Uber has burned through its AI budget, and companies' compute spending is overtaking payroll

Artificial intelligence is quickly moving from experimental spending to one of the heaviest lines in the corporate budget. At a number of companies, the cost of models, tokens and compute now competes not with software licenses but with payroll. A telling signal came from Uber: the company's chief technology officer said its annual AI budget for 2026 is already exhausted, and the main reason is the high cost of tokens. What was recently seen as a flexible tool for speeding up teams' work is now a full-fledged operating expense that cannot be ignored.

The problem is that AI economics do not scale like classic enterprise software. A subscription to an ordinary service is usually predictable, but spending on generative models grows with every request, every new user and every additional use case. The more a company builds AI into search, support, analytics, code, internal assistants and customer products, the higher the bill for tokens, cloud GPUs, data storage, monitoring and security. At the pilot stage these sums may seem tolerable, but once a service goes into production, the budget starts to jump, especially when thousands of employees or customers begin using it.

A significant part of the cost is also hidden: it is not in a single contract but spread across several lines, including APIs, cloud, vector databases, logging, data protection, fine-tuning, integration and maintenance. That is why AI in a corporate environment more and more often turns out to cost more than the first presentation or first invoice suggested.

Against this backdrop, the forecast for global IT spending looks logical: in 2026 it is expected to reach $6.31 trillion, 13.5% more than in 2025. A substantial share of the increase comes from AI infrastructure, software and cloud services. The money goes not only to access to models but to everything around them: server capacity, integrations, observability tools, tools for checking the quality of answers and data protection mechanisms. For business, this means AI adoption can no longer be treated as a local initiative of a single team. It calls for a systematic review of the IT budget.

Companies used to argue about whether to hire one more analyst, developer or manager. Now another question comes up more and more often: does a specific AI use case pay for its tokens and compute better than a human specialist or ordinary automation?

Management has to change accordingly. It is no longer enough for companies to simply let employees use neural networks or plug an API into a couple of products. They need financial limits, cost control for each use case, a choice between expensive and more compact models, caching of repeated requests and a hard assessment of where AI really creates value. Increasingly, companies are turning to routing requests by complexity, limits on context length, small models for routine tasks, and local deployments where confidentiality and predictable cost matter. In many cases it is cheaper to leave some processes to people or automate them with classic methods than to pay without control for every generation.

That is why the Uber example matters: even a large technology business with a strong engineering culture faces not a problem of access to AI but a problem of payback. The next stage for the market is not a race for the largest number of AI features but a shift to disciplined use. The winners will not be the companies that connect more models, but those that learn to calculate the cost of an answer, weigh it against revenue or time saved, and switch off expensive use cases with no clear effect in time.

AI is finally ceasing to be a fashionable add-on to the business and is becoming infrastructure. And infrastructure, as the Uber story shows, demands not only ambition but very strict budget control.