Intelligence Is in the Wild

Why the AI race may be turning intelligence from the world’s most valuable scarce asset into one of its most abundant commodities

For most of the artificial intelligence boom, the economic assumption was simple.

Intelligence would be expensive.

The best models would require enormous amounts of compute.

Only a handful of companies could afford to build them.

Those companies would control access.

Everyone else would rent intelligence from them.

That assumption justified extraordinary investment in data centres, GPUs, power and infrastructure.

The future appeared to belong to whoever owned the machines capable of producing intelligence.

But something unexpected is happening.

Intelligence is escaping.

Open models are proliferating.

Weights are being released.

Models are becoming smaller.

Inference is becoming cheaper.

Capable models are moving onto local hardware.

Routers can arbitrage between providers.

And personal agents are emerging that may eventually operate continuously on our behalf.

Artificial intelligence is no longer developing exclusively inside a handful of corporate fortresses.

It is in the wild.

The Original Assumption

The economics of frontier AI depended upon scarcity.

Only a few organisations possessed enough capital, compute and research talent to build the largest models.

If they could maintain that advantage, they could potentially maintain extraordinary pricing power.

Build something enormously expensive.

Protect it.

Control distribution.

Charge everyone else for access.

But technological scarcity only produces durable economic value while competitors cannot reproduce or substitute for it.

AI is becoming unusually effective at destroying precisely that condition.

Being First Is Not The Same As Remaining Scarce

Frontier laboratories still matter enormously.

The largest models remain difficult and expensive to build.

The best laboratories may continue discovering capabilities unavailable elsewhere.

But there is an important distinction between:

being first

and

remaining scarce.

A laboratory may spend billions pushing the frontier forward.

Once some of that knowledge escapes into the ecosystem, everyone behind it starts closer to the answer.

Architectural choices.

Training techniques.

Reasoning approaches.

Efficiency improvements.

Post-training methods.

Model outputs.

Weights.

Every release contains reusable information.

The frontier lab bears much of the cost of discovery.

Everyone following gets a better starting point.

Open Weights Change The Economics

There is a profound difference between accessing intelligence through an API and possessing the model itself.

An API says:

You may use our intelligence under our terms at our price.

Open weights say:

Here is the intelligence.

Optimise it.

Quantise it.

Fine-tune it.

Distil from it.

Specialise it.

Run it locally.

Build something else from it.

Open weights therefore do more than distribute intelligence.

They accelerate its diffusion.

Knowledge that once remained inside individual laboratories becomes reusable infrastructure for the wider ecosystem.

And once sufficiently capable intelligence escapes into the world, it becomes extraordinarily difficult to put it back behind the walls.

Then The Models Leave The Data Centre

This may be one of the most consequential developments.

Capable models increasingly run on hardware outside hyperscale data centres.

Consumer GPUs.

PCs.

Macs.

Eventually more phones, cars, robots and edge devices.

And we are now seeing models explicitly designed around local, always-on agentic workloads.

At almost exactly the same moment, Meta is describing a future in which everyone has a personal AI agent that understands their goals and what they care about — with free versions intended to reach billions of people.

Put those ideas together.

Personal agents are coming.

And increasingly, they may not need to live entirely in someone else’s data centre.

From Rented Intelligence To Owned Intelligence

An always-on agent could perform enormous numbers of cognitive tasks.

Observe.

Classify.

Remember.

Route.

Retrieve.

Summarise.

Decide.

Execute.

If every operation requires remote inference, the agent continuously consumes someone else’s infrastructure.

If much of that intelligence runs locally, the economics change.

The hardware has already been purchased.

The model has already been downloaded.

Electricity still costs money.

Hardware depreciates.

Frontier workloads will continue requiring enormous centralised infrastructure.

But much everyday cognition may not.

The intelligence moves:

Cloud → Device

API → Weights

Rented → Owned

Metered → Increasingly fixed-cost

Concentrated → Distributed

The intelligence hasn’t disappeared.

The metered transaction has.

Routing Attacks The Economics Again

The same pressure appears from another direction.

Model routers.

If dozens of capable models exist, an intelligent system doesn’t need the most expensive one for every problem.

It can ask:

What is the cheapest intelligence capable of resolving this task to the required confidence?

Simple task?

Use the small local model.

Harder task?

Use a stronger open model.

Specialist task?

Route to a specialist.

Exceptional problem?

Escalate to the frontier.

The economically important question changes.

Not:

Which model is smartest?

But:

What is the lowest cost at which this intent can be successfully resolved?

Cost Per Successful Resolution

Nobody ultimately wants tokens.

They want outcomes.

An answer.

A booking.

A piece of software.

A decision.

A transaction.

A completed task.

So perhaps the economically meaningful unit isn’t cost per token.

It is:

cost per successful resolution.

And intelligent systems will relentlessly attack that cost.

Use a cheaper model.

Retrieve less information.

Reuse memory.

Improve the prior.

Cache the result.

Route intelligently.

Run locally.

Reuse successful pathways.

Every optimisation reduces the amount of expensive intelligence required next time.

Better intelligence doesn’t merely become cheaper.

It becomes better at avoiding expensive intelligence.

Demand Can Explode While Unit Economics Collapse

This is why the AI infrastructure debate is so interesting.

AI demand can exceed almost everyone’s expectations.

