ultimately, whose law governs everything?

Different countries will answer those questions differently because they are protecting different things.

Europe is making one kind of bet: that privacy, auditability, legal predictability and trusted technology will become valuable in their own right.

China is pursuing much greater control across the technology stack.

Energy-rich countries may see an opportunity to convert reliable power into computing capacity.

Smaller countries may care less about building frontier models than about protecting national languages, public data and critical systems.

Sovereignty therefore does not necessarily mean having the world’s smartest model.

It may simply mean having enough control over the system that another country or company cannot switch off something essential.

What happens to countries built around cheap cognitive labor?

This is where the discussion becomes more uncomfortable.

Countries such as India and the Philippines built large service industries around a powerful economic advantage: educated workers could perform valuable cognitive tasks remotely at lower cost than workers in richer economies.

AI attacks exactly that category of work.

Routine software development, customer support, administrative processing, document review, translation and basic analysis are among the areas where cheaper machine intelligence could increasingly compete.

That does not mean these industries disappear. They may move upward.

Human workers may become less valuable for producing every routine output and more valuable for supervising machines, handling exceptions, checking compliance, managing relationships and applying domain knowledge where the agent cannot be trusted alone.

The value proposition could shift from “our people are cheaper” toward “our people know how to run, check and govern these systems.”

But there is an obvious question.

Will the new work appear at the same scale as the old work disappears?

We do not know.

And that is a much larger development question than whether a particular AI company wins the model race.

Thinking gets cheaper. The physical world does not.

This may be the most important economic tension in the whole story.

AI could produce two opposite price movements at the same time.

On one side is cognitive deflation. Routine drafting, translation, administrative work, standard contracts, basic analysis, customer service and some forms of software development become cheaper as capable models spread.

That is excellent news for anyone buying those services. It is less comfortable for companies and workers whose income depends on selling them.

Underneath that falling price of cognition, however, something very different is happening.

AI is becoming heavy industry.

Data centers need enormous amounts of electricity, but they also need chips, land, cooling systems, transformers, fiber connections and construction. None of those things expand at software speed.

A new application can reach millions of users in weeks. A power plant, substation or transmission line may take years.

That mismatch is economically important.

Even if AI systems become more efficient, the amount of computing being deployed may continue to increase demand for power and infrastructure.

So the same technological revolution may make certain kinds of thinking cheaper while making the physical systems required to produce that thinking more valuable.

The revolution appears on a screen, but the investment happens in concrete, copper, silicon and electricity.

That is why the next stage of AI may increasingly resemble heavy industry rather than the software industry we are accustomed to imagining.

Energy-rich regions could gain strategic importance. Utilities, grid operators, chip manufacturers, cooling companies and owners of land with access to reliable power may capture value from AI without ever building an AI model themselves.

Some of the biggest winners from artificial intelligence may not have “AI” anywhere in their names.

Intelligence per watt may matter as much as intelligence itself

Once electricity and computing capacity become constraints, the industry has an incentive to ask a different question.

Not simply: how intelligent is the system?

But: how much useful intelligence can it produce for the energy, hardware and money required to run it?

That creates pressure toward smaller models, more specialized systems, more efficient chips and more processing on local devices.

Again, this reinforces the idea that frontier performance is not the only race.

A country or company unable to afford enormous computing clusters may still build highly useful systems if it can deliver enough intelligence at much lower cost.

The best system on a benchmark may not always be the best system in an economy.

Who is responsible when an agent gets it wrong?

The legal question becomes much more serious once AI can act.

Suppose my agent orders the wrong material, violates a contract, moves money it was not allowed to move or causes a shipment to miss a legal requirement.

Who is responsible?

“The bot did it” cannot become a legal escape hatch.

The simplest principle may be that the agent can act, but someone must remain accountable for giving it that authority. That someone might be a person, a company, a bank or a government institution.

An agent could have a digital identity showing who it represents, what systems it is allowed to access and how much money it is authorized to spend. Important actions could leave records that allow investigators or commercial partners to reconstruct what happened.

This does not mean every error ends up in court.

Many disputes are mechanical. Was the payment made? Was the document delivered? Did the agent exceed its spending limit? Those questions can often be settled automatically.

More complicated cases would still require human judgment.

AI may allow commerce to move faster.

It does not eliminate ambiguity, responsibility or law.

