Bernstein: Enterprise AI is entering a "Token Shock" phase, and cost and ROI may be more constraining for adoption than technology itself.

Overview

Bernstein's global software team discusses whether "Token Shock" will harm AI adoption. The report argues that free, freemium, or low-cost unlimited subscription models of ChatGPT and early generative AI tools drove rapid adoption, but as AI companies move to token, context window, API call, and agent-task-based pricing to cover inference costs and reach profitability, the enterprise market is shifting to usage-based billing. Enterprises still recognize AI's capabilities, but cost uncertainty, budget overruns, and difficulty proving ROI are making deployment more cautious.

Core views

The key view is that AI demand has not disappeared; on the contrary, rising bills suggest adoption is moving into production environments. However, the bottleneck for enterprise AI expansion is becoming economics. Agentic AI, multi-step automation, and large context windows significantly amplify token and cloud resource consumption, moving enterprises from a "use first, monetize later" phase to a "prove ROI first, then scale" phase. Enterprises will manage spend through usage quotas, budget caps, departmental cost allocation, approval workflows, and AI FinOps; vendors, by contrast, need to reduce their own costs and pricing, or some initiatives may be delayed, reduced, or cancelled.

Analysis framework

The report uses a thematic industry research approach, combining observations from software, cloud, AI labs, enterprise applications, and consulting channels, with a focus on pricing-model shifts, inference-cost structure, enterprise budget behavior, and case feedback. The argument path is: first explain why free or low-cost subscriptions drove early adoption, then why token-based billing became inevitable, then show cost shocks through enterprise case studies, and finally discuss mitigation options for both enterprises and suppliers.

Methodology notes

• Cost and business modelToken Shock

  When AI demand and token consumption grow faster than token price declines and budget capacity, enterprises face budget shocks.

  The report defines Token Shock as the phenomenon where, after AI shifts from fixed subscriptions or free use to token/usage-based billing, enterprises find real production costs materially higher than pilot expectations.

• Operational governanceAI FinOps

  Applying cloud FinOps-style management to AI usage, budgets, cost allocation, and ROI.

  Enterprises need to monitor token consumption, allocate costs by business unit, set budgets and usage limits, and continuously measure the business value generated per dollar of AI spending.

• Investment judgmentROI constraint framework

  AI spending must match measurable revenue, savings, or productivity gains.

  The report emphasizes that tokens only measure compute activity and cannot directly measure business value; if increased AI spending cannot be shown to deliver higher revenue or savings, rising costs will naturally lead to reduced deployment.

Asset mapping & comparison

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

• ADOBE INC / ADBE.US

  A related software company identified in the report's entity recognition, cited as one example of an enterprise-application vendor with freemium AI capabilities.

  Strengths

  If AI features are embedded in high-value creative workflows and can prove productivity gains, Adobe may have pricing and retention advantages.

  Weaknesses

  If customers perceive AI features as creating incremental cost without clear ROI, adoption and monetization conversion may be limited.

  Comparison

  Compared with AI labs or pure agent tools, Adobe's AI commercialization depends more on existing usage scenarios and enterprise subscription relationships.

  Risks

  Rising generative AI costs, client budget constraints, and AI features being viewed as included in existing subscriptions could all compress incremental monetization potential.

• AI coding tools, including Claude Code, Codex, Cursor, GitHub Copilot

  The report sees coding tools as one of the most visible and likely most effective generative AI use cases, and the earliest enterprise context in which Token Shock is being felt.

  Strengths

  Productivity gains are visible, and adoption by enterprise product and IT teams has been fast.

  Weaknesses

  Using multiple tools in parallel, duplicate builds, and agent task chains can make token consumption hard to control.

  Comparison

  Compared with search or content generation, ROI from coding tools is easier to perceive, but high-frequency use also makes them more prone to triggering budget caps.

  Risks

  After enterprises set per-headroom, tool-level caps, or approval workflows, growth in usage may slow.

