The AI industry faces an uncomfortable reality: even as models become more capable and efficient, usage costs are climbing sharply for businesses deploying advanced agents. The culprit is a phenomenon called token amplification, where complex multi-step tasks consume millions of processing units behind a single user request.
The Mechanics of Token Amplification
Token amplification arises from a fundamental limitation of current AI models: they lack persistent memory. Each interaction in a conversation requires reprocessing the entire exchange, including all prior messages, uploaded files and model responses. This means that a simple follow-up query that costs 500 tokens may trigger an additional 600 tokens because it must reprocess the previous 500. Over a long conversation, costs accumulate dramatically. For agentic tasks that involve accessing Excel spreadsheets and a CRM system to generate customer reports, the token count can explode into the millions for a single workflow.
Why This Matters
The financial implications are severe for enterprises adopting AI agents. A single report generation task might use 280,000 tokens at a cost of $1. That seems trivial until it runs every 15 minutes for a dashboard: $96 per day or nearly $2,880 monthly. For more complex reports requiring millions of tokens, monthly costs can exceed $28,800 for one department. This has forced companies like Uber and Amazon to rein in AI spending, as agentic workloads can become costlier than human employees.
For AI providers, the challenge is twofold. They must reduce per-token costs through techniques like prompt caching, model routing and batch processing while also helping clients manage token consumption. The move to usage-based pricing by Anthropic, Microsoft and others in April reflected this reality, as fixed-rate plans proved unsustainable when power users burned through vast token allocations.
A New Cost Calculus
Businesses now treat token budgets as a critical operational metric, akin to cloud compute costs. A prompt redesign can become a margin event, and a poorly designed agent loop risks running up an exorbitant bill with no output. The era of treating AI as a cheap utility is over for enterprise deployments.



