Enterprise AI buyers ditch largest models for cheaper task-specific tools
European businesses are shifting their AI spending away from expensive frontier models toward cheaper, task-specific alternatives as monthly enterprise bills run into the millions.
The assumption that the most powerful artificial intelligence model always wins is collapsing across the corporate sector. Enterprise buyers are now selecting AI tools based on specific tasks, cost, and data control rather than chasing top-of-the-league benchmark performance.
The shift is driven by stark financial realities. At an enterprise scale, running these massive systems costs millions of dollars a month. Even though the price per individual token has plummeted, overall corporate AI bills have actually tripled because newer agentic tools consume vastly more tokens to complete a single task.
That price shock is forcing a strict change in purchasing behaviour. Companies are capping employee AI spending outright in a wave of internal "tokenminimizing." Palo Alto Networks chief executive Nikesh Arora recently captured the industry mood, stating that "token prices need to fall by as much as 90% for adoption to scale."
The new operating principle for IT departments is to use the cheapest model that clears the quality bar. Buyers have realised that most routine corporate tasks simply do not require a frontier system. To automate this logic, model routing has emerged as a standard practice, directing each request to the most appropriate model.
Specialised, industry-specific models are filling the remaining gaps. Research firm Gartner expects 40% of enterprise applications to embed task-specific AI agents by the end of 2026. This represents a dramatic jump from under 5% just a year earlier, signalling a permanent structural shift.
The inference margin
If raw capability is becoming a commodity, the financial margins in the AI sector will migrate to whoever can run inference the cheapest. This dynamic is sharpened by Chinese models, which are rapidly closing in on the capabilities of American frontier labs at a fraction of the price.
The arrival of cheap, open models effectively caps what any provider can charge for merely competent output. This creates a distinct problem for the broader AI scaling thesis. Tech companies have justified hundreds of billions of dollars in capital expenditure on the premise that bigger models would remain decisively better.
Frontier models are not obsolete, but buyers are now demonstrating that everyday work does not require the most expensive tool available. For European enterprises watching their technology budgets, the AI race is no longer about size, but about operational efficiency.
