IBM CEO Arvind Krishna told investors that open-weight models now make sense for enterprises that want to run AI in their own data centers.
On the third-quarter earnings call, Krishna described a fast shift. In 2024 closed frontier models were clearly ahead of open alternatives. On broad benchmarks, he said, the top model might score 70 while the open one scored 20 or 30. By late 2025 the gap had narrowed to 70 versus 50. According to the CEO, open-weight models now cost between 25% and 10% of the per-token price of frontier models while being about 90% as effective. That ratio changes the choice: summarising an email, reviewing a contract or answering a customer call rarely needs the most powerful model.
Image: alamy.com — https://www.alamy.com/photo-by-zzndzstar-maxipx-2024-9624-arvind-krishna-chairman-and-chief-executive-officer-of-ibm-at-the-grand-opening-and-ribbon-cutting-ceremony-of-the-new-ibm-new-york-flagship-office-headquarters-at-one-madison-avenue-on-september-6-2024-in-manhattan-new-york-city-nyc-image620538062.html
The reasoning fits with what AlexTech has already reported on the gap between open and closed models, now around four months, and with the move of US companies toward mixed model portfolios to contain costs.
The problem is infrastructure, not the model
Krishna argues that many companies can build useful AI infrastructure in just two or three racks, but will struggle to make it work. The questions that, in his view, trip enterprises are operational: who manages the kernels, who handles the underlying network, who ensures the model runs on GPUs close together in network proximity, who sets up result caching for fast responses. He did not have full answers, saying only that IBM will produce software to ease deployments. He also predicted that demand for on-premise AI will spur sales of Spyre, IBM's mainframe accelerator cards.
The quarter's numbers
The z17 mainframes delivered $1 billion more revenue than any previous model, yet mainframe revenue fell 20%, as IBM had expected eighteen months after the z17 launch. Infrastructure revenue grew 5% year on year to $3.7 billion, much slower than Dell, HPE and Lenovo. Software did better, up 9% to $7.8 billion. Red Hat led, growing 11% with annual recurring revenue up 30%. Its virtualisation products have already won $800 million in deals, with a further $750 million in the pipeline. Total quarterly revenue reached $17.1 billion, up 5%. The share price wobbled after hours and closed down a point or two.
The same day, a severity-one incident in IBM's cloud stopped users from provisioning new resources or virtual private clouds for 90 minutes. The fix for IBM's Classic Cloud, acquired from SoftLayer in 2013, was scheduled for 08:00 UTC Thursday.
No comments yet. Be the first!