IBM CEO says open weight models are now irresistibly good and cheap on-prem
Arvind Krishna tied enterprise interest in open-weight models to control of data, deployment and inference costs.
TL;DR
- IBM CEO Arvind Krishna said open-weight models are increasingly viable for enterprise use on company-controlled infrastructure.
- He linked the shift to lower inference costs and the ability to keep logs and data within private environments.
- The remarks describe IBM’s view of a market trend, not measured adoption across all enterprises.
Krishna made the argument during IBM’s third-quarter 2026 earnings call, where he discussed the economics behind open-weight models. The Register’s report captured his case for running them on-premises, while the earnings-call transcript preserves the broader exchange with investors. [1]
The core distinction is operational control. Krishna said companies can run models on premises or in private clouds and keep control over the underlying logs and data; that is a deployment option, not a claim that every open model matches every frontier model. [2]
Cost was another part of his comparison. On the call, Krishna described open-weight inference as substantially cheaper than frontier-model use for many routine workloads, but the transcript does not establish a universal cost ratio across model sizes, hardware, utilization or service contracts. [2]
IBM sells hybrid-cloud and enterprise infrastructure products, so its CEO’s assessment also aligns with the company’s commercial position. Independent reporting confirms the remarks, but broader uptake, performance trade-offs and the total cost of operating local GPU clusters remain questions for buyers and the market. [1]
Why it matters
If enterprises shift selected workloads to private infrastructure, the economics of AI may depend as much on deployment and data control as on the largest model’s benchmark lead.
Editor's note
Based on the earnings-call remarks and independent coverage; forecasts about eventual market share are excluded.