The AI Moat Illusion

"We have no moat. And neither does OpenAI."

When Google's internal memo leaked in 2023, many people dismissed it as pessimism.

Looking back, I think it was one of the most important observations made during the AI revolution.

Not because Google was losing.

But because it recognized something much bigger.

The model itself was already beginning to commoditize.

Three years later, that prediction seems increasingly difficult to ignore.

After attending the Paris Open Source AI Summit and speaking with founders, researchers, infrastructure companies and chip manufacturers, I left with a very different view of where AI is heading.

I no longer believe the long-term battle is about who owns the smartest model.

I think we're watching intelligence itself become infrastructure.

And history has shown us what eventually happens to infrastructure.

Technology has a remarkably consistent pattern.

Electricity was once a competitive advantage.

Today it's an expectation.

Owning servers was once a competitive advantage.

Cloud computing turned it into a utility.

Storage was expensive.

Bandwidth was scarce.

Compute was rare.

Eventually they all became abundant.

The value didn't disappear.

It simply moved somewhere else.

AI feels like the next chapter of the same story.

Today we obsess over benchmark scores.

Which model reasons better?

Which model writes better code?

Which model wins another leaderboard?

Those questions matter today.

I'm not convinced they'll matter nearly as much ten years from now.

The remarkable thing about open source isn't simply that it's catching up.

It's that it has fundamentally changed the economics of progress.

Every new model teaches the next generation.

Researchers build on each other's work.

Optimization techniques spread across the entire ecosystem.

Ideas move faster than companies.

History suggests that once knowledge escapes into the open, it rarely returns behind closed doors.

That doesn't mean proprietary models disappear.

It means maintaining a permanent lead becomes extraordinarily difficult.

The frontier keeps moving.

Everyone runs after it.

Eventually yesterday's frontier becomes tomorrow's baseline.

We've watched this happen repeatedly across technology.

There's little reason to believe AI will be the exception.

One assumption appears almost everywhere in AI today.

Better GPUs create better AI.

That is certainly true today.

But today's bottlenecks are not necessarily tomorrow's bottlenecks.

Software has always adapted faster than hardware.

The first computers needed entire rooms.

Now the phone in your pocket outperforms them.

Not because physics changed.

Because engineering improved.

AI is following exactly the same trajectory.

We're already seeing it.

Quantization.

Distillation.

Mixture-of-Experts.

Sparse architectures.

KV-cache optimization.

Better attention mechanisms.

Specialized inference engines.

Every few months someone discovers another way to achieve similar performance using fewer parameters, less memory and less energy.

That trend is unlikely to stop.

The opposite is far more likely.

The models of 2030 will probably feel absurdly efficient compared to those of today.

Not because hardware alone became dramatically better.

Because we finally learned how to use it properly.

This is why I don't believe the long-term winners are automatically today's hardware companies either.

NVIDIA has executed extraordinarily well.

AMD continues to improve.

Tomorrow another company may do the same.

Or perhaps the next breakthrough comes from China.

Or from Europe.

Or from a startup nobody has heard of yet.

Five years is an eternity in semiconductors.

Intel once looked untouchable.

IBM looked untouchable.

Cisco looked untouchable.

Sun Microsystems looked untouchable.

Technology has a habit of humbling incumbents.

Meanwhile, the software keeps improving.

Models become smaller.

Inference becomes cheaper.

Architectures become more efficient.

Eventually, companies stop asking for the largest GPU they can afford.

They start asking for the cheapest hardware capable of running their workload.

Then another optimization arrives.

Then another.

And another.

Hardware improves.

Software improves even faster.

The cycle repeats.

I suspect the next wave won't simply produce better models.

It will produce models optimized for specific hardware.

Apple Silicon.

AMD.

NVIDIA.

Qualcomm.

RISC-V accelerators.

Custom ASICs.

Entire companies will emerge whose only business is making one family of models dramatically faster on one family of chips.

That is how technology evolves.

General-purpose systems eventually become specialized.

Specialization drives efficiency.

Efficiency drives commoditization.

And commoditization changes where value is created.

Today, many frontier AI companies make money by selling tokens.

That makes perfect sense while frontier intelligence remains scarce.

But scarcity rarely survives technological progress.

As models improve, become cheaper, and increasingly run locally, token pricing will come under pressure.

That doesn't mean inference becomes free.

Electricity isn't free.

Cloud storage isn't free.

Bandwidth isn't free.

Infrastructure still generates enormous businesses.

But infrastructure businesses rarely command monopoly economics forever.

Competition compresses margins.

Customers gain alternatives.

Reliability, integration and service become more important than exclusivity.

The same transition is likely to happen in AI.

Selling raw intelligence will become less differentiated.

The value will migrate toward products, workflows, trust, distribution and integration.

Perhaps the biggest mistake we make is assuming someone has already found a permanent moat.

History suggests otherwise.

Every technological revolution begins with scarcity.

Eventually it creates abundance.

Abundance changes what people value.

The internet made information abundant.

Search became valuable.

Cloud computing made infrastructure abundant.

Software became valuable.

AI may make intelligence abundant.

Something else will become valuable.

I don't know exactly what that "something else" is.

Perhaps it's distribution.

Perhaps it's trust.

Perhaps it's owning the customer relationship.

Perhaps it's something none of us has imagined yet.

What I am increasingly convinced of is this:

The model is not the destination.

It is becoming the foundation.

And foundations rarely capture the greatest share of value forever.

Five years from now, we may look back and laugh that we once believed intelligence itself could be a moat.

History rarely rewards people for owning yesterday's bottleneck.

It rewards those who recognize where value is moving next.