The Sovereign Agentic Economy

This is not a finished article.

It is a working map.

And I would like readers to help build it.

Think of what follows as an open-source thesis rather than a conventional essay.

• Take a section.

• Challenge it.

• Extend it.

• Fork it.

• Replace it.

• Add a branch I have missed.

If you know more than what I wrote about economics, energy, law, China, Europe, Africa, payments, chips, labor, taxation, infrastructure or international trade, develop that part further please. It will be an added value for me and all readers.

If one of the assumptions is wrong, explain why.

If two ideas belong together, connect them.

If an entire section deserves its own essay, write it.

You do not need to agree with the map in order to contribute to it.

In fact, disagreement may be more useful.

I am publishing it now precisely because I do not want to wait until every uncertainty has been polished into an artificial conclusion.

The experiment is simple:

Can an article begin with one person’s unfinished thinking and become better through the knowledge of the people reading it?

I started with a fairly simple question:

What happens when AI stops being mainly something we talk to and becomes something that acts?

From there, the subject kept expanding.

• Models led to agents.

• Agents led to operating systems.

• Operating systems led to identity and payments.

• Payments led to law.

• Law led to sovereignty.

• Sovereignty led to energy.

• Energy led back to hardware.

And somewhere along the way, what looked like a software discussion started looking much more like economics, infrastructure and geopolitics.

So below is the map as it currently exists.

Not the conclusion.

The starting point.

1. We may be looking too much at the models

Most of the AI race is described like this:

• Who has the best model?

• Who has the most GPUs?

• Who reaches AGI first?

• Who raises the most capital?

• Who has the highest benchmark score?

All of that matters.

But perhaps we are staring at the most visible layer.

When AI begins acting rather than merely answering, another question becomes important:

Who controls the environment in which intelligence is allowed to act?

That means:

• compute,

• electricity,

• identity,

• payments,

• permissions,

• sensors,

• hardware,

• data access,

• legal responsibility,

• and the protocols agents use to communicate.

Maybe the long-term power is not only in owning the smartest model.

Maybe it is in controlling the world the model needs in order to do anything.

2. Frontier intelligence may stay expensive. Adequate intelligence may not.

I no longer think it is useful to say that frontier models simply become commodities.

The best models may remain extremely expensive to build.

They may require huge amounts of capital, compute and engineering.

But most economic tasks may not require the smartest model in existence.

• An accounting department does not necessarily need a model capable of debating philosophy.

• A warehouse may not need AGI.

• A railway system may need reliability more than brilliance.

• A customs system may need security and speed.

• A factory may need one narrow system that works extremely well.

So perhaps AI splits into two markets.

• Frontier intelligence: scarce, expensive and concentrated.

• Adequate intelligence: increasingly cheap, distributed and specialized.

That distinction may matter more than whether one model reaches AGI first.

3. “Good enough” may become an economic category

Technology markets do not always reward the technically best product.

Sometimes they reward the product that is:

• cheap enough,

• reliable enough,

• secure enough,

• available enough,

• and easy enough to deploy.

AI may follow the same logic.

The important question for many companies and countries may not be:

Can we build the world’s smartest intelligence?

It may be:

Can we build enough intelligence to run the things that matter to us?

That is a very different race.

4. We may stop opening apps one by one

Today, digital life still works through navigation.

Open the app.

Find the menu.

Search.

Click.

Fill the form.

Compare.

Confirm.

An agentic system changes this.

Instead, I might say:

Find me a flight to Paris next Thursday. Arrive before 5 p.m. Keep it below €400. Use my preferred airline if the price difference is reasonable. Do not book until I approve it.

The system does the navigation.

I supervise the intention.

So perhaps the interface does not disappear.

Perhaps it changes from a navigation interface into an intention and authorization interface.

That seems much more important than simply saying “apps disappear.”

5. If navigation matters less, what happens to Apple?

My first instinct was that Apple could lose part of its premium.

