As enterprises adopt AI tools on a large scale, new problems are beginning to emerge: the issue isn't whether the models are powerful enough, but rather that costs are rising too rapidly. Multiple technology and internet companies have found that despite a decline in the price of individual tokens, the total consumption of AI coding, automated assistants, and intelligent agent tools is still rising rapidly due to their widespread adoption.
Many companies have overdrawn their budgets in advance.
TechCrunch reports that some companies have already overdrawn their 2026 AI budgets. Uber had already used up its entire year's AI coding budget by April; Microsoft revoked the Claude Code licenses of some developers after several months of opening it up; and a Priceline employee said that Cursor's regular renewal quotes are 4 to 5 times higher than before.
This shift is related to the release of more powerful models in recent months. Anthropic, OpenAI, and Google have successively launched new models more suitable for intelligent agent scenarios since November of last year, driving a continued surge in usage. One company even incurred a Claude bill of up to $500 million due to the lack of employee usage limits.
Productivity increases may not cover costs
Alexander Embr, head of OpenAI's enterprise business, stated that six months ago, clients were more concerned with whether the model's capabilities were sufficient; now, the focus has shifted to expenditure visibility, auditing capabilities, token control, and model efficiency. The question for enterprises purchasing AI is changing from "what can it do" to "how much money is spent and is it worthwhile?"
The industry is beginning to recalculate the return on investment (ROI) for AI coding tools. A Faros AI survey of 20,000 developers in March showed that while development output is increasing, bugs and rework are also on the rise. Research from engineering management platform Jellyfish shows that engineers who heavily use AI are about twice as productive as those who use it less, but consume 10 times more tokens.
- Heavy AI users are approximately twice as productive as low-level users.
- The corresponding token consumption is approximately 10 times higher.
- A single developer's data consumption increased approximately 18.6 times within 9 months.
Cost management tools are being developed at an accelerated pace.
As billing issues expand, the market for tools surrounding AI cost management is heating up. This week, the Linux Foundation announced the establishment of the Tokenomics Foundation, hoping to create a unified language and management standards for AI token spending, much like FinOps in the cloud cost management field.
The organization plans to develop open standards and unified metrics for token usage and billing, as well as new cost-efficiency metrics such as "cost per unit of intelligence" or "tokens per watt." The official launch is expected in July, with more members to be announced at next week's FinOps X conference.
Meanwhile, both startups and established vendors are accelerating their deployments. Companies like Pay-i and Paid focus on AI cost tracking, measurement, and optimization; Jellyfish, Waydev, and Faros AI provide AI agent monitoring services; and Ramp, Datadog, and New Relic are also adding AI spending management, token-level observability, and GPU monitoring capabilities.
Model routing becomes a direction for cost reduction
Some investors and corporate executives predict that such capabilities will increasingly appear at the application layer or model routing layer in the future. For example, Factory, an enterprise AI startup, launched a model router this week that automatically selects the most suitable model for a task to reduce the cost of calling it. Similar practices have already appeared in some enterprise billing statements, where even when calling high-end models, the system will allocate some requests to cheaper models for processing.
Additional information:Goldman Sachs predicts that global token usage will increase 24-fold by 2030. For companies already in the high-investment phase, controlling costs while expanding AI usage is becoming a real challenge for the next stage of deployment.
