AI is getting cheaper. But is it becoming more valuable?
That question has been on my mind for some time.
The AI industry has become exceptionally good at measuring consumption. We measure tokens, API calls, GPU hours, model prices, latency, infrastructure utilisation and increasingly energy consumption.
These measures are useful. We need them to operate AI systems.
But they leave a much bigger question unanswered:
What did all that consumption actually accomplish?
This question became the starting point for my new book, AI Economics: Building and Scaling Intelligence.
The book is not really about the price of AI. It is about what happens when intelligence itself becomes an increasingly accessible, scalable and executable resource.
And I believe that changes the economics of technology in ways we are only beginning to understand.
The Token Is the Wrong Place to End
Much of today’s discussion about AI economics starts with tokens.
How much does a million tokens cost?
Which model is cheaper?
How much can we reduce inference costs?
These are legitimate engineering and procurement questions. But tokens tell us what AI consumed. They do not tell us what AI accomplished.
Imagine two AI systems.
The first consumes one million tokens and produces thousands of outputs, many of which need correction, trigger retries or are never used.
The second consumes two million tokens but reliably supports a decision that saves an organisation millions of dollars.
Which system is more efficient?
A token-based view might favour the first.
An economic view probably would not.
This is why I believe the measurement chain needs to move upward:
Compute → Output → Tasks → Successful Tasks → Outcomes → Value
Each step changes what we are measuring.
Compute tells us what resources were consumed.
Outputs tell us what the system produced.
Tasks tell us what work was attempted.
Successful tasks tell us whether that work actually succeeded.
Outcomes tell us whether something changed.
And value asks whether that change mattered.
The closer we get to value, the closer we get to the reason the AI system exists in the first place.
Agentic AI Makes the Economics More Interesting
This becomes even more important as we move from generative AI toward agentic systems.
With a conventional model call, the amount of work is relatively bounded.
An agent can decide to retrieve more information, call another tool, invoke another model, retry a failed step, ask another agent, verify its answer or continue reasoning.
The relationship between a user request and the amount of machine work required to satisfy it becomes variable.
That means agentic AI turns economics partly into a control problem.
The important metric can no longer simply be cost per request.
We need to understand things such as:
Cost per successful task
Energy per successful task
Failure and retry rates
Verification cost
Human intervention
Outcome achieved
An inexpensive agent that repeatedly fails may be considerably more expensive than a sophisticated agent that succeeds the first time.
Cheap AI that fails is expensive AI.
From Models to Skills
There is another economic transition happening.
As access to capable general-purpose models becomes increasingly widespread, simply having access to a powerful model becomes less differentiating.
Value starts moving elsewhere.
I describe this progression in the book as:
Model → Capability → Skill → Domain Skill → Industry Skill → Enterprise Skill
A model provides general intelligence.
A skill packages capability so it can be reliably reused.
A domain skill incorporates professional expertise.
An industry skill understands the particular requirements and context of an industry.
And an enterprise skill combines intelligence with the proprietary data, processes, policies, history and institutional knowledge of a particular organisation.
That last layer is particularly interesting.
If everyone can access similar models, competitive advantage increasingly comes from what the organisation knows, how effectively it can encode that knowledge, and how reliably it can apply it.
Expertise starts becoming executable.
What Happens When Expertise Becomes Executable?
This has implications far beyond software architecture.
Consider IT and professional services.
For decades, much of the industry’s economic model has been built around a relatively simple equation:
People × Utilisation × Rate = Revenue
Expertise scales by adding people.
Revenue scales through billable capacity.
AI changes that equation.
If a task that previously required ten hours can now be completed in two, the productivity improvement is obvious.
But under an hourly commercial model, something unusual happens:
the supplier becomes more productive while potentially generating less revenue.
This is what I call the Billable-Hour Paradox.
The future delivery model increasingly looks more like:
People + Agents + Skills + Platforms + Expertise → Outcomes
That changes how we think about utilisation, pricing, margins, offshore delivery, managed services, talent pyramids and intellectual property.
It also raises a fascinating question:
Who owns executable expertise?
If a consulting or technology-services firm works with a client’s experts to encode years of institutional knowledge into an AI skill, what exactly has been created?
