The Economics of Intelligence: What Building With AI Taught Me About Tokens, Models and Energy

Over the past year, I have spent a significant amount of time building with AI. Not experimenting with a few prompts or asking ChatGPT to improve an email.

I have been building HEA-World with AI deeply involved in the engineering process — there is no way I could have pulled this off by myself — while simultaneously designing AI-based systems that may one day process large volumes of interactions between organizations and people.

HEA-World operates around 10 models in production today, while maintaining a 28-model catalog across 6 providers and a 90-path qualification matrix for safe model rotation.

My perspective comes from building almost continuously with a fairly broad mix of AI models and tools. GPT has been a constant companion, spanning GPT-4 through GPT-5.5. I experimented with Grok for a while. I have used Codex extensively for engineering, Google's Antigravity with Gemini and Claude, and Cursor, more recently including Composer.

Thousands of conversations, architecture discussions, code reviews, implementation cycles, broken builds, rewrites, content creation and, occasionally, some very frustrating evenings.

So what follows is not a scientific comparison of AI models. It is my REX, my return on experience, from putting them to work on the same increasingly complex system.

I have not solved the economics of AI. But I have clearly learned a lot. And one conviction is becoming stronger:

Token efficiency does not mean using the cheapest model. It means allocating the right intelligence to the right task.

The more I work with AI, the less I think of models as a simple ranking from less intelligent to more intelligent. Reasoning matters. Speed matters. Context matters. Tool use matters. Emotional intelligence matters. And, increasingly, the energy required to deliver all of this matters too.

Perhaps we need to stop asking which model is the best. The more useful question may be: best for what?

The Cheapest Model Can Produce the Most Expensive Outcome

My experience with AI-assisted engineering has changed significantly as models have improved. For complex system design and engineering work, I increasingly want access to the most advanced model available.

The reason is simple: the cost of sloppiness is too high.

A cheaper or less capable model can move incredibly fast. The first iteration can look impressive. But in a complex system, small architectural misunderstandings accumulate.

Context gets lost. Existing design principles are ignored. A quick fix creates another problem somewhere else.

Then comes rework. And rework is expensive. Not only in tokens. It costs human review time, debugging, lost confidence and the mental energy required to understand what the AI actually changed.

I have experienced this directly while building HEA-World. Sometimes a faster model feels extremely productive for the first few hours. Then, as iterations accumulate and the system complexity catches up, I find myself spending more time reviewing, correcting and cleaning.

The token price looked attractive. The outcome was not necessarily cheaper.

A more advanced model that understands the system better and gets the architecture right can ultimately cost significantly less.

The cost of intelligence should not be measured per token. It should be measured against the cost of the outcome.

This is why I believe demand for frontier AI models will continue to grow. As more engineers, researchers, creators and organizations use AI for increasingly complex work, the best reasoning models will remain in high demand.

At HEA-World, I expect us to continue using the most advanced models available for engineering, complex system thinking and selected creative tasks where quality matters disproportionately. Video is probably a good example. But I do not think this is how most AI tokens should ultimately be consumed.

Most Business Problems Are Not Intellectually Difficult. They Are Contextually Heavy.

While building HEA-World, I have become increasingly interested in another category of model.

• Extremely fast.
• Able to absorb a large amount of context.
• Reliable at using tools.
• But not necessarily requiring frontier-level reasoning.

Many business AI tasks are not intellectually difficult. They are contextually heavy.

An AI may need to understand the organization, its services, the customer, the conversation history, the current system state, permissions and the tools it is allowed to use. That is potentially a lot of context. But once the context is understood, the task itself can be relatively straightforward.

• Retrieve.
• Decide.
• Call the right tool.
• Execute.
• Respond.

I suspect this type of fast, high-context operational intelligence may become the real workhorse of AI-based automation.

At one end of the spectrum, we will have frontier intelligence for difficult reasoning, engineering and high-value creation. At the other end, nano intelligence for classification, extraction, routing, scoring and highly bounded repetitive tasks. And between them, perhaps a massive layer of operational intelligence: models designed to understand rich context and reliably act.