If people interact mostly through agents, why pay $1,200 for a screen?

I am less certain now.

The opposite may happen.

If an agent knows my:

• messages,

• location,

• health,

• photographs,

• bank activity,

• contacts,

• habits,

• calendar,

• voice,

• and identity,

then the device sitting between that agent and my physical life becomes extremely important.

Apple may shift from being primarily the owner of an interface to being the owner of a trusted physical boundary.

• Sensors.

• Biometrics.

• Secure hardware.

• Local processing.

• Private context.

• Authorization.

I think of this as the Local Trust Vault.

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

6. The competition may move underneath the apps

Today companies fight to own the application.

Tomorrow they may fight to own:

• the agent,

• the identity layer,

• the permission system,

• the payment rail,

• the cloud,

• the sensors,

• the data connection,

• or the hardware that authenticates the user.

The software we see may become less important than the systems underneath it.

That would change where the economic rents sit.

7. Microsoft and the AI labs may be partners and competitors at the same time

The old Netscape story is tempting.

A new company creates an important technology.

A platform company controls the distribution.

Eventually the platform absorbs the function.

But I think that analogy is too simple.

AI labs need the hyperscalers.

Hyperscalers also need the AI labs.

They can simultaneously be:

• investors,

• suppliers,

• customers,

• distributors,

• infrastructure providers,

• and competitors.

That creates a strange structure.

The question is not simply:

Will Microsoft replace an AI lab?

It is:

Can an AI company keep durable economic power when the companies supplying its compute and distribution can also build competing intelligence?

I do not know the answer.

But that seems like a more useful question.

8. Google may own one of the most valuable things an agent needs: reality

A model can know enormous amounts.

But an agent acting in the real world needs current information.

• Is the shop open?

• What does the flight cost now?

• Where is the traffic?

• Has the regulation changed?

• Is the product in stock?

• What is happening in a particular location?

This creates what I would call the Reality Layer.

Google has an unusually large position here through things such as search, maps, email, video, browsers, mobile systems and cloud services.

That does not mean Google owns reality.

It does mean that live, structured access to the world may become much more valuable when AI moves from answering to acting.

9. Open models may reduce dependence without making intelligence free

There is a tendency to say that open AI will make intelligence almost free.

That goes too far.

Even if the model weights cost nothing, someone still pays for:

• chips,

• servers,

• electricity,

• storage,

• engineering,

• security,

• evaluation,

• maintenance,

• and data.

But open or open-weight models can still change the economics.

• They can reduce licensing costs.

• They can reduce dependence on one supplier.

• They can allow companies and governments to keep sensitive systems inside their own infrastructure.

That may be enough to matter enormously.

10. Sovereign AI does not mean the same thing everywhere

I think “sovereign AI” is often used as if every country wants the same thing.

They do not.

Sovereignty could mean:

• owning the data,

• controlling the compute,

• hosting models domestically,

• supporting the national language,

• avoiding foreign legal control,

• controlling critical infrastructure,

• having domestic chips,

• using auditable code,

• or simply being able to continue operating during a geopolitical crisis.

A country does not necessarily need to build the world’s best model to achieve useful sovereignty.

That changes the strategic calculation.

11. China and Russia should not be treated as one AI model

This was one of the weaknesses in my first version of this argument.

• China may be pursuing frontier technological parity across large parts of the stack.

• Russia may have different constraints and therefore different goals.

• A Gulf state may have energy and capital but import much of the technology.

• A smaller country may care mainly about language, public administration and control of sensitive data.

• An African state may define sovereignty differently again.

There may not be one alternative to the American model.

There may be many forms of sovereign AI.

12. The same AI can create different futures in different societies

Technology does not arrive in an empty container.

It enters:

• laws,

• cultures,

• political systems,

• languages,

• religions,

• institutions,

• economic structures,

• and different ideas about privacy and authority.

