# IBM CEO: rationing AI is a slow way to lose

> Arvind Krishna argues that capping AI spending kills productivity. The real answer is engineering efficiency, open models, and owning your own cost layer.
- Title: IBM CEO: rationing AI is a slow way to lose · Fortune
- Summary: Arvind Krishna argues that capping AI spending kills productivity. The real answer is engineering efficiency, open models, and owning your own cost layer.
- Keywords: arvind krishna, ibm, ai inference costs, ai operating layer, open-source models, enterprise ai, token rationing, ai caching, cost optimization, Fortune
- Source: Fortune — https://fortune.com/2026/11/05/ibm-arvind-krishna-ai-inference-cost-curve-rationing-tokens
- Author: Arvind Krishna
- Read time: 2 min
- Topics: arvind krishna, ibm, ai inference costs, ai operating layer, open-source models, enterprise ai, token rationing, ai caching, cost optimization
## Running AI now costs more than training it
For the first time this year, the world will spend more money running AI systems than training them. Running costs are an operating expense that lands on every company's P and L every day.
## Companies are rationing AI to control costs
Facing production costs, companies are capping tokens, putting spending meters on screens, and reserving capable models for approved users. IBM CEO Arvind Krishna says this approach is counterproductive.

> Rationing the most productive technology you have ever put in your people's hands is a slow way to lose.
## Three questions determine AI economics
Krishna identifies what drives cost, who sets the price, and what companies actually spend. Engineering efficiency, not model capability, is the key to affordability at scale.
## AI agents reread the same data repeatedly
AI agents writing software read roughly six hundred words for every one they write. Nearly all that reading is repetitive, meaning companies pay to reread their own documents again and again.
## Caching cuts AI costs dramatically
IBM, Red Hat and Google built an open-source system that keeps what AI has already read and sends requests to machines holding it. On IBM systems, ninety-eight percent of needed data was already cached.
## Caching and optimization multiply capacity
With caching and parallel processing optimizations, the same hardware handled seven times more work. The software is open source and available to any company today.
## Open models offer cost control
Open-weight models run on infrastructure companies control, giving them ownership of cost structure rather than relying on vendor discounts. IBM ran an open-weight model powering three thousand AI agents simultaneously.
## Open models cost far less for some tasks
IBM's open model delivered a fifty-two-fold design speedup for twenty-six dollars in inference costs. A leading proprietary model produced a two hundred four-fold speedup for two hundred ninety-nine dollars, over ten times the cost.
## AI breaks traditional cost tracking
The same AI request might cost two cents one time and two dollars the next because the system takes different routes. Without control, the machine decides what work costs, not your company.
## Own your AI operating layer
An AI operating layer connects systems, governs model access, decides where workloads run, and keeps regulatory records. Owning this layer means owning the economics of AI and bending the cost curve yourself.
## Key takeaway

IBM CEO Arvind Krishna warns that rationing AI tokens to control costs is self-defeating; companies should instead invest in engineering efficiency and infrastructure control to manage expenses at scale.