This is already influencing how I think about the architecture of HEA-World. But there is another dimension that makes the allocation problem even more interesting.

Intelligence Is Not Only Reasoning

People talk about intelligence all the time when comparing AI models.

• Reasoning benchmarks.
• Coding benchmarks.
• Mathematics.
• Scientific problems.

But one of my favourite capabilities of an AI model is much harder to put on a leaderboard.

Emotional intelligence and empathy.

I have not done enough research in this field to make strong scientific claims. This is purely an observation from my own experience.

But I would absolutely choose one model over another based on its emotional intelligence and its capacity for empathy, at least for models designed to interact with humans.

Over the past year, I have worked almost non-stop with AI models while building HEA-World. They have been present through the excitement of a breakthrough, the satisfaction of seeing a complex system finally come together, the frustration of a broken build and the moments when I seriously questioned some of my own product decisions.

When you work with models at this intensity, their emotional qualities, and their defects, become amplified.

• A model that does not recognize frustration and enthusiastically suggests ten more ideas can become exhausting.
• A model that agrees with everything can reinforce bad decisions.
• A model that becomes overly cautious at the wrong moment can destroy momentum.
• And a model that understands the difference between exploration, doubt, frustration and execution can become a remarkably effective working partner.

I want this in my own daily work with AI. But more importantly for what I am building, I want it for the people who will interact with an HEA.

• A customer asking a simple product question may only need a fast and accurate answer.
• A frustrated customer does not.
• Someone hesitating before an important purchase may need clarity without pressure.
• A person contacting an organization during a difficult moment may care as much about how the answer is delivered as about the information itself.

Perhaps the right model should not only be determined by the reasoning complexity of a task.

Perhaps the emotional complexity of the interaction should matter too.

A nano model may classify the request. An operational model may retrieve the customer history and available options. But the system may recognize that the conversation has become emotionally sensitive and allocate a different kind of intelligence to the interaction.

Not because the question suddenly became intellectually difficult. Because the human context changed.

We spend a lot of time benchmarking artificial intelligence on mathematics, coding and reasoning. For human-facing AI, I suspect we will become increasingly demanding about something much harder to benchmark:

How does this intelligence make us feel when we work with it?

For me, that is already becoming a model selection criterion.

AI Automation Is an Intelligence Allocation System

This is why I increasingly see AI-based automation as more than a productivity tool. It is a governance layer for intelligence.

A human opens an AI interface and says:

“Here is my problem. Think about it.”

The model has to interpret the problem, understand intent, decide what matters, reason and respond.

A well-designed automation system can be much more deliberate.

• Classify this interaction.
• Understand its context.
• Assess its complexity.
• Recognize the available tools.
• Consider the permissions.
• Perhaps even understand the emotional state of the conversation.
• Then allocate the right model.

The same architecture that controls which data an AI can see and which actions it can execute can also decide which intelligence should handle the task, how much context it needs and how many tokens the task deserves.

Quality guardrails and token-efficiency strategies become part of the same system. I increasingly see this as a new design discipline.

Tokenomics is becoming the art of allocating intelligence.

The Energy Question May Also Be an Architecture Question

My professional background is in energy, so it is difficult for me to look at the expansion of AI without thinking about electricity demand.

The current data-centre build-out is extraordinary. We are designing increasingly large and power-dense infrastructure to support AI growth. And I understand why.

If millions of engineers, researchers, creators and businesses increasingly delegate complex work to advanced AI models, demand for frontier intelligence will grow substantially. I am contributing to that demand myself. But I keep wondering about the long-term workload mix.

What if 80% of future AI traffic eventually does not require frontier intelligence?

The number is deliberately hypothetical. It is a thought experiment, not a forecast.

But imagine that most AI interactions eventually happen inside well-designed automation systems.

• Classification.
• Extraction.
• Routing.
• Monitoring.
• Routine customer interactions.
• Structured content generation.
• Business workflow execution.

A huge amount of this work may eventually be handled by nano or fast operational models.

