IBM CEO says open weight models now 90% as effective at 10% cost

Open weight models close performance gap

During the Q3 earnings call IBM chief executive Arvind Krishna said that in 2024 closed frontier models were clearly ahead, scoring about 70 on broad benchmarks while open weight models scored 20 to 30. By late 2025 the gap had narrowed to roughly 70 versus 50. Krishna added that open weight models now cost 10 to 25 percent of the token price of frontier models while delivering about 90 percent of their effectiveness.

He argued that most enterprises will choose the cheaper option because tasks such as summarizing email, reviewing contracts, handling customer calls and composing messages can be handled equally well by the lower-cost models.

On-premises AI infrastructure demand rises

Krishna reported booming interest in on-premises AI infrastructure. He noted that a useful AI cluster can be built in two or three racks, but many organizations lack the kernel engineers and network expertise needed to place GPUs close together, manage caching and achieve fast response times. IBM plans to supply software that simplifies these deployments and expects the trend to boost sales of its Spyre mainframe accelerator cards.

Mainframe and cloud updates

The current z17 mainframe generation has generated one billion dollars more revenue than any previous model, and the new rackable models achieved the strongest midrange launch in nearly twenty years. Mainframe revenue still fell 20 percent as the product is eighteen months into its cycle. Infrastructure revenue grew five percent year over year to 3.7 billion dollars, slower than rivals Dell, HPE and Lenovo. Software revenue rose nine percent to 7.8 billion dollars, with Red Hat growing eleven percent and increasing annual recurring revenue by thirty percent. Red Hat virtualization products have secured 800 million dollars in deals with a 750 million dollar pipeline. Total quarterly revenue reached 17.1 billion dollars, up five percent, while the share price slipped a point or two after hours.

Separately, IBM Cloud suffered a severity one incident caused by emergency maintenance that prevented provisioning of new resources and virtual private clouds for ninety minutes. Classic Cloud customers were told provisioning would be restored by 08:00 UTC Thursday, meaning the outage will last at least sixteen hours. The maintenance began shortly after critical PHP bugs were disclosed, and IBM has been asked whether the patching is related.

Sources: The Register