IBM CEO Arvind Krishna told investors on the Q3 earnings call that open weight models now make strong sense for on-prem enterprise AI.
Krishna said that in 2024 closed frontier models clearly beat open alternatives on broad benchmarks (e.g. top model ~70 vs open ~20–30). By late 2025 the gap narrowed (e.g. ~70 vs ~50). He said open models now cost roughly 25–10% of frontier token pricing while delivering about 90% of the effectiveness.
Enterprises face a trade-off: pay ~10× for marginal gains on tasks like email summarization, contract review, customer calls, and email composition—most workloads do not need frontier models, he argued.
Krishna reported booming interest in on-prem AI (often two or three racks) but said many firms lack kernel engineers, networking expertise, GPU clustering, and caching skills. IBM will ship software to ease deployments; he expects demand for Spyre mainframe accelerator cards.
Other Q3 highlights: z17 mainframes delivered 3.7B (slower than Dell, HPE, Lenovo). Software +9% to 800M deals with 17.1B (+5%). Share price dipped after hours.
The Register also notes a severity-one IBM Cloud incident from “emergency maintenance” blocking new Classic Cloud provisioning for at least ~16 hours, with possible relation to critical PHP bugs (IBM did not confirm).