TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
IBM CEO Arvind Krishna told investors that open-weight AI models are now close enough to leading closed models for many enterprise tasks, while costing a fraction as much to use. He said that economics are increasing interest in on-premises AI, but businesses may need specialized infrastructure and software expertise to deploy it.
IBM CEO Arvind Krishna told investors that open-weight AI models have become effective enough for many business tasks to justify running them on company premises, at a reported 10% to 25% of frontier models’ token costs. His comments point to a growing business case for on-premises AI, though he said companies can face practical challenges in building and operating the required infrastructure.
Krishna described a sharp narrowing in the performance gap between open models and closed frontier models. In 2024, he said, a broad assessment might have given a leading model a score of 70 while an open model scored 20 or 30. By late 2025, he said, the comparison had shifted to about 70 versus 50. He characterized current open-weight models as “90 percent as effective” as the best models, without identifying a particular benchmark, model, or evaluation method.
Krishna said open-weight models cost “anywhere from 25 percent to 10 percent” of the token cost of frontier models. He argued that this trade-off suits common workloads such as summarizing email, examining contracts, answering customer calls and composing messages. Some tasks may still call for more capable models, he said, but he expects most businesses to favor the cheaper option when it performs well enough.
He linked the economics to rising interest in on-premises AI infrastructure. Krishna said many businesses could build useful systems in two or three racks, but may lack the people needed to configure GPU clusters, networking and caching. IBM plans to produce software to ease AI deployments, he said, and he expects demand for on-premises AI to support sales of its Spyre mainframe accelerator cards. The earnings-call remarks did not include specific deployment or sales forecasts.
The Cost Case for On-Prem AI
The comments offer a view of how IBM expects enterprise AI purchasing to develop: rather than sending every request to a high-cost frontier model, companies may use less expensive open-weight systems for routine work and reserve premium models for tasks that need them. If the cost and performance estimates hold for a company’s own workloads, that division could reduce ongoing AI usage costs.
Running models on company premises can also give organizations more control over where workloads and data are processed. But Krishna’s remarks highlight that owning or hosting hardware does not by itself make an AI system easy to operate. GPU placement, network performance, caching and engineering skills can affect whether an installation delivers useful response times. IBM’s opportunity, in his account, includes software and infrastructure support as well as hardware.
The figures are Krishna’s broad estimates, not a reported independent comparison across named models or tasks. Enterprises would need to test performance, operating costs and reliability against their own requirements before treating the proposed savings as a forecast.
Products Worth Considering
Jackery HomePower 3600 Plus Portable Power Station, 3584Wh, 3600W
Anker SOLIX C200 DC Portable Power Station, 192Wh, 200W
Jackery Explorer 300 Portable Power Station, 292Wh
As an affiliate, we earn on qualifying purchases.
IBM’s Earnings and AI Business
The comments came during IBM’s Q3 earnings call. The company reported quarterly revenue of $17.1 billion, up 5%. IBM’s infrastructure revenue rose 5% year over year to $3.7 billion, while software revenue grew 9% to $7.8 billion. Red Hat revenue increased 11%, and IBM said annual recurring revenue at Red Hat rose 30%.
Krishna also discussed IBM Z systems. He said the current-generation z17 mainframes had generated $1 billion more revenue than any prior model and described the launch of new rackable models as the strongest midrange launch in nearly two decades. Mainframe revenue nevertheless fell 20%, which IBM attributed to the z17 launch having taken place about 18 months earlier.
IBM’s remarks place its AI argument alongside a broader enterprise infrastructure business. The company is positioning on-premises deployments as one response to the cost of repeated model use, while its results show different growth rates across infrastructure and software. The earnings figures do not establish how much of IBM’s revenue or future growth will come from open-weight AI deployments.
What the Cost Estimates Leave Open
IBM did not identify the open-weight or frontier models behind Krishna’s estimates, or explain which benchmarks, prompts, hardware, or inference settings were used. The statement that open models are “90 percent as effective” therefore remains a high-level assessment attributed to the CEO, not a universal measurement. Performance can vary by model and task, and the remarks do not establish that every open-weight model will match a frontier system on enterprise work.
The 10% to 25% cost range refers to token costs. Krishna did not provide a full comparison of total ownership costs, including hardware purchases, electricity, staffing, maintenance, software and model updates. He also did not quantify expected demand for on-premises systems or forecast Spyre card sales tied to that demand.
IBM’s planned deployment software was not described in detail during the comments summarized in the source report. Its release timing, capabilities and pricing remain unclear, as does how many enterprises will choose to operate models themselves rather than use hosted services.
IBM’s Next AI Deployment Steps
Krishna said IBM will produce software intended to make AI deployments easier, but did not give a product name, release date or technical specification in the reported remarks. Those details will help show how IBM plans to address the engineering and operations barriers he identified.
For customers, the next practical test is whether lower token costs translate into lower total costs while meeting the performance, privacy and reliability needs of specific workloads. Further disclosures from IBM about its software, Spyre accelerator demand and customer deployments could clarify how much of the interest Krishna described is translating into purchases.
The earnings call provides IBM’s current case for on-premises AI, but not independent validation of its model comparisons. More detailed benchmarks and deployment figures would help establish how broadly the stated cost-performance balance applies.
Key Questions
What did IBM’s CEO say about open-weight models?
Arvind Krishna said open-weight models have become close enough to frontier models for many enterprise tasks and may cost 10% to 25% as much per token, based on his estimate on IBM’s Q3 earnings call.
What does “90 percent as effective” mean?
It is Krishna’s broad characterization of current open-weight models compared with the best frontier models. IBM did not provide a named benchmark, model list or evaluation method in the reported comments, so the figure should not be read as a result for every model or task.
Why might a company run AI on-premises?
Krishna argued that lower token costs can make on-premises use attractive for routine work. Running systems locally may also give a company more control over its deployment, but it can require specialized hardware and staff to manage networking, GPU clusters and performance.
What remains unknown about IBM’s AI plans?
IBM did not give details on the software it plans to create, including its timing or features. The company also did not quantify how many customers are moving on-premises or how much that demand may contribute to Spyre accelerator sales.
Source: hn
