Token Shock and Its Impact on Enterprise AI Adoption

Analysts at Bernstein describe “token shock” as the budgetary strain that arises when enterprise consumption of generative and agentic AI expands faster than the decline in per‑token pricing. To protect margins, many AI software providers are moving away from fixed, unlimited subscriptions toward usage‑based pricing models that directly reflect inference costs.

Agentic AI Amplifies Cost Pressures

A study by EY found that converting a straightforward large‑language‑model prompt into an agentic workflow—where multiple model calls, larger context windows, and iterative reasoning are required—can raise the cost per interaction by roughly thirty‑fold.

Corporate Responses to Rising Expenses

  • Uber Technologies (NYSE:UBER) exhausted its 2026 AI coding budget within four months and responded by capping spending at $1,500 per employee for each agentic coding tool.
  • Walmart (NYSE:WMT), after initially offering employees unlimited token access to an internal AI agent, has now imposed a cap on that access.
  • A major French insurer reduced its usage of Anthropic’s Claude model just two months after launch because the incurred costs exceeded its expectations.
  • HubSpot (NYSE:HUBS) has begun charging customers based on outcomes rather than token consumption, signaling a shift toward performance‑linked pricing.

Project Delays and Limited Returns

Discussions with technology and consulting providers reveal that many clients have postponed, scaled back, or cancelled at least one AI project due to cost concerns and uncertain economics. One provider estimated that fewer than 25 % of its pilot projects generated positive returns.

Cost Structure of AI Models

Inference expenses can represent about 80 % of an AI model’s total lifetime cost, shifting the financial focus from the upfront training expense to the ongoing cost of running models in production.

Strategies to Manage Inference Costs

Enterprises are increasingly reserving the most advanced models for complex tasks while routing simpler workloads to smaller, cheaper alternatives. Open‑source models and private infrastructure are gaining traction for high‑volume, predictable workloads.

Outlook

While the overall outlook remains guardedly optimistic, the pace of enterprise AI adoption may be slower than anticipated. Software vendors may respond with specialised models, improved model‑routing mechanisms, and outcome‑based pricing structures, as exemplified by HubSpot’s recent initiative.