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 can deliver about 90% of the effectiveness of frontier models at 10% to 25% of their token cost, making them attractive for many on-premise business tasks. He said enterprises are showing strong interest in on-prem AI infrastructure, while acknowledging that deploying and operating it requires specialized engineering skills.
IBM CEO Arvind Krishna told investors that open-weight AI models can deliver roughly 90% of frontier models’ effectiveness at 10% to 25% of their token cost, a trade-off he said is driving enterprise interest in running AI on-premise. His comments, made during IBM’s Q3 earnings call, frame cheaper models as a practical option for routine business tasks, while leaving the most demanding work to more expensive frontier systems.
Krishna described a rapid narrowing in the performance gap between open and closed models. In 2024, he said, a broad benchmark might put a leading model at 70 points and an open alternative at 20 or 30. By late 2025, he said, the comparison had moved closer to 70 versus 50. He characterized current open models as “90 percent as effective” as the best frontier models, but did not identify a specific benchmark, model, or evaluation method supporting that estimate.
The CEO said open-weight models cost “anywhere from 25 percent to 10 percent of the token cost of the frontier models.” He argued that this price-performance balance suits tasks such as summarizing email, reviewing contracts, responding to customer calls and drafting messages. Some workloads still require leading frontier models, he said, but most routine tasks may not justify paying substantially more.
Krishna said IBM is seeing increased interest in on-premise AI infrastructure. He suggested many businesses could build useful systems in two or three racks, but warned that operating them can be difficult. Companies need people who can manage kernel-level software, networking, GPU clustering and caching, he said. IBM plans to provide software to ease deployments, though he did not give details about the products, release dates or costs.
Lower-Cost Models Shift Enterprise Choices
If Krishna’s cost and capability estimates hold for particular workloads, enterprises could run more AI applications without paying frontier-model prices for every request. On-premise deployment may also appeal to organizations that want greater control over where models and business data are processed. The remarks point to a possible division of labor: lower-cost open models for common tasks, with more expensive systems reserved for work that needs their additional capability.
The economics do not remove the technical burden. Companies still need suitable compute, network capacity, software and staff to deploy and maintain models. Krishna’s emphasis on shortages of specialized engineering skills suggests that the cost of infrastructure and operations could affect whether the projected savings are achievable. His comments are a company executive’s assessment, not an independent comparison of model performance or total deployment costs.
Products Worth Considering
Bose QuietComfort Headphones (2nd Gen), Wireless Noise-Cancelling Over-Ear Headphones, Eucalyptus Green
Samsung 27-inch Odyssey OLED G6 QD-OLED QHD Gaming Monitor
Sony WH-1000XM4C Wireless Noise-Cancelling Headphones, Black
As an affiliate, we earn on qualifying purchases.
IBM’s Earnings and AI Infrastructure
Krishna’s comments came as IBM reported Q3 revenue of $17.1 billion, up 5% year over year. Infrastructure revenue rose 5% to $3.7 billion, while software revenue increased 9% to $7.8 billion. Red Hat grew 11%; IBM also said its virtualization products had generated $800 million in deals, with a further $750 million in the pipeline.
IBM is positioning its hardware and software for enterprise AI workloads, including through its Spyre mainframe accelerator cards, which Krishna said could benefit from on-prem demand. He also cited strong sales for newer IBM Z products. The company said mainframe revenue fell 20%, as expected about 18 months after the z17 launch. These results provide the business setting for Krishna’s remarks, but IBM did not report a quantified on-prem AI revenue figure in the supplied account.
Benchmarks and Deployment Costs Remain Open
The basis for the 90% comparison is not specified in the reported comments. Krishna did not name the models, benchmark suite, tasks, inference settings or date of testing behind the estimate. Model performance varies by task, so the figure should be understood as his broad characterization, not a verified result for every enterprise use case.
The cost comparison also refers to token costs, not the full expense of buying and operating on-premise systems. The reported account does not provide hardware prices, power use, staffing costs, expected utilization or a direct comparison with cloud services. It is also unclear which IBM software will support deployments, when it will be available, and how much implementation expertise customers will still need.
IBM’s Deployment Software and Sales
IBM’s next steps will be to develop the software Krishna said could make AI deployments easier and to pursue demand for on-prem infrastructure, including systems using Spyre accelerators. The company has not provided a launch schedule or product specifications in the reported remarks. Future earnings updates may show whether the interest Krishna described translates into measurable infrastructure sales.
For customers evaluating the approach, the practical test will be whether open models meet the accuracy and reliability requirements of specific tasks while reducing total costs after hardware and operations are included. IBM has not published those comparative results as part of the comments covered here, so the performance and savings claims remain to be assessed workload by workload.
Key Questions
What did IBM’s CEO say about open-weight models?
Arvind Krishna said open-weight models can be about 90% as effective as frontier models while costing roughly 10% to 25% as much per token, according to his comments on IBM’s Q3 earnings call.
Which business tasks did Krishna say could use cheaper models?
He cited summarizing email, examining contracts, answering customer calls and composing email as examples of tasks that may not need the most expensive models.
Does the 90% figure come from a named benchmark?
Not in the reported remarks. Krishna did not identify the models, benchmark, task mix or evaluation method behind the estimate, so it is an executive claim rather than a fully specified comparison.
Why might on-prem AI be difficult for businesses?
Krishna said deployment can require specialized skills in kernel engineering, networking, GPU clustering and caching. The reported account does not quantify how much staff or infrastructure a typical organization would need.
What is IBM planning to do next?
Krishna said IBM will produce software intended to ease AI deployments and predicted that on-prem demand could support sales of its Spyre mainframe accelerator cards. He did not give product release dates or sales forecasts in the reported comments.
Source: hn
