IBM CEO Arvind Krishna says rationing AI is a losing strategy

For years, the big AI expense was building the models. In 2026, the bill for actually using them is expected to be bigger.

IBM CEO Arvind Krishna, writing in Fortune, argues that companies are responding to that shift in the wrong way. Their instinct is to ration AI. He calls that a losing strategy.

The inference bill comes due

Global spending on inference, the cost of running AI systems once they are deployed, is expected to reach $23.3 billion in 2026. Training spend is projected at $19 billion.

That would mark the first time running AI costs more than building it.

Krishna frames the distinction in accounting terms. Training was a capital bet made by a small group of companies. Inference is an operating expense, and it shows up on the books of every business that uses the technology, every single day the system stays switched on.

Why rationing feels right but isn’t

So companies have started capping tokens, the units of text that AI models process and bill by. Some have put spending meters on employee screens, and others reserve their most powerful models for a short list of approved users.

Krishna says he understands the impulse. He has long pushed companies to pick three or four focused AI projects and fund them properly, rather than spraying money across dozens of experiments.

Krishna’s alternative is disciplined, value-driven adoption. The idea is to manage costs by choosing the right technology, not by restricting who gets to use it.

The open-weight math

His preferred lever is open-weight models, AI systems whose underlying parameters are publicly released so companies can run and customize them on their own infrastructure. According to Krishna, these models deliver roughly 90% of the performance of frontier models.

They do so at 10-25% of the cost per token, by his estimate.

This is the core of his argument against rationing. Rather than limiting access to an expensive model, companies can widen access to a cheaper one and save the premium options for work that truly needs them.

Krishna also favors hybrid-cloud setups and targeted enterprise deployments over sweeping infrastructure bets. That fits IBM’s long-running pitch to corporate customers, which is worth keeping in mind when weighing the advice.

The $8 trillion question

Krishna’s skepticism extends well beyond individual corporate budgets. He has pointed to the scale of AI data-center commitments announced by hyperscalers, the giant cloud providers that build and rent out computing capacity.

By his estimate, those announced investments total about $8 trillion. Servicing the interest on that sum could require $800 billion in annual profits, he has said.

He is also notably unenthusiastic about the industry’s grandest ambition. Krishna puts the odds of reaching artificial general intelligence through current large language model approaches at somewhere between 0% and 1%.

What this means for enterprises and investors

The shift from training to inference changes who carries the cost of AI. When training dominated, the financial risk sat mostly with the handful of firms building frontier models. Now it spreads to every company running AI in production, which means AI budgets will face the same scrutiny as payroll and office leases.

For investors, the more consequential piece may be his data-center math. An $8 trillion commitment carrying a potential $800 billion annual interest burden invites closer questions about how quickly those investments can generate returns.

There is a tension worth watching here. Hyperscalers need heavy, sustained AI usage to pay for their buildouts, while enterprise customers are actively hunting for ways to pay less per unit of that usage.

For business leaders, Krishna’s message comes down to a choice between two kinds of cost control. One limits how much AI people can use, and the other lowers what each unit of AI costs.