IBM CEO says open weight models are now good and cheap on-prem

Open-weight AI models can now be run on-premise for 90% of frontier model performance at 10% of the cost.

IBM CEO Arvind Krishna announced that open-weight AI models have become a viable and cost-effective option for on-premises enterprise deployment. During IBM's Q3 earnings call, he stated that while closed frontier models were superior in 2024, the performance gap has significantly narrowed by late 2025. Krishna highlighted that open-weight models now offer 90% of the performance of top-tier models at only 10% of the token cost. He believes this economic advantage will lead most enterprises to choose open-weight solutions for common tasks such as email summarization, contract review, and customer service interactions. Krishna also noted a substantial increase in demand for on-premises AI infrastructure, though he acknowledged that many businesses will face challenges in implementation, leading IBM to develop software solutions to ease deployment. He anticipates this trend will also boost sales of IBM's z17 mainframe accelerator cards.

IBM's CEO asserts that the cost-performance ratio of open-weight AI models has reached a tipping point, making them an 'irresistible' choice for on-premises enterprise solutions. This shift is driven by a dramatic reduction in token costs, enabling companies to achieve nearly comparable results to proprietary models at a fraction of the price. The accessibility and affordability of these models are expected to democratize AI adoption within organizations, allowing them to deploy powerful AI capabilities without the prohibitive expenses typically associated with frontier models. This development directly challenges the market dominance of closed-source AI providers and signals a potential acceleration in the on-premises AI infrastructure market.

The growing viability of on-premises open-weight models presents both opportunities and challenges for the enterprise. While cost savings are a primary driver, the practical implementation of these systems requires specialized expertise in hardware, networking, and model optimization, areas where IBM plans to offer solutions. The increased demand for on-premises AI infrastructure is likely to spur innovation in specialized hardware like IBM's mainframe accelerators, potentially reshaping the competitive landscape for AI hardware vendors. Enterprises must now weigh the immediate cost benefits against the need for internal expertise or reliance on vendors for deployment and management.