The use of AI by enterprises is moving from “first-to-first-to-first” to budgetary constraints, but overall inputs are not significantly cooling. SemiAnalysis, after researching more than 50 businesses during the Databricks AI Summit, stated that recent cases of over-expenditure of concern were more of an out-of-control of individual management than of an overall imbalance in corporate AI spending.

Extreme overexpenditures are not common.

In the first half of this year, some of the major companies were concerned about the high staff consumption of Token. Meta has encouraged staff to use the AI tool as much as possible, and Uber has been tightening the budget for short periods of time. SemiAnalysis argues that such incidents are mainly related to incentive design and internal licensing, and do not represent the enterprise as a whole.

According to Ramp data, the average annual AI expenditure per head 1 per cent of clients was approximately $90,000, the top 10 per cent was about $7300, while the median client was only $136. Even in enterprises with higher technology adoption, most of their staff did not enter the HF call phase.

Anthropic's developer files also show that Claude Code spends approximately $150 to $250 a month, and only a few users spend more than $30 a day. This means that the company AI is still concentrated in part of the job, rather than in the whole.

Enterprises start to set clear limits

Most of the enterprises visited had a ceiling on the use of AI by their employees, but standards varied widely and there was no uniform approach. Some manufacturing and pharmaceutical companies have a monthly limit of between $250 and $500, while others such as Workday and Stripe have a monthly budget of approximately $2,000 for their employees.

There are also enterprises that allocate their budgets by post and project. A large tourism technology company has by default allocated a lower level of staff, which is then increased by duty. A large United States airline incorporated Token ' s costs directly into the project ' s financial model, with the team allocating itself within the total budget, rather than as a separate IT expenditure.

With a tight budget, enterprises are also adjusting their use patterns. Common practices include the use of Microsoft Copilot for basic tasks and then calling Claude or Codex to perform more complex tasks; or switching default models from more expensive versions to lower-cost versions to reduce unit call costs.

Low price models and API markets continue to expand.

SemiAnalysis argues that budget management does not amount to a reduction in mobilization, and that the enterprise is more concerned with cost efficiency. The task that can be done with cheaper models will not allow enterprises to rely on high-price models for long periods of time, which will change the procurement structure but will not make demand disappear.

According to the Agency, the Token-as-a-Service and API endpoint markets of front and open source models are still growing. After taking AWS Bedrock into account, its judgement on the growth rate of AWS-related operations during the season was higher than the market-wide expectations. Total annual recurrent income from suppliers such as Together, Fireworks and Baseten has also exceeded $4 billion.

From the application scene, coding remains the strongest driver of AI income for enterprises. SemiAnalysis estimates that OpenAI and Anthropic have more than 70% of annual recurrent income in this direction. The next stage of growth may come from a broader scenario such as cybersecurity and white-collar knowledge work.

Efficiency gains have been made, but they're also increasing.

It is generally acknowledged that AI tools have led to significant efficiency gains. In the case of recruitment, data analysis and other processes, work that would have taken weeks or even months has been condensed to a few hours or months and completed in shorter cycles.

However, efficiency gains have not been translated directly into a lighter work tempo. Some staff indicated that, with the help of AI, the completion of the task was faster, and that the company had simultaneously increased output expectations, with the result that the pace of work had been tightened.

SemiAnalysis concludes that Enterprise AI is moving from an unbounded test phase to a precision management phase. Budget ceilings are becoming the norm, but coding, automation and enterprise-level API procurement are still expanding, suggesting that this round of AI inputs has not slowed, but has shifted from broad use to cost optimization.