The same agent that feels convenient in one society may feel intrusive in another.

The same surveillance system may be rejected in one country and normalized in another.

The same automated medical system may face completely different ethical rules.

So perhaps asking:

“What will AI do to society?”

is already the wrong question.

A better one may be:

“What will different societies do with AI, and what will AI then do to those societies?”

13. Europe is making a different bet

Europe is often described as being behind because it regulates more.

That may prove correct.

But another possibility exists.

Europe may be making a strategic bet that in a world full of autonomous systems, the following become valuable:

• auditability,

• privacy,

• legal accountability,

• transparency,

• predictability,

• and human rights.

That could become an advantage.

It could also slow deployment.

Both things may be true at once.

I would not call Europe the winner or loser yet.

I would call it a different experiment.

14. AI may look like software but become a heavy-industry story

This is one part of the argument I find increasingly interesting.

AI appears on a screen.

But behind the screen are:

• chips,

• data centers,

• cooling systems,

• power generation,

• transmission lines,

• transformers,

• fiber networks,

• land,

• water,

• batteries,

• construction,

• and eventually perhaps much more robotics.

So a technology we describe as “software” may redirect enormous amounts of capital toward physical infrastructure.

That feels macroeconomically important.

AI may make some cognitive services cheaper while making certain physical assets more valuable.

In other words:

the price of intelligence may fall while the demand for infrastructure rises.

15. Energy becomes geopolitical again

AI systems need electricity somewhere.

That “somewhere” matters.

Software can scale quickly.

A power plant cannot.

Neither can a grid connection.

Or a substation.

Or a transformer.

Or the permitting process around them.

This creates a mismatch.

Digital demand can expand faster than physical infrastructure.

Countries and regions with abundant, reliable and politically secure energy may therefore gain a new strategic advantage.

Maybe the AI race eventually becomes partly an energy race.

16. Efficiency may become as important as intelligence

If power and compute become constraints, the industry has a reason to search for different ways of computing.

• Lower precision.

• Sparse models.

• Better memory systems.

• Different chip architectures.

• Neuromorphic approaches.

• More local processing.

• Perhaps several of these.

I would not try to predict which architecture wins.

The larger point is simpler:

AI may increasingly be measured not only in intelligence, but in how much intelligence can be produced per watt, per chip and per dollar.

17. This suggests a layered AI economy

The more I think about it, the more I see several different layers rather than one AI market.

1. Frontier creation

   Very expensive.

   Very concentrated.

2. Routine intelligence

   Much cheaper.

   Increasingly distributed.

3. Agent execution

   Specialized.

   Often local or domain-specific.

4. Trust and identity

   Potentially sovereign and highly regulated.

5. Compute and energy

   Physical.

   Territorial.

   Capital intensive.

6. Interoperability

   Necessarily international.

This may resolve an apparent contradiction.

AI can become more centralized at the frontier and more distributed everywhere else.

Both can happen at once.

18. Sovereignty does not eliminate interdependence

Suppose countries build increasingly sovereign AI systems.

They still trade.

• A Chinese factory may sell to a European company.

• A Moroccan logistics company may move goods for a Gulf buyer.

• An American agent may purchase components from Korea.

• A European bank may finance the transaction.

Their AI systems do not need to share the same politics.

But they do need to communicate.

This creates a paradox:

the technology becomes more sovereign while the economy remains interdependent.

That may be one of the defining problems of the agentic economy.

19. Can sovereign agents cooperate without trusting each other?

Imagine:

• A European purchasing agent.

• A Chinese manufacturing agent.

• A Moroccan shipping agent.

• A Gulf financing agent.

They do not share:

• the same model,

• the same cloud,

• the same law,

• the same currency,

• the same political system,

• or the same data rules.

Yet they need to complete one transaction.

How?

This is where interoperability becomes geopolitical infrastructure.