The international problem may turn out to be surprisingly boring

Much of the global AI debate currently focuses on large questions: what values should AI respect, how dangerous systems should be controlled, how privacy should be protected and how governments should respond to the technology.

Those questions remain important.

But an agentic economy creates another category of problem that is much less dramatic and perhaps just as important:

How does an autonomous system legally conduct business across borders?

Imagine a European purchasing agent buying from a Chinese manufacturing agent, using a Moroccan logistics company and financing the transaction through a Gulf bank.

The systems may use different models, different clouds, different currencies, different laws and different rules about data.

Yet the transaction still has to work.

That requires identity, payments, insurance, audit trails, customs rules, liability and ways to resolve disputes.

In other words, the international architecture of AI may eventually involve a great deal of boring plumbing.

But economies are built on boring plumbing.

Shipping containers, accounting standards, bank-clearing systems, commercial law and customs documentation rarely attract the excitement of a breakthrough technology. Without them, however, global trade does not function.

AI agents may eventually need their own equivalent.

The real divide may be between what can be copied and what cannot

This leads to the question underneath the whole article.

If intelligence itself becomes increasingly abundant, what remains scarce?

Models improve quickly. Software features can be copied. Open systems can spread. A breakthrough developed by one company can often inspire similar systems elsewhere.

But a power plant cannot be copied overnight.

Neither can a semiconductor factory, a trusted identity system, a national payment network, years of enterprise relationships, a legal jurisdiction, a live map of the physical world or a secure device sitting in hundreds of millions of pockets.

Some of these assets are slow to build precisely because they exist partly outside software.

That may be where durable economic power moves as intelligence becomes easier to obtain.

The economy beneath AI

The early AI story was easy to understand: build the smartest model and capture the value.

The next stage looks messier.

Some value will remain with frontier models. Intelligence still matters, and superior reasoning can be extremely valuable.

But once intelligence starts acting in the world, the surrounding system matters much more.

The agent needs electricity to run, chips to compute, trusted hardware to receive permission, live information to know what is happening, access to payment systems to transact, common standards to deal with other agents, and a legal framework that determines who remains responsible for what it does.

That gives us a very different economic map.

AI may push down the price of some kinds of cognitive work while increasing the value of infrastructure. It may weaken some software gatekeepers while strengthening hardware, identity and data gatekeepers. It may make intelligence more global while making governments more concerned about technological sovereignty. And it may make autonomous systems more capable while making human accountability more important rather than less.

So perhaps the scarce thing in the next phase of AI will not simply be intelligence.

It may be trusted access to the physical and institutional world.

Energy, identity, permission, current information, infrastructure and responsibility do not improve at the speed of a model release.

And the smartest model in the world is not very useful if it cannot get electricity, prove who it works for, access the world it is supposed to act in or pay for anything.

This is a fork

This essay began as one branch of The Sovereign Agentic Economy, the open-source map I published earlier.

The point of that experiment was to invite readers to take one part of the map and push it further.

This is what my own first attempt produced.

It also changed some of my thinking along the way.

I started with a stronger suspicion that models themselves would rapidly lose most of their economic premium. I now think the picture is more complicated. Frontier intelligence may remain scarce and valuable while routine intelligence becomes much cheaper. At the same time, the surrounding layers—energy, trust, access, identity and interoperability—may capture more value than I initially expected.

That is exactly why I wanted the original map to remain open.

A branch is useful only if following it can change the map.

So consider this one contribution, not the answer.

If another part of the original map interests you more—labor, energy, sovereignty, payments, law, developing economies, hardware—take that branch instead.

And if this fork is wrong, fork it again.

A few technical notes for readers who want them

Several developments sit underneath this argument, but they do not need to be understood to follow the essay.

Standards such as MCP and A2A are emerging to help AI systems connect to tools and communicate with other agents. New payment approaches are being developed so software can authorize and settle transactions without relying entirely on payment processes designed for humans. Governments are experimenting with different versions of sovereign AI, ranging from local data storage to much broader control over cloud infrastructure, models and computing capacity.

At the same time, data-center electricity demand is rising rapidly, while many of the physical systems that support it—generation, transmission, transformers, cooling and advanced chips—expand much more slowly. If agents are increasingly allowed to take consequential actions, identity, permissions, spending limits and audit trails are also likely to become part of the ordinary infrastructure around them.

The technical details will keep changing.

The economic question underneath them is more durable:

If intelligence becomes cheaper and more abundant, where does the real moat move?

Karim Spring