The Next AI Race Will Not Only Be About Bigger Models

The recent Fable 5 drama feels like a turning point for AI.

Not because of one model. Not because of one company. But because it made something very visible: frontier AI is no longer just software. It is infrastructure. It is power. It is policy. It is national strategy.

For a long time, Europe could look at the AI race from a slightly uncomfortable distance.

The United States had the biggest labs. China had the scale, ambition and state coordination. Europe had research talent, strong engineering, industrial companies and a lot of regulation.

That was not nothing. But it was not enough either.

The gap in large language models is real. Europe is behind at the frontier-model layer. Pretending otherwise would be useless.

But being behind is not the same as being finished.

Europe still has talent. Europe still has deep technical competence. Europe has hard industrial problems worth solving. Europe has companies that care about reliability, accountability, energy, privacy and long-term trust.

The question is whether European states, institutions, investors and builders can agree on what kind of AI capability Europe actually needs.

Because sovereign AI cannot only mean servers located in Europe. It cannot only mean regulation. It cannot only mean funding a few giant compute projects and hoping the rest solves itself.

Europe needs models. Europe needs infrastructure. Europe needs talent. Europe needs companies willing to build useful, trusted systems around AI.

And maybe most importantly: Europe needs a sharper idea of performance.

The brute-force AI race has limits

The first phase of the AI race has been brutally simple:

• Bigger models
• More compute
• Longer context windows
• More tokens
• More retries
• More agents doing more things in more loops

It works. A lot of progress comes from scaling.

But there is a limit to treating every problem as a reason to send more text into a larger model.

The cost becomes real. The infrastructure pressure becomes real. The environmental cost becomes real. And inside companies, the waste becomes real too.

AI agents can burn through huge amounts of context, tool calls, retries and generated text before producing something useful. Sometimes that is worth it. Often, it is not.

A future where every business process becomes ten AI agents burning tokens by shouting long prompts at each other is not intelligence.

It is expensive noise.

The next race will not only be about who can build the biggest model.

It will also be about who can produce the same or better result with less:

• Less irrelevant context
• Fewer retries
• Fewer tokens
• Lower cost
• Lower energy use
• Better traceability
• Better review
• Better control

That is not a step backward. That is maturity.

Europe should care about efficient intelligence

This is where I think Europe has an actual opening.

If Europe tries to copy the US and China only by chasing the largest possible model with the largest possible infrastructure, it may always be late.

But if Europe focuses on trusted, efficient, industrial AI systems, the picture becomes more interesting.

Europe understands constrained systems. Energy matters here. Privacy matters here. Traceability matters here. Procurement matters here. Responsibility matters here. Human approval matters here.

These things are sometimes mocked as friction. And yes, they can become friction when badly designed.

But they can also become an advantage.

Because the future of AI inside serious organizations will not only be about raw model capability. It will be about whether the organization can safely use that capability without losing control.

Who gave the AI the context? Where did that context come from? Was it the right context? Was it allowed? What did the AI produce? Who reviewed it? What changed? What was rejected?

Can we understand why a decision was made six months from now?

That is where the real work begins.

Prompting is not enough

Today, a lot of AI work still starts with a prompt.

Someone writes a long instruction. Then they paste background. Then they add documents. Then they explain the same thing again. Then the model misses something. Then they retry.

This is powerful, but it is also primitive.

Prompting puts too much burden on the human operator.

The human has to remember what matters, decide what to include, avoid sending too much and then interpret the result afterward.

That is not a sustainable foundation for AI-assisted work at scale.

AI does not need bigger prompts.

It needs better context.

What Nolta is trying to do

This is one of the reasons I am building Nolta.

Nolta is not trying to be another chatbot. Nolta is not trying to replace the model layer.

Nolta is trying to become the context, trust and execution layer around AI work.

Before AI acts, Nolta should help define what context it receives.

Not everything. Not a giant dump of documents. Not a messy prompt assembled by someone in a hurry.

But scoped, relevant context from the actual living structure of the work: projects, decisions, risks, history, relationships, documents, previous proposals, review notes and current state.

Then, when AI produces something, it should not disappear into a chat window.

It should come back as a proposal.

A human should review it. The useful parts should be approved. The weak parts should be rejected. The changes should become trusted only when approved.

And the whole thing should remain traceable afterward.

Nolta's principle is:

AI proposes. Humans approve. Nolta remembers.

Better context, less waste

The next step is making AI work more efficient without destroying traceability.

That means helping teams send the context that matters, not every document that happens to be nearby.

It means reducing irrelevant material, preserving links back to source evidence and making it clear what a human can inspect later.

The goal is not to hide complexity.

The goal is to remove waste without removing accountability.

This is not glamorous. But this is where real AI efficiency will come from.

Better AI is not only a model problem

We often talk about AI progress as if it only happens inside the model.

Better weights. Better training. Better reasoning. Better benchmarks.

But in real organizations, a lot of AI performance will come from everything around the model.

The context layer. The approval layer. The retrieval layer. The governance layer. The cost layer. The review layer. The memory layer.

The system that decides what the model sees before it acts.

The system that decides what happens after it answers.

A weaker model with the right context can beat a stronger model with the wrong context.

A smaller model with clean scope can be more useful than a frontier model drowning in irrelevant material.

A governed workflow can be more valuable than an impressive demo.

That is where I think Europe should focus its energy.

Not only on catching up with the largest model, but on building the systems that make AI usable, efficient, auditable and trustworthy in real work.

The next race

The next AI race will still involve compute. Of course it will.

But compute alone will not be enough.

The winners will not only be the ones who can spend the most tokens.

They will be the ones who waste the fewest.

The ones who can get excellent results with cleaner context. The ones who can prove what happened. The ones who can make AI useful without making organizations blind.

That is the race I want Nolta to be part of.

Less prompting. Less rework. Better context. Better proposals. Lower cost. Lower waste. More trust.

That feels like a European race worth entering.