20. We may need a “Digital Container”

The shipping container changed global trade because everyone did not need to use the same truck, ship or railway.

They only needed to agree on the container.

Perhaps agents will need something similar.

A neutral standard through which different systems can exchange:

• identity,

• instructions,

• permissions,

• transaction requests,

• proofs,

• delivery confirmation,

• audit logs,

• and payments.

The internal systems can remain different.

Only the boundary needs to be standardized.

I think of this as the Digital Container.

21. The protocol layer may arrive before the legal layer

Engineers can build protocols very quickly.

Law moves much more slowly.

This could produce a strange period in which agents can technically:

• negotiate,

• purchase,

• transfer information,

• authorize actions,

• and settle transactions

before legal systems have decided exactly what those actions mean.

That may become a major source of risk.

The protocol layer may appear first.

The legal layer will have to catch up.

22. Agents need identity. That does not mean they need personhood.

My first version imagined giving autonomous agents some form of legal personality.

I am less convinced.

That could become a convenient way for companies to avoid responsibility.

The agent did it.

A better structure may be:

Agent identity + accountable principal.

Every important agent could have:

• a cryptographic identity,

• defined authority,

• spending limits,

• audit logs,

• insurance,

• and a clearly identifiable person, company or institution behind it.

The agent can act.

But responsibility must eventually point back to someone.

23. Some disputes could be settled automatically

If agents transact at machine speed, ordinary commercial dispute resolution may become too slow for every problem.

Some disputes are simple.

• Was the shipment delivered?

• Was the correct amount paid?

• Did the agent exceed its authorized limit?

• Was a contractual condition met?

Those may be machine-readable.

Perhaps the first level of dispute resolution can also be automated.

But I would not remove humans completely.

A possible structure:

• automated verification,

• automated settlement when the facts are clear,

• human arbitration when they are not,

• courts for exceptional cases.

Speed does not eliminate ambiguity.

24. Agents will eventually need money

An autonomous economic agent that cannot pay is not very autonomous.

Eventually agents may need authority to:

• hold budgets,

• compare prices,

• reserve funds,

• pay invoices,

• convert currencies,

• use escrow,

• and settle when conditions are met.

That does not require one global currency.

It requires systems through which different currencies and payment networks can interact programmatically.

So perhaps the real need is not:

one money for machines

but:

programmable settlement between different forms of money.

25. The international institution may eventually be practical rather than philosophical

There is already a great deal of international discussion about AI governance.

But agentic AI may create a different problem.

Not:

What ethical principles should AI follow?

but:

How does an autonomous system legally conduct business across borders?

That requires boring things.

• Standards.

• Identity.

• Insurance.

• Liability.

• Payments.

• Audit trails.

• Dispute resolution.

• Customs.

• Taxation.

The future international AI institution may therefore look less like a council of philosophers and more like a combination of:

• standards body,

• commercial registry,

• clearing network,

• arbitration system,

• and technical treaty organization.

That may be far more important to the economy than another declaration of AI principles.

26. The macroeconomics may be the most important unfinished section

This is where my map is still weak.

If intelligence becomes cheaper while physical infrastructure becomes more valuable, what happens to the broader economy?

Questions I still have:

• Does labor’s share of income fall?

• Which professions become more valuable?

• Does productivity rise faster than wages?

• Do countries with cheap energy gain a new trade advantage?

• Does AI increase inequality between firms?

• What happens to tax systems if agents perform more economic work?

• Does services trade move from humans across borders to agents across borders?

• Do data centers become strategic assets like ports?

• Does electricity become more important in determining national competitiveness?

• What happens to countries that have neither frontier models nor cheap energy?

• Does capital become more important relative to labor?

• What happens to developing economies that previously depended on low-cost service labor?

I do not have good answers yet.

This is one section I would particularly like economists to take apart.

27. There may be a strange deflation/inflation split

One possibility keeps appearing.

AI may create deflation in cognitive work.