Software?
Intellectual property?
A reusable methodology?
A client asset?
A supplier capability?
These are not simply legal questions. They are economic questions about who creates value and who ultimately captures it.
Intelligence May Become Abundant. Its Complements Will Not.
There is an even bigger shift underneath all of this.
For most of computing history, sophisticated machine intelligence was scarce.
That scarcity forced prioritisation.
When something was expensive, organisations had to decide whether it was worth doing.
AI is progressively removing some of that constraint.
But abundance does not eliminate scarcity.
It relocates it.
The scarce resources increasingly become things such as:
expert attention,
high-quality proprietary context,
evaluation capability,
human review,
governance and accountability,
infrastructure capacity,
energy,
water,
hardware,
and ultimately the organisational ability to determine which problems are actually worth solving.
This creates an interesting paradox.
As the cost of producing intelligence falls, organisations may consume far more of it.
Efficiency can therefore increase total demand rather than reduce it.
Which means AI economics cannot stop at financial cost.
Intelligence Has a Physical Economy
AI can feel intangible because its output appears on a screen.
Its production is anything but intangible.
Behind an AI response is a physical chain:
Capital → Materials & Hardware → Energy & Water → Compute → Intelligence → Outcome → Value
AI requires accelerators, servers, networking, storage, data centres, electricity, cooling systems, semiconductor manufacturing and eventually hardware replacement.
That means there is another side to AI economics.
We should ask not only:
What value did this AI create per dollar spent?
but increasingly:
What value did it create per unit of energy?
What value did it create relative to its carbon impact?
What value did it create relative to water consumption, in the places where that water was consumed?
These are not arguments for using value to justify unlimited resource consumption.
Absolute reduction still matters.
A high-value AI workload does not make its energy, carbon or water impact disappear.
Instead, these measures give organisations another decision lens:
Of the resources we do consume, are we directing them toward outcomes that matter, and can we deliver those outcomes with progressively lower impact?
That connects two conversations that are still too often separated:
how much AI consumes and what AI actually accomplishes.
From Efficiency to Allocation
This leads to perhaps the most important idea in the book.
AI economics is ultimately an allocation problem.
If compute is constrained, which workload receives it?
If expert attention is constrained, which AI system deserves review?
If energy capacity is constrained, which workload should scale?
If capital is constrained, which AI investment deserves the next dollar?
Traditional efficiency asks:
Resource / Outcome
How efficiently are we producing this result?
Allocation reverses the question:
Value / Resource
Where should the next unit of a scarce resource go?
Both views matter.
And neither removes the need to measure absolute consumption.
Why I Wrote AI Economics
I wrote AI Economics: Building and Scaling Intelligence because I think we are entering a phase where the economics of AI becomes as important as the capability of AI.
The first phase was about whether AI could perform the task.
The next was about making it cheaper and faster.
The emerging question is larger:
Which intelligence should we build, where should we deploy it, how much should we allow it to consume, and what value should we expect in return?
The book follows that journey across AI cost, infrastructure, agentic systems, context, failure, reusable skills, domain and enterprise expertise, outcomes, attribution, value, energy, carbon, water, hardware, IT services and ultimately the operating model required to manage intelligence as an economic resource.
But the entire argument can be reduced to two questions:
How much does our AI consume?
What does that consumption accomplish?
I think organisations that can answer both will have a very different understanding of AI economics from those that can only answer the first.
And as intelligence becomes easier to produce, what we choose to do with it may become more important than how much of it we can produce.
Read AI Economics
If these questions resonate with you, I explore them in much greater depth in my new book:
AI Economics: Building and Scaling Intelligence
The book develops a practical framework for understanding AI from compute and cost to skills, outcomes, resources, and value, and what that transition means for enterprises, AI practitioners, technology leaders, and the IT services industry.
AI Economics is now available in paperback from Amazon worldwide.
I would then keep your existing final thought immediately before this CTA:
And as intelligence becomes easier to produce, what we choose to do with it may become more important than how much of it we can produce.
That gives you a strong intellectual ending, followed by a clean conversion to the book rather than making the purchase link itself the conclusion
