SemiAnalysis breaks down enterprise AI budget: Meta once consumed 70 trillion tokens in a single month, but the real risk is not that customers no longer use AI According to BlockBeats, on July 1st, the use of enterprise AI is shifting from "maximizing usage" to "limited usage". SemiAnalysis stated in its Token Budgeting report released on July 1st that tokenmaxxing, which was popular at the beginning of the year to encourage employees to consume AI tokens as much as possible to increase productivity, is being replaced by a more realistic budgeting system. But the agency believes that media narratives about companies cutting AI spending have been exaggerated, and OpenAI and Anthropic's API businesses did not face substantial budget risks in the second half of this year. The SemiAnalysis team stated that after communicating with over 50 enterprise clients through Slack, phone calls, and VNet AI Summit, they found that most companies have indeed started setting limits on AI usage, but no unified standard has been established. The low-end budget may only be $250 to $500 per person per month, while the high-end budget can reach $2000 or even tens of thousands of dollars per month. A large American aerospace and defense manufacturer sets the monthly quota for some employees at $250, while a large pharmaceutical company sets it at $500; Enterprises with more advanced technologies such as Workday and Stripe have employee budgets of approximately $2000 per month. This contrasts with the "token maximization" at the beginning of the year. The report mentions that companies such as Meta and Salesforce have encouraged employees to extensively use AI tools. Meta even had a dashboard called 'Claudeeconomics' internally, ranking the top 250 heavy users of the company. Data shows that Meta employees consumed over 6 trillion tokens within 30 days, with a single highest user consuming approximately 280 billion tokens. The dashboard was closed two days after the relevant reports. Uber has also been reported to have exhausted Claude Code and Codex's annual budget within four months, and subsequently set a limit of $1500 per person per month, with excess requests requiring case by case approval. But SemiAnalysis believes that these extreme cases more reflect loose incentive mechanisms and management, rather than the overall peak of enterprise AI spending. The report states that the top 10% of high spending customers contribute the majority of AI lab revenue, and the risk of these customers cutting API spending for the remainder of this year is low. Even though Meta consumed approximately 70 trillion tokens per month in February and spent nearly $50000 per employee per year based on pricing, SemiAnalysis estimates that it still only accounts for 3% to 5% of Anthropic's revenue. The distribution of corporate expenditures is also highly uneven. SemiAnalysis cited Ramp data, stating that the top 1% of customers spend nearly $90000 per employee on AI per year, the top 10% of customers spend around $7300, and the median customer spends only $136. The agency also stated that many technology leading Fortune 500 companies still have an average annual AI expenditure of less than $2000 per employee, and large expenditures are mainly concentrated in the engineering and data science departments. This means that there is still a lot of room for growth in the S-curve of enterprise AI usage. The rise of budget systems is changing the way employees use them. Some companies switch the default model from Opus to Sonnet, turning off advanced models or fast mode; Some employees first draft and summarize using Microsoft 365 Copilot, and then use more expensive Claude or Codex tokens for critical tasks. A global tourism technology company spends nearly $10 million annually on AI and recently changed its default Claude model from Opus to Sonnet, but still allows employees to switch to Opus proactively. Some positions have a default budget of only $200 per month, but engineers or senior employees can apply for higher amounts. The conclusion of SemiAnalysis is that budget management will exist for a long time, but it does not necessarily mean a decrease in demand. On the contrary, companies are incorporating AI from experimental tools into formal cost management. Encoding is currently the strongest demand vertical domain, and SemiAnalysis estimates that over 70% of ARR in OpenAI and Anthropic can be attributed to encoding scenarios. In the future, cybersecurity, white-collar knowledge work, enterprise collaboration, and automated office may replicate the growth path of Claude Code, Codex, and Copilot in the developer market. This means that the AI market is entering a new stage. Early stage enterprises may be willing to pay vague bills for "trying AI"; Now, the finance department is starting to demand budgets, quotas, and ROI. But as long as the improvement of employee efficiency can offset the cost, the company will not stop purchasing tokens. For AI modeling companies, the risk is not that customers suddenly stop using AI, but that they must prove that every dollar spent on tokens can be converted into faster code, shorter recruitment processes, higher sales efficiency, or less manpower investment.