US companies seek cheaper AI, FT reports

New reporting points to interest in open-weight alternatives. Earlier spending data does not establish a broad migration away from proprietary models.

The Financial Times, a business newspaper, reported on 27 September that US companies were turning toward cheaper open AI models as technology expenses increased.

Companies are looking for less expensive ways to use AI, according to the report. Open-weight models offer another option. The evidence reviewed here does not establish how much business has moved or what buyers have saved.

Why it matters: Interest in an alternative and a measured change in spending are different findings. A buyer evaluating the report needs to know whether it describes executive intentions, a limited deployment or a sustained replacement of an existing supplier.

PYMNTS, a business-news publication summarizing the FT, said mentions of open-weight or open-source models in earnings calls and investor conferences increased sixfold in August and September against the same months of 2025. It attributed that count to AlphaSense, a research platform. The underlying search dataset was not available in this review, so the figure remains unverified here.

Even if reproduced, a mention count would measure what executives discuss. It would not show the share of tasks completed with those models, the proportion of spending they receive or the quality of the resulting work. A company can talk about an evaluation before making a substantial operational change.

The distinction also matters for language. Open weights are the trained numerical parameters of a model made available for use. Downloadable weights create a deployment option, but the term alone says nothing about the buyer’s complete operating expense or whether a particular workload will perform well.

This is why the useful comparison is a successful task under specified conditions. An evaluator should decide in advance what counts as success, include corrections and retries, and apply the same standard to each option. Otherwise, a lower advertised rate can be mistaken for a lower cost of getting the work done.

There is also a timing question. A new article may bring together several developments from earlier weeks. Its publication date makes the reporting current, but it does not make every underlying observation a newly measured event or establish that separate datasets cover the same buyers.

Earlier spending evidence limits the broad claim

Ramp, a business-spending platform, published relevant counterevidence on 9 September. Its analysis put use of routing platforms associated with open models at 6.4% of AI-spending businesses in its sample and explicitly warned that those platforms also serve closed models.

Ramp’s report said cheaper standard models, rather than open-model adoption, explained its observed shift in token usage. That is older context, not a new result from the past day. It limits a sweeping interpretation of the FT story without disproving that particular businesses are exploring alternatives.

The two sources can describe different stages of the same purchasing process. Discussion may precede deployment; deployment may initially cover a narrow task. That is a possible reconciliation, not a finding demonstrated by linking the two datasets.

The practical question for a technology team is correspondingly narrow: which of its tasks can move without unacceptable changes in output quality, response time or operational work? A trial should answer that question directly instead of treating either a newspaper trend or a market-wide percentage as a substitute.

For now, the defensible conclusion is that new reporting describes interest in cheaper open models. It does not establish universal savings, the end of proprietary services or an industry-wide replacement rate. Those stronger conclusions require evidence beyond the material verified here.

Next, watch for named deployments with comparable before-and-after task costs, and for publication of the methodology behind the executive-mention count. Those details would make the reported shift easier to measure and distinguish durable adoption from an active search for alternatives.

Verification

  • PARTIALLY VERIFIED — Reporting: The FT’s original publication confirms the story’s subject; full reporting was access-restricted.
  • UNVERIFIED — Adoption scale and sixfold count: Underlying corporate records and AlphaSense data were not inspected; reported via PYMNTS. No primary dataset was retrieved to validate the count.
  • VERIFIED AS PUBLISHER-REPORTED — Counterevidence: Ramp’s 9 September analysis supplies the sample-specific percentage, proxy limitation and interpretation; this is background, not today’s news.
  • ANALYSIS — Measurement distinctions, suggested tests and possible reconciliation are editorial reasoning, not observed savings.

Glossary candidates

  • Open weights: Downloadable trained model parameters.
  • Proxy: An indirect measurement used to estimate another activity.

Cold-reader sentence: The FT reports growing corporate interest in cheaper open AI models, but available evidence does not establish the scale or savings of a broad migration.