$3,000 a month, per engineer: the AI cost problem hiding in plain sight

For years, the cost of AI was hidden in the model training phase: millions of dollars in computing power that only the largest tech companies could afford. That paradigm is now reversing. According to industry projections, inference spending, that is, the cost of using already-trained models in real operations, will reach $23.3 billion USD this year and will surpass training spending for the first time.

The consequence is direct: an increasingly larger share of AI costs will arrive once companies are already using the models, not before. And that expense, unlike traditional software, does not behave predictably.

Gorilla Logic, a software engineering company with development centers in San José, Costa Rica, and Medellín, Colombia, decided to put that discussion on the table.

Its recent article laid out a scenario that could be difficult for any CFO to anticipate: allowing an engineer to use two agent-based programming tools could drive potential costs up to $3,000 USD per month per person, or $36,000 USD per year. With an organization of 4,000 engineers, potential spending could scale to about $144 million USD annually.

When tokens stop being a technical detail

The difference compared to traditional software lies in the nature of consumption. A conventional license has a relatively stable price per user.

An AI agent, by contrast, can consume radically different amounts of tokens depending on the task: a simple process may require little use, while a complex task may involve thousands of tokens of context, generation, and reasoning.

That variability turns technology spending into an unpredictable line item within the income statement. "AI token costs have gone from a minor technology detail to a board-level budget problem in under a year. Twelve months ago, 'token cost' wasn't a line item on most CFOs' radar — now it's showing up," noted Gorilla Logic.

Bob Graham, Chief Growth Officer of Gorilla Logic, argues in the article that the problem goes beyond an unexpectedly high bill.

"There is a lesson to be drawn here," said the executive.

"If the biggest technology companies in the world can be surprised by this expense, so can anyone. Token consumption has become a genuinely new expense category, one that did not exist on any P&L two years ago."

The challenge is beginning to appear in external research as well. McKinsey noted in July 2026 that some companies are already exhausting their annual AI budgets within months, forcing them to renegotiate contracts or seek additional financing.

Gartner, for its part, warns that AI is turning technology spending from a relatively fixed model into a variable one, forcing CIOs and CFOs to control not only licenses and contracts but also consumption.

It is a fundamental shift in the way technology is managed. The budget stops being an annual exercise of negotiation with vendors and becomes a problem of continuous monitoring, closer to managing a payroll or energy consumption than to buying software.

Wrote Bob Graham, "Permitting the use of two agentic coding tools together (say Claude and Cursor) and that ceiling doubles to $3,000 a month, per engineer $36,000 a year for one person's AI usage."

For Gorilla Logic, the conversation about costs takes place on top of an engineering infrastructure built over more than a decade in Latin America.

The company opened its nearshore development center in San José, Costa Rica, in 2014. Five years later, in 2019, it added a second center in Medellín, taking advantage of the city's growing tech ecosystem.

Today, the company presents Costa Rica as one of the most consolidated software engineering ecosystems in Latin America, while describing Colombia as one of the fastest-growing markets for technical talent in the region.

The gap that defines the new AI economy

For CFOs, the challenge is learning to manage a cost that behaves more like a payroll than like a license. For CIOs, it is building the architecture that prevents consumption from spiraling out of control.

And for companies in general, the underlying challenge is closing the gap between what the tools consume, what engineers produce, and what ultimately reaches the business.

The new AI economy is not defined solely by how much it costs to use it, but by how much value it generates. And that account, for now, remains open.