The race for AI in business is entering a new phase

Two years after the widespread adoption of AI assistants in companies, the time for discovery is over; now is the time for massive industrialization.

Autonomous agents connected to business tools, co-pilots integrated into almost all software suites, continuous automation, large-scale content generation: by 2026, AI is no longer a test. It has become an operational layer of the daily functioning of organizations.

But as usage becomes more widespread, another issue is now emerging in finance departments and infrastructure teams: the real cost of this growing dependence on AI .

Because despite the continued drop in model prices, the overall cost is skyrocketing in many companies. And so is the environmental footprint of the infrastructure.

The challenges needed to support this massive consumption remain far from being resolved.

By 2026, the race for AI had become a race for its application.

For several months, major technology players have been actively pushing companies to increase their consumption of AI.

The issue is no longer limited to “adopting AI”, but to using it everywhere, all the time:

• assistants integrated into messaging systems;

• agents capable of performing tasks independently;

• automatic code generation;

• automated reporting;

• permanent syntheses;

• document automation;

• enhanced search within internal databases.

In some organizations, token consumption has even become an implicit indicator of technological engagement.

The debate surrounding " tokenmaxxing "—the practice of intentionally maximizing the use of AI—has intensified following several internal revelations at tech giants. ( Tom's Hardware )

At Meta, an internal dashboard ranking employees based on their AI tool usage sparked controversy before being removed. ( Fortune )

What was still a weak signal in 2024 has become a real governance issue in 2026.

The costs of AI in business are no longer falling fast enough to offset usage.

The paradox is now well identified by finance departments: the cost of the token is falling, but AI spending is still increasing.

For what ?

Because the uses have changed scale.

Companies are no longer just paying for a few subscriptions to ChatGPT or Claude. They are now funding:

• autonomous agents operating 24/7;

• complex multi-model workflows;

• GPU infrastructures;

• security and compliance pipelines;

• massive storage;

• business integrations;

• human supervision systems;

• high availability architectures.

In other words: AI has become a permanent operational infrastructure.

Several IBM and Deloitte analyses published in recent months show that many companies are now struggling to stabilize their AI spending and demonstrate a truly measurable ROI at scale. ( IBM Think ) ( Deloitte )

The subject becomes particularly sensitive with the arrival of autonomous AI agents.

According to several industry analyses, some agents consume significantly more computing power than typical conversational uses, particularly when they make numerous model calls, perform checks, and execute tools continuously. ( Tom's Hardware )

After the AI euphoria, finance departments are taking back control

In 2026, faced with the explosion of AI costs , many companies are now entering a phase of rationalization.

The narrative has changed.

Just a year ago, the main challenge was “not to miss the AI revolution”.

Today, senior management is primarily asking for:

• which uses actually create value;

• which automations truly replace tasks;

• which tools justify their cost;

• and which projects should be stopped.

  
  

Some companies have already started to reduce internal licenses, limit certain uses, or redirect their AI projects towards more targeted and profitable use cases.

The topic of FinOps applied to AI — that is, the precise financial management of AI usage — is gradually becoming a standard in large organizations.

The environmental impact remains a major blind spot

Meanwhile, the environmental issue continues to grow in importance.

Because the mass industrialization of AI inevitably leads to an explosion in infrastructure needs:

• new data centers;

• increasing electricity consumption;

• intensive cooling;

• pressure on water resources;

• accelerated renewal of equipment;

• increased need for rare metals.

The problem is that the exact impact remains difficult to measure transparently.

Major suppliers provide little information on the detailed consumption of their models and on the real environmental cost of large-scale inference.

However, several studies and analyses now converge on one point: if current usage growth continues without major optimization, the energy footprint of AI could become a structural issue for both businesses and governments. ( Arxiv )

The argument that “models are becoming more efficient” is no longer enough to reassure some observers.

Because in practice, the efficiency gains are largely absorbed by the explosion in volumes consumed.

Towards a more restrained AI?

Faced with this situation, a new topic is gradually emerging in companies: that of AI sobriety.

The goal is no longer just to deploy AI quickly, but to do so in a way that is economically and energetically sustainable.

Several levers are beginning to emerge:

Governance of uses

Companies are now looking to:

• track costs per team;

• to cap certain uses;

• measure the real ROI;

• avoid high-consumption “gadget” uses.

Technical optimization

Many organizations are also reassessing their architectures:

• smaller models;

• reduced contexts;

• caching;

• Local or hybrid AI;

• finer selection of tasks that are truly automated.

Integration of environmental criteria

Some IT departments are also starting to integrate:

• the energy consumption of AI pipelines;

• the source of electricity for data centers;

• the lifespan of the equipment;

• and carbon impact indicators in technological trade-offs.

AI is entering its age of economic maturity

By 2026, AI is no longer an experimental subject.

It has become a major budget line, an infrastructure issue and now a question of sustainability.

The debate is therefore no longer solely about the capabilities of the models.

It also focuses on the ability of organizations to:

• to sustainably finance this increase in capacity;

• avoid inflationary practices;

• measure the actual value created;

• and limit the associated environmental impacts.

In other words: after the rush towards AI, perhaps comes the time for arbitration.