# IBM's CEO Says Rationing AI Tokens Is a Slow Way to Lose. His Fix Is Making Inference 7 Times Cheaper > Arvind Krishna argues that companies capping AI token usage to control costs are competing from a losing position, and that caching and engineering improvements, the kind that let IBM handle seven times more work on the same hardware, are the real answer to AI's rising inference bill. - Source: Continental - Canonical URL: https://continental.today/article/ibm-s-ceo-says-rationing-ai-tokens-is-a-slow-way-to-lose-his-fix-is-making-inference-7-times-cheaper - Author: staff - Section: Business - Published: 2026-11-05T18:11:46.430Z - Updated: 2026-11-05T18:11:46.430Z - Tags: IBM, Arvind Krishna, Artificial Intelligence, Inference Costs --- IBM chairman, president, and chief executive Arvind Krishna has argued that companies responding to rising AI costs by rationing token usage are making a mistake that will cost them competitively over time. In his view, capping how much employees can use AI models is a slow way to lose, since it treats AI as an expense to be contained rather than a capability to be made more efficient. Krishna's criticism is aimed at a defensive pattern he sees spreading across large companies: spending caps on AI usage, hard limits on the number of tokens employees can consume, and restricting the most capable models to a narrow group of approved users, all driven by concern over how expensive running these systems has become in production. His alternative is to attack the cost structure itself rather than access to it. Krishna points out that AI agents typically read around 600 words for every word they actually generate, and that most of that reading is repetitive, the same context and data processed again and again. Caching, which keeps previously processed data available instead of reprocessing it from scratch, is one fix IBM has used internally, and Krishna says it allowed IBM's own systems to handle seven times more work on the same hardware. He offers a concrete example of how much engineering choices can move the cost curve. For a quantum chip design task, an open-weight model produced a 52-fold speedup at an inference cost of $26, while a leading proprietary model produced a 204-fold speedup at a cost of $299, more than ten times as expensive for roughly four times the speed gain. The example is meant to show that the most capable model is not automatically the most cost-effective one for a given job. Krishna's broader argument is that companies should own the AI operating layer running underneath their applications rather than depending entirely on a vendor's platform, since ownership is what gives a company real control over its own pricing and cost structure. The stakes are rising: 2026 is the first year enterprises are expected to spend more running AI systems in production than they spend training them in the first place, making inference cost management, rather than model access, the more urgent problem for most businesses to solve. --- Originally published by Continental. Free to cite with attribution and a link to https://continental.today/article/ibm-s-ceo-says-rationing-ai-tokens-is-a-slow-way-to-lose-his-fix-is-making-inference-7-times-cheaper.