Enterprise AI spending has not materially hit a wall; token budget management is becoming the new normal

Overview

This article discusses how enterprises are moving from the tokenmaxxing phase of "consuming as many tokens as possible" to a more mature phase of token budget management. Based on conversations with more than 50 enterprise customers via Slack, phone calls, and the Databricks AI Summit, SemiAnalysis believes that market concerns about budget overruns have been amplified by the media. In reality, enterprises are still increasing AI spending, but are beginning to manage monthly bills through hard caps, soft limits, model selection, and project-based budgets.

Core views

The report's core view is that enterprise AI spending is not undergoing a systemic contraction; budgetization is simply the normalization of governance after higher adoption. Cases such as Meta and Uber reflect imbalanced incentives and insufficient oversight, not the exhaustion of high-ROI use cases. Demand from programming tools, enterprise knowledge work, TaaS/API endpoints, AWS Bedrock, and similar areas continues to drive revenue growth for AI labs, hyperscale cloud providers, and the application layer.

Analysis framework

The report combines frontline enterprise interviews, the SemiAnalysis Tokenomics Model, Ramp data, and public product documentation to assess the true intensity of enterprise AI spending. Interviews covered Fortune 500 companies as well as enterprises in aerospace and defense, pharmaceuticals, finance, cybersecurity, travel technology, HR software, online retail, and cloud services.

Methodology notes

• Demand and revenue modelTokenomics Model

  Estimation of token consumption and ARR for AI labs, hyperscale cloud providers, and the application layer

  This model is used to estimate coding-related spending, AI Labs revenue, demand for TaaS/API endpoints, and the ARR and profit margins of application-layer products such as Cursor and GitHub Copilot.

• Frontline researchEnterprise customer interviews

  Validating the tokenmaxxing and budget-management narrative through more than 50 enterprise interactions

  SemiAnalysis collected examples of enterprise budget caps, model usage, employee behavior, and ROI cases through interviews at the Databricks AI Summit, on Slack, and by phone.

• Spending distribution analysisRamp customer spending distribution

  Using percentiles to compare differences in per-employee enterprise AI spending

  Ramp data shows that top customers spend far more than the median, indicating that AI spending is highly skewed and that extreme cases in the news are not representative of typical enterprises.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• Anthropic and Claude/Claude Code

  One of the core beneficiaries of enterprise AI APIs and programming use cases

  Strengths

  High B2B mix, strong coding use cases, and enterprise customers willing to pay higher token budgets for high-ROI workflows.

  Weaknesses

  Premium models such as Opus may be downgraded by enterprises to default options such as Sonnet, and some employees may shift to alternative tools that are not counted against budget.

  Comparison

  Compared with OpenAI, the report believes Anthropic has a higher B2B mix and therefore greater exposure to enterprise coding spend.

  Risks

  Tighter budget approvals, model downgrades, pricing pressure, and stronger enterprise governance may affect the intensity of premium-model usage.

• OpenAI, Codex, and related API business

  Major beneficiary of enterprise token consumption in programming and knowledge work

  Strengths

  Products such as Codex, Cowork, and Computer are expected to continue penetrating enterprise white-collar knowledge work.

  Weaknesses

  Budget controls at some enterprises may limit unconstrained consumption, and employees may prioritize free or annual-subscription tools for drafting content.

  Comparison

  Compared with Anthropic, OpenAI has a higher B2C mix, but it still benefits from expansion in enterprise APIs and programming use cases.

  Risks

  If enterprises strictly tie AI spending to project revenue, low-ROI use cases may be compressed.

• AWS Bedrock and AI services from hyperscale cloud providers

  Indirect beneficiaries as enterprises consume model APIs through cloud platforms

  Strengths

  The report says estimated AWS Bedrock momentum this quarter is driving AWS's overall growth rate above market expectations.

  Weaknesses

  Cloud model consumption is affected by enterprise budget cycles, model pricing, and usage governance.

  Comparison

  Compared with single-model applications, cloud platforms can aggregate demand across multiple models, multiple customers, and infrastructure needs.

  Risks

  If enterprises choose open-source self-hosting or cheaper endpoints, high-margin API consumption on cloud platforms may come under pressure.

• TaaS providers such as Together, Fireworks, and Baseten

  Beneficiaries of demand for cheap tokens and open-source model API endpoints

  Strengths

  The report expects strong demand in the TaaS/API endpoint market, with combined ARR of relevant providers already exceeding $4 billion.

  Weaknesses

  Competition may be intense, and long-term margins will depend on pricing and differentiation capabilities.

  Comparison

  Compared with frontier closed-source models, TaaS providers benefit more directly from enterprise demand for low-cost, large-scale tokens.

  Risks

  Falling model prices, expansion of built-in services from cloud vendors, and customers building their own inference capabilities may compress the opportunity.

