SemiAnalysis Breaks Down Enterprise AI Budgets: Meta Once Burned Through 70 Trillion Tokens in a Single Month, but the Real Risk Isn't Customers Abandoning AI

BlockBeats News, July 1st. Enterprise AI usage is shifting from "maximizing usage" to "budgeted usage." According to a Token Budgeting report released by SemiAnalysis on July 1st, the once-popular practice of tokenmaxxing, which encouraged employees to consume AI tokens as much as possible to boost productivity, is being replaced by a more realistic budgeting system. However, the organization believes that media narratives about enterprises cutting AI expenditure have been exaggerated, and OpenAI and Anthropic's API businesses did not face substantial budget risks in the latter half of this year.

The SemiAnalysis team stated that after communicating with over 50 enterprise clients via Slack, phone, and the Databricks AI Summit, they found that most companies are indeed starting to set limits on AI usage, but there is no unified standard. Lower-end budgets may be as low as $250 to $500 per person per month, while higher-end budgets can range from $2,000 per month to tens of thousands of dollars. A large U.S. aerospace and defense manufacturer set a monthly limit of $250 for some employees, a major pharmaceutical company set it at $500; tech-forward companies like Workday and Stripe allocate around $2,000 per month for some employees.

This contrasts with the early-year "token maxing" approach. The report mentioned that companies like Meta and Salesforce had encouraged employees to heavily use AI tools. Meta even had a dashboard internally called "Claudeconomics" that ranked the top 250 heavy users in the company. Data showed that Meta employees consumed over 600 trillion tokens in 30 days, with the highest individual user consuming around 280 billion tokens. The dashboard was shut down two days after related reports were published. Uber was also reported to have exhausted the annual budget for Claude Code and Codex in four months, then set a limit of $1,500 per person per month, with overages requiring individual approval.

However, SemiAnalysis believes that these extreme cases more reflect incentive mechanisms and loose management rather than a peak in overall enterprise AI spending. The report stated that the top 10% high-consumption customers contributed most of the AI lab's revenue, and these customers have a very low risk of reducing API expenditure for the remainder of this year. Even though Meta consumed approximately 700 trillion tokens per month in February and the estimated annual cost per employee based on list price was close to $50,000, SemiAnalysis estimates that this would still only account for 3% to 5% of Anthropic's revenue.

Corporate spending distribution is also highly uneven. SemiAnalysis cites Ramp data indicating that the top 1% of customers spend almost $90,000 per employee per year on AI, the top 10% spend around $7,300, and the median customer spends only $136. The organization also notes that many technology-forward Fortune 500 companies still have AI spending per employee below $2,000, with large expenditures primarily concentrated in the engineering and data science departments. This means that there is still significant room for growth in the S-curve of enterprise AI adoption.

The rise of budget discipline is transforming employee usage patterns. Some companies are transitioning from default Opus models to Sonnet, turning off premium or fast modes; while some employees first draft and summarize using Microsoft 365 Copilot before leveraging more expensive Claude or Codex tokens for critical tasks. A global travel technology company spends nearly $10 million on AI annually, recently switching its default Claude model from Opus to Sonnet, but still allowing employees to voluntarily switch back to Opus. Certain roles have a default budget of only $200 per month, but engineers or senior staff can request higher limits.

SemiAnalysis concludes that budget management will persist in the long term, but it does not equate to shrinking demand. Instead, businesses are formalizing AI into their cost structures from an experimental tool. Coding is currently the strongest demand vertical, with SemiAnalysis estimating that over 70% of OpenAI and Anthropic's ARR can be attributed to coding scenarios. In the future, areas such as cybersecurity, white-collar knowledge work, enterprise collaboration, and automated office tasks may replicate the growth trajectory of Claude Code, Codex, and Copilot in the developer market.

This indicates that the AI market is entering a new phase. Early adopter enterprises might have been willing to pay a vague bill for "trying AI"; now, finance departments are starting to demand budgets, limits, and ROI. However, as long as the improvement in employee efficiency can offset costs, businesses will not cease token purchases. For AI model companies, the risk is not that customers will suddenly stop using AI, but rather they must demonstrate that every dollar of token consumption can translate into faster code, shorter recruitment processes, higher sales efficiency, or reduced human effort.