Intelligence Has Reached the “Too Cheap to Meter” Point

We’ve been arguing for some time that this is where the economics of AI were heading.

Not because intelligence would stop improving.

Because intelligence would become abundant.

Capability would keep rising while the cost of producing it collapsed.

And now the trajectory is becoming difficult to ignore.

Models are getting:

Better → Smaller → Cheaper → More efficient → More open → More ubiquitous.

We are seeing models approach or match frontier capability at a fraction of the previous inference cost.

Context windows are expanding.

Architectures are becoming radically more efficient.

Compute requirements are falling.

Open-weight models are closing the gap.

And inference costs are collapsing.

This is the trajectory we’ve been documenting.

And it matters far more than simply making AI cheaper.

Because when the cost of intelligence approaches the marginal cost of running the machine that produces it, intelligence stops being the scarce resource.

It becomes infrastructure.

Like bandwidth.

Like storage.

Like electricity.

You don’t build a business around rationing intelligence anymore.

You build around what you can do with an effectively abundant supply of it.

And this is where the moat moves.

For the last few years, the perceived moat in AI has largely been:

Models → Compute → Capital.

The biggest models.

The biggest clusters.

The biggest training runs.

The biggest balance sheets.

That logic made perfect sense while intelligence was scarce and expensive.

But scarcity is moving.

When models become increasingly interchangeable, open, cheap and capable, the model itself becomes less defensible.

When compute becomes more efficient, the amount of capital required to access useful intelligence falls.

And when inference becomes extremely cheap, the economics of selling raw intelligence deteriorate.

The moat moves above the model.

It moves into the layer that compounds intelligence.

Context.
Memory.
Data.
Workflow.
Distribution.
Trust.
Defaults.
Relationships.
Judgement.

And ultimately:

The ability to reliably turn abundant intelligence into a decision.

That’s a very different moat.

The instrument is becoming good enough.

You don’t need infinitely better guitars.

At some point, another 5% improvement in the guitar matters less than what the musician does with it.

AI is beginning to look like this.

The question changes from:

Which model is smartest?

to:

What have you built around intelligence?

What context does it have?

What does it remember?

What does it understand about the user?

What defaults does it have?

What systems can it act through?

What decisions can it make?

What does the user trust it to do?

And most importantly:

How does every interaction make the system better at serving the next one?

That last part is enormous.

Because the real moat isn’t simply having intelligence.

It is compounding its use.

A system that learns a user’s preferences, history, context and behaviour can become dramatically more useful without necessarily becoming dramatically more intelligent.

That creates a different kind of advantage.

Not a model moat.

A relationship moat.

And this is why defaults become so powerful.

If intelligence is abundant, choice becomes abundant too.

There may be thousands of models capable of doing the job.

The user doesn’t want to evaluate thousands of models.

They want the right answer.

The right action.

The right recommendation.

The trusted default.

So the scarce asset becomes confidence about what to choose.

This is why trusted interfaces, brands, platforms, ecosystems and agents become increasingly important.

They compress an enormous amount of available intelligence into a decision.

Abundant intelligence creates abundant choice.

Abundant choice increases the value of trusted selection.

That is the paradox.

The cheaper intelligence becomes, the more valuable judgement can become.

And the implications for the AI capital cycle are enormous.

If intelligence becomes dramatically cheaper to produce, then some of the assumptions underpinning the current AI infrastructure boom have to be revisited.

You have to ask:

How much compute do we actually need?

How much inference capacity?

How much does each workload really cost?

How much pricing power can model providers maintain?

How much capital can be justified by selling intelligence when competitors can provide comparable intelligence for a fraction of the price?

And what happens when open models can run increasingly capable intelligence locally?

The answer isn’t necessarily that data centres become worthless.

Compute will still matter enormously.

But the required quantity of expensive compute per unit of useful intelligence can fall dramatically.

That’s the distinction.

The world can consume vastly more intelligence while requiring less compute per unit of intelligence.

That’s a very different trajectory from simply extrapolating today’s infrastructure demand forward.

And there is another consequence.

When intelligence becomes cheap enough, you can afford to waste it.

This may be one of the biggest changes of all.

You can run multiple agents.

Generate ten approaches instead of one.

Critique your own work.

Simulate alternatives.

Research continuously.

Personalise everything.

Run intelligence in the background.

Give every employee an AI collaborator.

Give every application an intelligence layer.

Give every person an effectively unlimited cognitive toolkit.

The economics stop forcing you to ask:

“Is this worth spending inference on?”

And increasingly become:

“Why wouldn’t we?”

That is what “too cheap to meter” really means.

Not literally zero cost.

Economically negligible cost relative to the value of using it.

And once you cross that threshold, usage explodes.

Which brings us back to the bigger picture.

The AI revolution may not ultimately be about creating a handful of super-intelligent machines.

It may be about making intelligence ambient.

Always available.

Embedded everywhere.

Personalised to the individual.

Cheap enough to use continuously.

And increasingly capable of acting.

At that point, intelligence becomes a utility.

And utilities don’t capture all the value.

The value accrues to the systems built on top of them.

The electricity isn’t the business.

The internet isn’t the business.

The intelligence won’t necessarily be the business either.

The moat moves to whoever can turn abundant intelligence into trusted outcomes at scale.

That is the transition we’ve been mapping.

And if 26 August 2026 is the point at which intelligence visibly crossed the too-cheap-to-meter threshold, then we should probably stop asking:

“Who has the smartest model?”

and start asking the much more important question:

Who owns the layer that makes abundant intelligence useful?

Because that is where the next moat is being built.