The AI revolution created itself based on an economic principle that was straightforward: the smarter the machines become, the lower the price for their intelligence will be.
It looked like a rational assumption because better models, less costly microprocessors, and more efficient software would lead to lower costs of generating units of machine intelligence. As the internet has done with the distribution of information, AI would enable cheaper execution of cognitive tasks.
However, enterprise AI has revealed another truth.
While the price of single tokens has become dramatically lower, businesses still experience an increase in their expenditures on AI. It is not the case because AI has stopped being cheaper. In fact, it is because businesses demanded more from machines.
So, if the cost of artificial intelligence is falling because it is becoming cheaper to produce, then why is it becoming more expensive for businesses to utilize?
Because the cost of AI does not depend anymore on the cost of a token. It also depends on how much work businesses expect from a machine.
The Token Is Getting Cheaper. The Work Is Getting Bigger.
Here’s how an ordinary chatbot interaction happens: a person poses a question; the model interprets it, comes up with an answer and stops.
Agentic AI alters this equation.
Rather than requesting the model for an answer to a certain question, a business can assign an AI-powered system a task. Then the system can split that task into sub-tasks, search for the needed data, invoke the corresponding software services, check and correct its actions and re-perform them before giving a final outcome.
This means that one single request may generate dozens or even hundreds of calls to the models.
According to TechNewsWorld, which cited a July report, the value of tokens dropped by 98% since early 2024, whereas enterprise AI invoices kept increasing. Moreover, the report estimated that the amount of operations generated by one single agent instruction is 5-30 times higher and may be hundreds of times larger.
The math here is quite simple.
If the cost of each particular action drops by 90%, but you perform 20 times more actions, the total invoice still rises.
Intelligent solutions don’t necessarily imply economical AI costs.
Why Does Agentic AI Cost More?
Because companies are not buying answers anymore. They are buying processes.
Think about a straightforward process, like reviewing a contract.
A human goes over the document, highlights important clauses, checks if those clauses agree with the company policy, and writes a summary.
An AI agent takes a roundabout approach. It extracts the contract, finds important clauses, queries its knowledge base, checks each clause against the company policies, finds past agreements, sends the findings for another model to review, fixes discrepancies, and generates a final report.
The machine will do the job faster, but with more computation.
Gartner spotted this exact issue back in August 2026, calling it “The Inference Paradox.”
The research showed that the economics of foundational models are getting better, whereas the more complex the AI processes become, the more tokens they use, resulting in higher overall inference costs.
Gartner predicts that the inference costs per agentic workflow will grow by more than five times through 2028.
The cheaper individual inference becomes, the easier it is to justify using more of them.
That is the paradox.
Where Is the Real Cost of AI?
This is only part of the equation for the cost.
All AI responses require a physical computing infrastructure. This infrastructure demands processors, memory, servers, networking hardware, data center space, electricity, and cooling.
As more workloads in the form of AI get done, there is more demand for all of this infrastructure.
This is where the cloud stops becoming an airy nothing.
The cloud is concrete, steel, silicon, electricity, and cooling working 24/7.
Why should you care about that?
Because even highly efficient software can’t solve the problem of physical limitations.
An efficient algorithm still needs some space to run. An inexpensive inference still needs electricity. More AI means more server and network capacity and more cooling capacity to support it.
More AI demand means more physical inputs for intelligence.
What Does the Nvidia–OpenAI Deal Tell Us?
The size of these recent infrastructure commitments gives an indication of how rapidly the physical aspects of the AI landscape are growing.
Nvidia entered into an agreement in August 2026 to make guarantees amounting to $105 billion connected to OpenAI’s 20-year lease of a large-scale data-center project in Ohio. This number does not include the cost of constructing just one building, and Nvidia isn’t merely writing a $105 billion cheque.
The eventual capacity of the campus could reach 8 gigawatts. Reuters noted that the finance deal involves certain leasing and power supply commitments for the project, which will aid in the infrastructure economics of the project.
