Executives told CNBC that companies are scrutinizing artificial-intelligence costs more closely after encouraging broad use, with finance chiefs slowing spending, seeking value and considering cheaper open models or different systems for workloads that do not require the most advanced frontier tools. [1]

The shift brings the Fed's account of AI costs preceding uncertain productivity gains inside the company, where electricity and chips become model bills but a return appears only when a defined task produces more speed, quality, savings or revenue after implementation and review costs.

CNBC's evidence remains executive testimony: Nebius calls the change rationalization, and Cerebras expects workloads to migrate among models, but the article supplies no comparable before-and-after task table, adoption sample, error rate, labor denominator, audited saving or revenue attribution. [1]

No qualifying X status survived the recorded CNBC, Nebius and Reuters searches, so token-volume boasts and failure stories cannot be treated as a measured enterprise consensus, while an empty X stack also cannot prove that finance chiefs have rejected the technology.

The honest return calculation must name the task, model, period, full cost, human review, correction burden and output measure for each team and workload inside the company under a comparable accounting period that management can defend publicly, because slower spending does not prove failure, continued spending does not prove value and an executive anecdote cannot substitute for audited return on investment.

-- THEO KAPLAN, San Francisco