• Programming tools such as Cursor, GitHub Copilot, Claude Code, and Codex

  The core vertical application of current enterprise AI spending

  Strengths

  The report believes that more than 70% of current ARR for OpenAI and Anthropic can be attributed to coding use cases, showing that willingness to pay has already been validated in programming scenarios.

  Weaknesses

  Enterprises may set daily or monthly budgets and make premium-model usage an explicit choice rather than the default.

  Comparison

  The programming vertical has taken off ahead of other white-collar knowledge-work scenarios.

  Risks

  If developer budgets are too low or approval processes too strict, usage intensity and expansion in renewals may slow.

• Enterprise AI users such as META, UBER, WDAY, AMZN, and MSFT

  Representative samples of enterprise AI adoption and budget governance

  Strengths

  AI tools can significantly shorten hiring, report generation, data analysis, and engineering workflows, and some enterprises are willing to raise budgets in exchange for output.

  Weaknesses

  Improper incentives may lead to wasteful token consumption, while management may also underestimate the value of automation scenarios such as email.

  Comparison

  Technology-frontier enterprises and engineering departments spend significantly more than ordinary employees and slower-adopting industries such as finance.

  Risks

  Poor spending governance can cause budget waste, while overly strict governance may suppress the AI leverage of highly productive employees.

Key data

• Enterprise interview sampleMore than 50 enterprise customersInterview channels included Slack, phone calls, and the Databricks AI Summit.
• Meta tokenmaxxing caseMore than 60 trillion tokens in 30 days, with the single highest user at about 280 billion tokensThe related dashboard was shut down within two days after media coverage, and the report believes this case is not broadly representative.
• Uber budget limit$1,500 per employee per monthUber reportedly set this limit after exhausting its annual Claude Code and Codex budget within four months; exceptions require case-by-case approval.
• Enterprise monthly budget rangeAbout $250 to tens of thousands of dollarsLower-end examples include $250 at an aerospace and defense company and $500 at a pharmaceutical company; higher-end examples include about $2,000 at Workday and Stripe, with some roles allowed higher budgets.
• Ramp spending percentiles99th percentile about $90,000 per person per year, 90th percentile about $7,300, median $136The report emphasizes that Ramp's customers themselves are skewed toward the technology frontier, while ordinary Fortune 500 media-type customers are still below $100 per person.
• Anthropic Claude Code average usage$150 to $250 per developer per monthAnthropic documentation shows that only 10% of users spend more than $30 per day.
• Source of AI Labs ARRMore than 70% of current ARR for OpenAI and Anthropic can be attributed to coding use casesThe report believes Anthropic's spending is higher than OpenAI's because more than 90% of its mix is B2B.
• Scale of TaaS providersTogether, Fireworks, Baseten, and others have combined ARR of more than $4 billionThe report believes demand is growing for both frontier-model and open-source-model API endpoints.
• Travel technology company case800 engineers out of 1,500 employees, with annual AI spending slightly below $10 millionThe company switched its default Claude model from Opus to Sonnet, and most employees have a default monthly budget of $200.

Impact & implications

For investors, budget caps do not equate to collapsing demand; rather, they indicate that enterprise AI penetration has entered a governance phase. If the report's judgment is correct, Anthropic, OpenAI, AWS Bedrock, TaaS providers, and AI applications for programming and white-collar knowledge work still have room for revenue growth. The risk is that enterprises may reduce premium-model consumption through model downgrades, free Copilot substitutes, department-level approvals, and project-based budgets, but this is more likely to change the consumption mix than to end token demand growth.

Risks

• Extreme tokenmaxxing cases in media coverage may continue to influence enterprise management attitudes toward AI budgets.
• By downgrading default models, disabling advanced modes, and setting hard caps, enterprises may reduce the intensity of premium-model token consumption.
• If AI usage cannot be tied to project revenue or employee output, budget approvals may tighten further.
• Tools such as Microsoft 365 Copilot with annual subscriptions or free quotas may divert some metered token usage away from Claude, Codex, and similar products.
• The TaaS/API endpoint market may face price competition, substitution from open-source models, and pressure from built-in cloud-vendor services.

What to watch

• Whether the monthly net new growth rate of Anthropic's and OpenAI's API businesses is maintained in the second half of 2026.
• Whether enterprise budget caps shift from soft limits to stricter hard limits and project-based approvals.
• Whether downgrading behavior such as switching Claude's default model from Opus to Sonnet expands.
• Revenue growth at TaaS/API endpoint providers such as AWS Bedrock, Together, Fireworks, and Baseten.
• The speed of penetration of Cyber, Cowork, Copilot, Codex, and Computer products in white-collar knowledge work beyond programming use cases.
• Whether enterprises can use measurable ROI to prove the output gains generated by AI spending.