IBM chief executive Arvind Krishna told investors on 22 October that open-weight AI models now deliver 90 percent of frontier models’ effectiveness at 25 percent to 10 percent of their cost per token, The Register reported. Speaking on IBM’s third-quarter 2026 earnings call, he argued that the economics favour running open-weight models on companies’ own infrastructure for many everyday tasks.
Key points
- Krishna said open-weight models have reached 90 percent of frontier-model effectiveness at a substantially lower cost per token.
- He estimated that two or three racks could provide useful AI infrastructure for an average large enterprise.
- IBM’s quarterly revenue rose 5 percent to $17.1 billion; infrastructure revenue grew 5 percent to $3.7 billion.
Krishna’s comparison with frontier models
Krishna described the earlier performance difference through a hypothetical broad-based assessment. In 2024, he said, a leading frontier model might have scored 70 while an open model might have scored 20 or 30. By late 2025, he put the possible comparison at 70 against 50. His current estimate of 90 percent effectiveness is a further narrowing of that gap, rather than a score from the same assessment.
Krishna placed the earlier comparison in late 2024, Benzinga reported. It put the present cost range at 10 percent to 25 percent of the cost per token of frontier models.
Krishna told investors that some work still requires frontier-model capability. For most routine enterprise tasks, including summarising emails, examining contracts, answering customer calls and drafting messages, he said cheaper models can provide an adequate answer. That distinction matters for buyers deciding which requests warrant a more expensive model, rather than applying one choice to every task.
Two or three racks for enterprise AI
For an average large enterprise, Krishna estimated that two or three racks would be enough to begin running open-weight models. He linked interest in such installations to companies’ preference for keeping sensitive data, logs and model activity under their own control.
Operating the equipment is a separate challenge. Krishna said enterprises need people who can manage kernels and networks, place GPUs close enough together for clustered workloads and configure caching to improve response times. IBM plans software to make those deployments easier, according to The Register. Krishna also predicted that demand for on-premises AI would support sales of IBM’s Spyre accelerator cards for mainframes.
His case for local infrastructure rested on the estimated cost of running models and the resources needed to operate them. A two- or three-rack installation would still require the specialist work he described. For IBM, that puts deployment software and infrastructure alongside the cheaper models in its pitch to enterprise customers.
IBM revenue rises to $17.1 billion
IBM’s third-quarter revenue rose 5 percent from a year earlier to $17.1 billion, The Register reported. Infrastructure revenue increased 5 percent to $3.7 billion, slower growth than at Dell, HPE and Lenovo. The comparison places Krishna’s expectations for on-premises AI alongside a hardware business growing more slowly than those rivals.
Software revenue climbed 9 percent to $7.8 billion. Red Hat revenue grew 11 percent, while its annual recurring revenue increased 30 percent. IBM chief financial officer Jim Kavanaugh also reported 10 percent growth in transaction processing, driven by mainframe software, Benzinga reported.
Mainframe revenue fell 20 percent from a year earlier, a decline IBM expected at a point 18 months after the z17 launch. Krishna said the current z17 generation had brought in $1 billion more revenue than any previous model. He also described the launch of IBM’s new rackable midrange models as its strongest in that segment in nearly two decades.
Red Hat’s virtualisation products, launched in 2024, have won $800 million in deals and have a further $750 million in the pipeline. IBM sees scope for more business if VMware customers move away from Broadcom’s licensing changes.