Would we design the same data-centre infrastructure if most AI demand did not require maximum reasoning? I am not convinced we would.

The total demand for AI could still explode. Efficiency usually creates new usage, and cheaper intelligence will almost certainly create applications we cannot anticipate today. But the shape of infrastructure demand could be very different.

We may be planning AI infrastructure around peak intelligence demand while much of future AI consumption becomes something closer to base-load intelligence.

• Frontier models handle difficult, high-value and unpredictable problems.
• Operational models absorb rich context and execute workflows.
• Nano models process enormous volumes of repetitive and bounded tasks.
• And AI automation becomes the dispatch layer deciding which intelligence resource to call.

Coming from the energy sector, that analogy is increasingly difficult for me to ignore.

Distributed Energy. Distributed Intelligence.

There is one more parallel I keep thinking about.

The energy system is becoming increasingly distributed. Solar generation is spread across millions of assets. Wind generation is geographically dispersed. Batteries are appearing at grid, commercial and residential levels. The future electricity system is becoming more dynamic and increasingly dependent on our ability to match flexible demand with variable supply.

Could the evolution of AI infrastructure eventually follow a compatible path?

Frontier models will probably continue to require enormous, concentrated computing infrastructure. Complex reasoning, engineering and scientific workloads may justify highly specialized clusters with extraordinary power density.

But a world dominated by nano and operational models creates a different possibility.

• Many AI tasks are bounded.
• Some are latency-sensitive.
• Many are not.

A CRM analysis can run overnight. A content brief can wait a few minutes. Classification, extraction, monitoring and background processing can potentially be queued, distributed and scheduled.

In energy terms, these are flexible loads.

If the compute itself becomes more distributed, an interesting architecture starts to emerge. Smaller AI inference workloads could potentially be located closer to distributed renewable generation and battery capacity. Workloads could move geographically or shift in time depending on energy availability, grid constraints and carbon intensity.

Instead of only building ever-larger power systems around concentrated AI data centres, perhaps we will also learn to distribute part of the intelligence demand around the energy system we are already building. I see a potentially powerful structural match.

Distributed renewable energy.
Distributed storage.
Distributed intelligence.

And again, automation could become the orchestration layer. The same system that decides which model deserves a task could eventually consider when and where that task should be executed.

• Not every token is equally urgent.
• Not every model requires the same infrastructure.
• Not every human interaction requires the same emotional capability.
• And perhaps not every AI workload should be computed in the same place or at the same time.

The energy industry has spent years learning how to orchestrate millions of distributed assets. AI may be about to create millions, or billions, of distributed intelligence workloads. The two worlds may fit together better than we currently assume.

Learning to Allocate Intelligence

I do not believe the answer to AI's growing energy demand is to stop using advanced models. Quite the opposite.

My own experience has convinced me that using the best available model for difficult engineering work can be more efficient than using a cheaper model and paying for rework afterwards. But we should be equally demanding about where frontier intelligence is actually required.

• Use advanced reasoning where complexity and quality justify it.
• Use fast, high-context operational intelligence where the challenge is understanding context and reliably executing.
• Use nano models for high-volume, well-defined tasks.
• For human-facing interactions, consider the emotional complexity of the moment.
• Build governance systems that can route between them.
• Design quality guardrails so that efficiency does not quietly become mediocrity.
• Escalate when a system reaches the limits of the intelligence it has been allocated.
• And perhaps, eventually, make time, location and energy availability part of that orchestration decision too.

Maybe the future of AI efficiency will not be defined by a single breakthrough model or a more efficient GPU. Maybe a significant part of it will come from better system design.

In the energy world, we learned that managing demand can sometimes be as important as building new generation capacity. I suspect we may eventually learn something similar about artificial intelligence.

Before asking how much AI infrastructure we need, perhaps we should first ask how intelligently we intend to consume intelligence.

#ArtificialIntelligence #Tokenomics #AIModels #AIAutomation #AIAgents #Energy #DataCenters #RenewableEnergy #DistributedEnergy #HumanEnhancedAI #HEAWorld #BuildInPublic