• Enterprise AI and Agentic AI platforms

  The core asset class discussed in the report, covering search, customer service, content generation, office workflows, coding, and multi-step agent workflows.

  Strengths

  These can automate processes and improve efficiency, helping AI move from pilot to production.

  Weaknesses

  Costs are driven by tokens, API calls, context window size, inference, CPU, storage, bandwidth, and governance, making budget predictability poorer.

  Comparison

  Compared with traditional seat- or subscription-based software pricing, usage-based AI billing shifts more variable-cost risk to clients.

  Risks

  ROI is difficult to quantify, regulatory and industry governance costs are high, infrastructure constraints can be tight, and projects may be delayed or cancelled.

• AI infrastructure and inference platforms

  The report views this as a direct beneficiary of AI usage growth, while also being a source of enterprise cost pressure.

  Strengths

  Production token growth can translate into platform revenue, as shown by the Fireworks AI case where customers and revenue grew quickly.

  Weaknesses

  Greater client price sensitivity may force platforms to cut prices or help customers optimize consumption.

  Comparison

  Compared with training costs, inference and serving costs more continuously determine enterprise AI economics.

  Risks

  If token price declines lag enterprise budget pressure, tension may emerge between platform revenue growth and customer retention.

Key data

• Pylon AI bill increaseabout 3.5x to approximately $1.4MM in annualized AI billingsAfter enterprise pricing was triggered, seats no longer include tokens, and standard API rates make management focus more on usage with positive ROI.
• EY Agentic AI cost amplificationabout 30x higher cost per interactionAfter moving from simple search/summarization LLM prompts to multi-step Agent tasks, cost per interaction increased significantly.
• Uber AI coding budget2026 AI coding budget was exhausted in 4 months; set a $1,500 per-employee cap per Agentic coding toolThe report cites media coverage that Uber managed Anthropic Claude Code and Cursor budgets through per-user spending caps.
• European bank inference cost assessmentInference costs may account for about 80% of total AI model lifecycle costThe report argues that markets focus too much on training cost, while enterprise AI long-term economics increasingly depend on service and operating costs in production.
• Fireworks AI scaleabout 300,000,000,000,000 tokens/day, over 10,000 customers; annualized revenue rose from about $305m at end-2025 to about $800m in MayHigher costs can also be interpreted as evidence that production-scale adoption is expanding.
• Rating distribution disclosureOutperform 51.2%; Market-Perform/Neutral 35.8%; Underperform 13.1%Bernstein disclosed its global rating distribution as of 2026-06-30, and this is not a rating of the company discussed in this report.

Impact & implications

For software and AI industries, Token Shock may slow adoption of undifferentiated AI features and raise client demands for demonstrable ROI, cost transparency, and governance capabilities. Beneficiaries may be vendors that lower inference costs, provide cost monitoring and optimization, or tie AI features to high-value workflows; pressure points may be AI applications that depend on high usage, low pricing, or unclear ROI. For investors, it is important to monitor both AI revenue growth and client budget capacity, not just usage volume or model capability.

Risks

• Token consumption grows faster than price declines, leading to enterprise budget overruns.
• Multi-step Agentic AI workflows significantly amplify cost per task.
• Enterprises struggle to map token usage directly to revenue, cost savings, or productivity gains.
• Regulated industries require governance, auditability, explainability, and compliance controls, increasing total cost of ownership.
• Infrastructure supply constraints may raise model access costs or limit availability.
• Enterprises may control risk by applying quotas, spending caps, scaling down deployments, or canceling projects.

What to watch

• Whether enterprises can maintain acceptable ROI after moving from pilots to production.
• The adoption pace of AI FinOps, token governance, and cost visibility tools.
• Whether major AI vendors lower inference costs and pass savings through to customers.
• Whether coding tools and Agentic AI continue to remain high-growth but highly contested categories in enterprise AI budgets.
• Whether customer contracts shift from unlimited subscriptions to finer-grained usage caps and approval mechanisms.
• Whether incremental revenue from AI features at software companies can cover inference costs while sustaining margins.