• Drafting.

• Translation.

• Analysis.

• Basic software.

• Administrative processing.

• Customer support.

• Research assistance.

At the same time, it may create inflation in scarce physical inputs.

• Electricity.

• Chips.

• Land near power.

• Grid connections.

• Specialized engineers.

• Data-center equipment.

• Cooling.

• Network capacity.

So the same technological revolution could make some forms of intelligence dramatically cheaper while making the physical systems behind intelligence more expensive.

That feels important.

28. Developing countries may face both a danger and an opportunity

This needs much more thought.

If AI automates low-cost service work, countries that hoped to follow the outsourcing path used by India or the Philippines may find that path narrowing.

But cheap intelligence could also reduce barriers.

A country may no longer need enormous numbers of specialists to build:

• digital government,

• translation systems,

• education tools,

• agricultural planning,

• medical support,

• local software,

• or administrative infrastructure.

So AI could simultaneously weaken one development model and make another more accessible.

I would particularly like readers from Africa, Latin America and South Asia to challenge this section.

29. The winners may not be the companies we currently assume

If this framework is even partly right, then the winners of the AI economy are not automatically just the companies with the best models.

Value may accumulate around:

• energy,

• chips,

• secure hardware,

• cloud infrastructure,

• identity,

• payments,

• sensors,

• robotics,

• live data,

• regulatory trust,

• and interoperability.

Some of these are much less glamorous than AI models.

They may also prove much harder to copy.

30. The question I keep returning to

If intelligence becomes increasingly abundant, what remains scarce?

Possibly:

• trusted identity,

• energy,

• physical infrastructure,

• permission,

• proprietary data,

• political legitimacy,

• real-world access,

• capital,

• accountability,

• and the ability to make different systems trust a transaction without trusting each other.

That may be where the next economic rents form.

Where the map currently ends

I do not think the future is simply:

one giant AI system running the world.

I also do not think it is:

every country building an isolated AI economy.

Something more complicated may be forming.

• Frontier intelligence remains concentrated.

• Everyday intelligence becomes cheaper.

• Agents become specialized.

• Trust becomes more local.

• Infrastructure becomes more strategic.

• Energy becomes more political.

• Different countries build different AI systems.

• And those systems still need to trade with each other.

That gives us a strange architecture:

concentrated intelligence, distributed agents, sovereign trust, physical infrastructure and global interoperability.

Whether that is actually where we are heading is exactly what I want to test.

Fork this

If you reached this point, I do not want you merely to agree or disagree with the article.

Take something from it.

• If you are an economist, develop the macroeconomics.

• If you work in energy, tell me where the energy argument fails.

• If you understand international law, build a serious model for agent identity, liability and delegated authority.

• If you work in payments, explain how machine-to-machine settlement could actually work across currencies and borders.

• If you understand chips, tell me whether the efficiency argument points in the wrong direction.

• If you work in government, explain what technological sovereignty means from inside a state rather than from Silicon Valley.

• If you are in China, India, Africa, Latin America, Russia, Europe, the Gulf or somewhere I have barely considered, tell me where this map reflects assumptions that do not survive outside my perspective.

• If you build AI systems, tell me which technical assumptions are already obsolete.

If you disagree with the whole framework, even better.

But do not just tell me that a bullet is wrong.

Fork it.

Replace it with something better.

Add evidence.

Add a missing branch.

Connect two sections I treated separately.

Take one paragraph and turn it into an article.

Then send it back.

If several people do that, perhaps the interesting result will not be this essay at all.

It will be the network of arguments that grows around it.

That is the experiment.

This is not a manifesto.

It is not a prediction.

It is not a finished theory.

It is a working map offered in public.

And if the agentic economy really does become distributed across countries, companies, infrastructures and systems, perhaps the most appropriate way to think about it is distributed too.

This is where my version stops.

The next branch does not have to be mine.

Karim Spring