Arvind Krishna delivered a concise thesis on IBM's Q3 2025 earnings call: the open-weight AI model gap has collapsed from catastrophic to negligible, and enterprises will overwhelmingly choose the cheaper option. His numbers trace a rapid convergence — open models scoring 20-30 against frontier's 70 in 2024, closing to roughly 63 against 70 by late 2025, at one-tenth to one-quarter the per-token cost. For the vast majority of enterprise tasks — email summarization, contract review, customer calls — the premium is no longer worth paying. The strategic implication Krishna draws is straightforward: if enterprises don't need frontier models, they don't need frontier cloud providers either. Two or three racks of on-prem hardware running open-weight models becomes the default architecture. IBM is positioning itself as the integration layer that makes this work — the company that solves kernel engineering, GPU proximity networking, and result caching so enterprises don't have to hire those specialists themselves. The financial results tell a mixed story. Total quarterly revenue hit $17.1 billion, up five percent year-over-year. Software grew nine percent to $7.8 billion, with Red Hat as the standout at eleven percent growth and a 30 percent increase in annual recurring revenue. Red Hat's virtualization products, launched in 2024 to capture VMware refugees, have already closed $800 million in deals with $750 million more in the pipeline. Infrastructure revenue grew a more modest five percent to $3.7 billion — notably slower than Dell, HPE, and Lenovo, all of whom are riding the AI hardware wave harder. Mainframe revenue fell 20 percent, which IBM attributes to the natural cycle 18 months after the z17 launch, though the z17 has delivered $1 billion more revenue than any prior mainframe generation. Krishna is betting that the Spyre mainframe accelerator cards will catch some of the on-prem AI demand. The honest gap in Krishna's narrative is the execution layer. He identified the exact talent shortage enterprises face — kernel engineers, network specialists, caching architects — then offered only the promise that IBM will produce software to ease deployments. No product names, no timelines, no pricing. The demand signal may be real, but the product to capture it remains vaporware on an earnings call. Meanwhile, IBM's cloud operations drew unwanted attention. A severity-one incident caused by emergency maintenance left users unable to provision new resources or virtual private clouds for 90 minutes, with Classic Cloud users facing at least 16 hours of disruption. The timing — hours after critical PHP vulnerability disclosures — raises questions about IBM's cloud operational resilience at exactly the moment the CEO is pitching infrastructure reliability. Krishna's core argument is sound economics: when the performance gap narrows enough and the cost gap remains wide enough, enterprises migrate. The question is whether IBM captures that migration or merely narrates it. The company is selling picks and shovels for an on-prem AI gold rush, but the picks aren't built yet.
