Corporate Reality After the Token Frenzy Recedes: Budget Control Becomes the New Normal, but the AI Train Hasn't Slowed Down

The era of "unlimited consumption" of AI for enterprises is coming to an end, but cost control does not mean slamming on the brakes.

In the first half of this year, "Tokenmaxxing" became a buzzword in the enterprise AI community. Companies like Meta and Salesforce actively encouraged employees to consume as many AI tokens as possible to boost productivity. Meta even had an internal leaderboard called "Claudeconomics" tracking the top 250 heavy users. Data shows that Meta employees consumed over 60 trillion tokens in 30 days, with a single top user reaching approximately 280 billion tokens. Employees competed for titles like "Token Legend" and "Cache Wizard," having AI agents spend hours on meaningless research to "burn tokens."

Subsequently, Uber made headlines for burning through the entire annual budget of Claude Code and Codex in four months, then imposed a monthly usage cap of $1,500 per employee, with any excess requiring case-by-case approval.

These incidents drew widespread attention. However, after conducting field research at the Databricks AI Summit and engaging in in-depth conversations with over 50 enterprise clients via Slack and phone, the SemiAnalysis team reached conclusions that differ sharply from the media narrative.

The firm believes that media reports have greatly exaggerated the prevalence of the problem. The extreme cases at Meta and Uber stem from misaligned incentives and lax internal management, not from an overall loss of control over enterprise AI spending.

Headline stories are exaggerated; the real data is more moderate

Media coverage may have overstated the crisis in enterprise AI budgets.

Key data supports this assessment. Citing Ramp's spending data, SemiAnalysis shows that the top 1% of customers spend an average of approximately $90k per employee per year on AI, the top 10% about $7,300, while the median for all Ramp customers is just $136. Ramp customers already have a much higher level of technology adoption than the average enterprise, yet the per-employee AI spending of Fortune 500 media clients remains well below $100.

Even at Meta, a major "token burner," annual spending per employee at list price is close to $50k, but SemiAnalysis estimates this accounts for only 3% to 5% of Anthropic's customer revenue.

Anthropic's own documentation confirms this: Claude Code developers spend an average of just $150 to $250 per month, and only 10% of users spend more than $30 per day.

SemiAnalysis argues: "The media's exaggerated reports are not true—enterprises are still investing continuously, and token consumption driven by new demand scenarios and verticals is pushing the AI train forward at a rapid pace."

This means enterprise AI adoption is still in a phase of uneven diffusion. Not all employees are using large models frequently. In many companies, only a few teams or roles are taking the lead.

Budget control becomes the new normal, but standards vary widely

Most of the 50+ surveyed enterprises have set hard caps on AI usage. However, standards differ significantly from company to company, with no industry consensus.

Low-end cases:

• A top-three U.S. aerospace and defense manufacturer: $250 per person per month cap; some heavy users exhausted their quota within four days of the first month.

• One of the world's largest pharmaceutical companies: $500 per person per month, with $1,000 available for special cases upon request.

High-end cases:

• Workday, Stripe: Monthly employee budget around $2,000.

• A publicly traded cybersecurity company: Junior employees $800 per month, senior employees $1,600 to $4,000, with data scientists receiving the highest limits.

• A major travel tech company (800 engineers out of 1,500 employees, annual AI spend near $10 million): Default $200 per person per month, which can increase to tens of thousands based on role.

The logic behind setting budgets also varies. One of the three major U.S. airlines has the most distinctive approach: token allocation is directly tied to specific projects and expected revenue. For example, for a project with an expected revenue of $10 million, the finance team approves a total budget of $1 million, and the team decides how much of that to use for tokens—AI costs are embedded into the project's financial model, not treated as a separate IT budget.

Employee survival tactics for "saving tokens"

Budget pressure has spawned a set of practical token-saving strategies.

The most typical is "Copilot arbitrage": Microsoft 365 enterprise subscribers can use the standard Copilot chatbot for free and unlimitedly, and this usage is not counted against the monthly AI budget. A large Dutch consumer goods and health technology company explicitly stated that employees first draft and consolidate ideas using Copilot, then call on Claude or Codex for final tasks, thereby saving measured tokens.

Model downgrading is also common. The global travel tech company has switched all employees' default Claude model from Opus to Sonnet; Opus is still available but must be selected manually. The aerospace and defense manufacturer has directly "disabled" Opus 4.8 and fast mode.

In response, the SemiAnalysis team offered a direct critique of management's logic: "Management believes that giving employees larger token budgets would prompt them to automate tasks that should not be automated, such as writing emails. We find this anti-automation perspective naive."

Demand for cheap tokens is still growing; the TaaS/API endpoint market has not cooled down

Budget management does not mean reducing usage. Enterprises care more about unit cost.

Demand for cheap tokens remains strong. Both the token-as-a-service/API endpoint markets for frontier models and open-source models are growing. After including AWS Bedrock in calculations, SemiAnalysis estimates AWS's overall growth rate for this quarter is higher than market expectations.

TaaS providers are also expanding. Companies like Together, Fireworks, and Baseten have a combined ARR exceeding $4 billion.

This shows that enterprise budget pressure will change the procurement structure. Problems that can be solved with cheap models will not always use the most expensive models. Downgrading default models does not mean reduced AI usage, but rather a re-optimization of the cost curve.

Coding remains the strongest demand; the AI train has not slowed due to budget caps

Coding scenarios are still the biggest driver of AI revenue, with OpenAI and Anthropic generating over 70% of their ARR from this area. Anthropic's B2B share exceeds 90% (OpenAI about 60%), making its revenue structure more dependent on and stable with enterprise clients.

The next wave of growth is expected to come from cybersecurity and white-collar knowledge work. As Cowork, CoPilot, Codex, and Computer products further penetrate enterprises, the path by which coding markets drive AI lab ARR growth will be replicated in broader scenarios.

Currently, most Fortune 500 enterprises spend far less than $2,000 per employee per year on AI, mainly concentrated in engineering and data science departments. This means that AI adoption in enterprises is still in its early stages, and growth potential has not disappeared—it has simply shifted form, from "uncontrolled spending" to "continuous investment within a budget."

The real ROI of AI: Efficiency gains, but output expectations are also rising

Among surveyed enterprises, cases of AI-driven efficiency improvement are real and significant.

• Amazon's recruiting department: The process from initial screening to team placement once took 6 to 9 months; with AI tools, it has been shortened to 3 to 4 months.

• A company that provides data analytics services to 85% of the Fortune 500: Work that previously took a week now takes just a few hours.

However, the flip side of efficiency gains is that output expectations also rise. An employee of a legal data and risk solutions company admitted that while a week's workload has been compressed to a few hours, "the company now expects her to complete more work, and as a result, she's busier than before."

SemiAnalysis points out that the token overspend incidents at Uber, Meta, and elsewhere are fundamentally due to misaligned incentives and lax oversight, not a lack of high-ROI use cases. Despite massive layoffs at Amazon, the efficiency gains from AI tools are enabling the company to hire new employees at a faster pace—this is the clearest illustration of AI as a "human leverage" tool.

Risk Warning and Disclaimer

        Market risk: Investment requires caution. This article does not constitute personal investment advice, nor does it consider the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their particular circumstances. Investment based on this content is at your own risk.