SemiAnalysis breaks down enterprise AI budgets: Meta once consumed 70 trillion tokens in a single month, but the real risk is not that customers stop using AI

Mars Finance News: On July 1, enterprise AI usage is shifting from “use as much as possible” to “use within a quota.” In its Token Budgeting report released on July 1, SemiAnalysis said that tokenmaxxing, which was popular at the start of the year—encouraging employees to consume as many AI tokens as possible to boost productivity—is being replaced by more realistic budgeting systems. However, the institute believes that media narratives about enterprises cutting AI spending have been exaggerated, and that the API businesses of OpenAI and Anthropic are not facing material budget risks in the second half of this year.

After speaking with more than 50 enterprise customers through Slack, phone calls, and the Databricks AI Summit, the SemiAnalysis team said they found that most companies are indeed starting to set usage limits for AI, but no unified standard has emerged. Low-end budgets may be as low as $250 to $500 per person per month, while high-end budgets can reach $2,000 per month and even tens of thousands of dollars. A major U.S. aerospace and defense manufacturer set some employees’ monthly allotment at $250, and a large pharmaceutical company set it at $500; more technology-forward firms such as Workday and Stripe have budgets of around $2,000 per month for some employees. This contrasts with the “token maximization” trend at the beginning of the year.

The report notes that companies such as Meta and Salesforce previously encouraged employees to heavily use AI tools. Meta even had an internal dashboard called “Claudeconomics” that ranked the top 250 heavy users. Data showed that Meta employees consumed more than 600 trillion tokens within 30 days, and the single highest user consumed about 280 billion tokens. The dashboard was shut down two days after related coverage. Uber was also reported to have used up the annual budgets for Claude Code and Codex within four months, after which it set a $1,500 per-person monthly limit, with over-quota requests requiring case-by-case approval.

However, SemiAnalysis believes these extreme cases reflect incentive mechanisms and lax management rather than an overall peak in enterprise AI spending. The report states that the top 10% of high-spending customers contribute most of the revenue for AI labs, and these customers face very low risk of cutting API spending for the rest of this year. Even though Meta consumed about 700 trillion tokens per month in February and, based on list prices, spent nearly $50,000 per employee per year, SemiAnalysis estimates that it still accounts for only 3% to 5% of Anthropic’s revenue.

Enterprise spending is also highly uneven. Citing Ramp data, SemiAnalysis said the top 1% of customers spend nearly $90,000 per employee annually on AI, the top 10% spend about $7,300, while the median customer spends only $136. The institute also said that many technology-leading Fortune 500 companies still spend less than $2,000 per employee annually on AI, and that large expenditures are mainly concentrated in engineering and data science departments. This means there is still substantial room for growth in the S curve of enterprise AI usage.

The rise of budgeting is changing how employees use AI. Some companies switch default models from Opus to Sonnet, disabling advanced models or fast modes; others have employees first use Microsoft 365 Copilot to draft and summarize, and then use more expensive Claude or Codex tokens for critical tasks. A global travel technology company spends nearly $10 million annually on AI; recently, it changed its default Claude model from Opus to Sonnet, but still allows employees to switch to Opus manually. Some roles have a default budget of only $200 per month, but engineers or senior employees can apply for higher quotas.

SemiAnalysis’s conclusion is that budget management will exist long term, but it does not equate to shrinking demand. Instead, enterprises are moving AI from experimental tools into formal cost management. Coding is currently the strongest demand vertical; SemiAnalysis estimates that more than 70% of the current ARR of OpenAI and Anthropic can be attributed to coding scenarios. In the future, cybersecurity, white-collar knowledge work, enterprise collaboration, and automated office work may replicate the growth path of Claude Code, Codex, and Copilot in the developer market.

This means the AI market is entering a new phase. Early-stage enterprises may have been willing to pay vague bills for “trying AI”; now, finance departments are starting to demand budgets, quotas, and ROI. But as long as improvements in employee efficiency offset costs, enterprises will not stop purchasing tokens. For AI model companies, the risk is not that customers suddenly stop using AI, but that they must prove that every dollar spent on token consumption translates into faster code, shorter hiring processes, higher sales efficiency, or less human effort.