Nvidia is also committing $1.5 billion towards SB Energy, which is SoftBank’s subsidiary that is working on this site.
Why is this significant?
Because this proves that the AI arms race requires physical capacity at a very large scale.
The model may exist in software. The economic underpinning most certainly does not.
Is AI Still a Software Business?
Yes.
However, larger scale deployments of AI make it increasingly resemble an industry that has economic features similar to that of the heavy industry sector.
The process of scaling up AI goes beyond writing more software. There also must be access to computing resources, electricity, cooling, land, and grid access.
These requirements are becoming visible now in Europe, as Reuters stated in August that data centre builders were relocating their projects from urban areas to those that have more affordable energy, land availability, and quicker connections to the grid.
Read Reuters about Europe’s move towards power-efficient data centre locations.
This way, the geography of AI is changing.
AI data centre’s ideal location is not necessarily the location with the biggest technological ecosystem anymore. It is rather the location that can provide fast access to physical resources like electricity, land, and infrastructure.
Are Companies Replacing Payroll With Compute?
Not always. The economics of the situation are actually far more complex than just swapping out a paycheck for an API invoice.
The cost of employing a human comes with wages and associated overheads. The cost of employing an AI comes with a different set of overheads in the form of models, cloud services, hardware, data, integration, and inference.
There can still be savings made by employing an AI that completes the task quicker, bigger, and with less labour.
But there isn’t a way to get around the cost of intelligence.
It’s just a matter of how you buy it.
And that’s why the concept of “compute payroll” is so important.
A business may find itself spending less on labour but becoming increasingly dependent on model providers, cloud, specialized chips, and power.
What Happens When Intelligence Becomes Cheap?
This is where the seeming paradox begins to resolve itself.
When a product gets cheaper, the number of ways in which it can be used tends to increase.
When storage got cheaper, it helped people generate huge volumes of data. When computation became cheaper, it helped companies to automate tasks for which expensive hardware and manual labour were needed.
AI too may follow the same logic.
As machine intelligence gets cheaper, companies can afford to employ it in many areas of their business.
Agents can be deployed to examine documents, systems, code, customers’ requests, anomalies, to prepare reports, and manage workflows.
What emerges is a much greater volume of machine intelligence activity.
The question thus ceases to be, “How expensive is one AI operation?”
Instead, what comes into play is the question, “How many AI operations will the companies undertake due to the cost of each one of them becoming low enough?”
This is where lower prices may lead to higher consumption.
Will AI Become Cheaper for Businesses?
Yes, in terms of per-unit.
But less expensive units do not necessarily imply lower total costs.
Future AI models are likely to become more efficient. More efficient hardware, more efficient software, more efficient routing, caching, and load balancing.
All of this might lower the cost of each particular task.
But businesses might react to that and leverage AI on a wider scale and make AI systems take on even more complex responsibilities.
Exactly what Gartner research highlights: improved unit economics drives the development of more expensive and more powerful AI systems, leading to increased total cost despite increased efficiency.
Read the Gartner full report about the Inference Paradox
In other words, businesses have to distinguish between unit economics and system economics.
A cheaper token is good. A cheaper workflow is better. But the ultimate metric is whether the additional economic value delivered by AI is higher than its cost.
The New AI Economy Is Physical
The first error would be seeing AI through the narrow lens of a software story.
AI’s next stage will be defined by algorithms and models, but also by hardware, data centers, power, cooling, land, and grid capacity.
This does not imply that AI is no longer growing more efficient.
It implies that efficiency has created new business demands on machines.
The token continues to get cheaper.
However, if companies deploy that cheaper intelligence to do ten times, one hundred times, or one thousand times more, total computation requirements can continue to increase.
This is the AI paradox.
We are not learning that artificial intelligence has turned out not to be cheap enough. We are learning that cheap intelligence can develop a hunger for far more intelligence.
The screen gives AI an appearance of weightlessness. The infrastructure presents a different picture.
Every new AI application will still require chips, servers, power, and cooling.
The intelligence may be digital. The bill is not.