I have spent the last couple of years working from two principles, and they still feel right.

First, AI is at its best when it amplifies people, not replaces them. It should sit inside an existing business and make decisions better, faster, and more consistently without pretending that a machine can simply substitute for judgement.

Second, a lot of enterprise work is better served by a governed local AI pipeline than by a generic cloud API. Not because local AI is always better, but because it gives the business more control, more visibility, and a cleaner cost model.

The second principle is not a claim that smaller models are always smarter. A frontier model can still be the right tool for a hard problem.

It is a claim that a careful system often produces better business outcomes than a clever prompt sent to a more capable model with no guardrails.

That is the difference between letting a model do work and designing a workflow that makes the work dependable.

And if the next six months go badly, it will be because too many organisations still treat AI like a utility they can bolt on without architecture, governance, or clear accountability.

The cloud argument is stronger than the market likes to admit

People talk about AI as if the only real question is whether inference is profitable. That is not the question that matters.

Inference is often profitable at the margin. That does not mean the whole business model is durable.

Every model provider is building huge infrastructure, paying for power, silicon, data centres, and R&D on a scale that cannot be ignored. The money is not just in the model run. It is in the capital stack behind it.

At some point the bill has to be paid by users, by enterprise budgets, by bundled software, by advertising, or by price increases. That is not a theory. It is arithmetic.

And once budgets tighten, the cost model changes. Service tiers change. Rate limits change. The “cheap AI” story becomes a little less stable than it looked at launch.

That does not mean cloud AI is a mistake. It means companies should not build their operating model around assumptions that are being made by the vendors and the market in real time.

Cheap tokens are not the same as cheap outcomes

One of the mistakes in the current AI conversation is that people equate lower token prices with lower business cost.

The unit price can fall while the number of tokens required rises. The system can become cheaper per token and more expensive overall.

Most agentic workflows do not run in a single call. They retrieve, route, reason, retry, summarise, and sometimes do the same task again. One business question becomes many model calls. Multiply that across a team, a department, or a company and the cost curve stops looking friendly.

That is why I keep coming back to the same idea: the business should measure cost by outcome, not by the headline price of a token.

A cheaper model that needs three attempts is not cheaper. A larger model that produces a confident but unsupported answer is not higher quality. The right question is what it costs to get a reliable result that can be trusted.

Local AI is a workload choice, not a moral position

I am not arguing for local AI because I have a romantic attachment to running things in-house.

I am arguing because different workloads need different controls.

Some work is fine in a shared cloud model. Some work should never leave the building. Some tasks require determinism, traceability, and a clear lineage of what data was used to produce the answer. Some do not.

That is where governance becomes part of the design, not a document tacked on later.

It is also where the hybrid estate starts to make sense. Local when control matters. Cloud when frontier capability matters. An explicit architecture, not a vague preference.

The real pressure is ROI, not hype

What may actually bite the market next is not whether AI is impressive. It is whether organisations can show a real return.

A lot of AI investments were justified by assumptions about labour replacement or dramatic productivity uplift. In many places that case has been weaker than the pitch suggested.

The companies that are doing well are the ones that used AI to improve a specific workflow, with clear ownership and clear metrics. Not the ones that bought a general-purpose “AI strategy” and hoped the economics sorted themselves out.

That is the real reason I think the next six months will be interesting. The market has spent a lot of money on enthusiasm. It is about to discover whether enthusiasm survives contact with accounting.

The value still sits with amplification

The evidence I keep seeing is not that AI is replacing people wholesale. It is that people who use AI well become more effective.

That is the version of the story I trust. Less dramatic than the replacement fantasy, but more useful. It matches what I have seen in real work: a good individual with a competent AI stack often accomplishes more than a much more expensive team without it.

That is not a slogan. It is a business model.

So the direction I still think is right is simple: design systems that own the reasoning pipeline, govern the data, and keep the important work under control. Use the cloud when it earns its place. Use local systems where the numbers and the risk say it should.