Billions of personal agents could generate extraordinary amounts of activity.

Token consumption could explode.

AI could become embedded in almost every product and process.

Yet none of that automatically tells us how much centralised infrastructure will be required per task or how much revenue will accrue to any particular model provider.

Because:

Volume and value are not the same thing.

A commodity can experience extraordinary demand while its unit price collapses.

A technology can transform the world while some of the capital invested in producing it earns disappointing returns.

AI can win while AI capital loses.

There is no contradiction.

The AI Race May Have Been A Race To Commoditise AI

And here lies the great irony.

The AI race appeared to be a competition to build the world’s most valuable scarce asset.

But competition itself may be producing the opposite outcome.

One laboratory pushes the frontier.

Another reproduces much of the capability.

Someone opens the weights.

Someone distils them.

Someone quantises them.

Someone gets the model running locally.

Someone builds a router that avoids using expensive intelligence unless necessary.

Hardware improves.

Inference gets cheaper.

The knowledge propagates.

And everyone begins the next race closer to the finish line.

Billions of dollars of competitive investment are effectively financing a global process that makes intelligence:

Better → Smaller → Cheaper → Open → Local → Ubiquitous

And this is no longer necessarily an unintended consequence.

Meta is explicitly arguing for widely distributed superintelligence, free access for billions where possible, and an open-source ecosystem capable of preventing excessive centralisation. Zuckerberg also acknowledges just how quickly advantage can diffuse, writing that AI innovations can be copied and absorbed within months.

Perhaps the lasting achievement of this extraordinary investment cycle won’t be that a handful of companies managed to own intelligence.

Perhaps it will be that, through furious competition with one another, they made intelligence extraordinarily difficult for anyone to own.

They thought they were racing to own the future of intelligence.

They may actually have been racing to give it away.

But Abundant Intelligence Creates Another Problem

Suppose everyone really does have a personal AI agent.

It understands you.

Your preferences.

Your objectives.

Your relationships.

Your history.

Your constraints.

That solves an enormous problem.

The agent knows you.

But it doesn’t automatically know what in the outside world deserves its confidence.

Which information is reliable?

Which company should it recommend?

Which supplier can deliver?

Which source is current?

Which transaction should it execute?

The intelligence may be local.

The world it needs to act upon isn’t.

And that creates the next bottleneck.

Trust Becomes The Scarcity

Imagine telling your personal agent:

Organise our leadership offsite for 30 people near London next month.

The agent may already understand the intent.

It knows the budget.

The preferences.

The team.

The calendar.

Previous experiences.

But now it must resolve the outside world.

It could search thousands of possibilities from scratch.

Retrieve enormous quantities of information.

Compare everything.

Verify every claim.

Resolve every contradiction.

Or it can use accumulated confidence.

Known sources.

Proven outcomes.

Trusted relationships.

Coherent entities.

Previous successful resolutions.

In other words:

trust networks.

The personal agent solves one side of the problem.

What does this person actually want?

The trust layer helps solve the other.

What in the world can reliably provide it?

Resolution connects the two.

Agent knows me → Trust network knows the world → Resolution connects them.

When Intelligence Becomes Abundant, Scarcity Moves

This is why abundance doesn’t eliminate value.

It moves it.

If models can be downloaded…

If inference can happen locally…

If routers can substitute between providers…

If yesterday’s frontier becomes tomorrow’s commodity…

Then intelligence itself becomes progressively harder to defend as the scarce layer.

Scarcity moves toward things that cannot simply be generated or downloaded.

Identity.

Context.

Memory.

Relationships.

Reputation.

Verification.

Coherence.

Proven outcomes.

Trust.

Knowing something becomes increasingly easy.

Knowing what to trust and what to do remains difficult.

And interestingly, Meta’s own vision arrives at trust as the practical threshold for personal agents: people will only delegate sensitive information and tasks to agents they believe are acting in their interests.

From The Model Economy To The Resolution Economy

The Model Economy asks:

Who owns the smartest intelligence?

The Resolution Economy asks:

Which intelligence should be used, what should it trust and what action should it take?

That second problem survives commoditisation.

In fact, commoditisation makes it more valuable.

Because once everyone has intelligence, merely possessing intelligence stops being particularly scarce.

Value migrates toward turning abundant intelligence into reliable outcomes.

From intelligence to selection.

From generation to verification.

From capability to confidence.

From search to resolution.

Intelligence Is In The Wild

The largest AI companies will continue building extraordinary models.

The frontier will continue moving.

Centralised compute will remain enormously important.

But something may already have changed permanently.

Models are open.

Weights are downloadable.

Capability is diffusing.

Inference is moving locally.

Routers are arbitraging providers.

Personal agents are emerging.

Every release moves the starting line forward for everyone who follows.

So perhaps the question is no longer:

Who wins the race to control artificial intelligence?

Perhaps it is:

What happens if nobody can control intelligence for very long?

Because billions of people may eventually possess intelligence that understands them.

The harder problem will be helping that intelligence understand what in the outside world deserves its trust.

The great irony of the AI race may be that one of the largest capital investment cycles in the history of technology ends up making intelligence too abundant for anyone to monopolise.

And once that happens, the next race begins.

Not who can produce the most intelligence.

Who can become something that intelligence trusts enough to choose.

The intelligence is escaping.

It is in the wild.