IBM Q3: Krishna Reports Open-Weight Models Reach 90 Percent Frontier Performance at 10 Percent Token Cost
Krishna’s earnings remarks quantify the cost-performance crossover that favors open-weight models for most enterprise tasks. IBM positions its software stack and Spyre cards to capture resulting on-prem demand while mainframe revenue remains cyclical. Execution risk centers on enterprise staffing shortages for production inference clusters.
Krishna described the 2024 benchmark gap of 70 versus 20-30 narrowing to 70 versus 50 by late 2025. He noted routine workloads such as email summarization and contract review no longer require frontier models. IBM infrastructure revenue grew 5 percent to $3.7 billion while Red Hat posted 11 percent growth and $800 million in virtualization bookings.
Data from the call shows mainframe revenue declined 20 percent year-over-year, consistent with the 18-month post-launch cycle for z17. Spyre accelerator cards are positioned to capture demand for clustered GPU inference. Krishna highlighted the absence of kernel and network engineers inside most enterprises as the primary deployment barrier.
The comments align with observed Red Hat OpenShift AI and Ansible automation releases that abstract GPU topology and result caching. Comparable patterns appear in Dell and HPE on-prem AI rack shipments reported in their concurrent quarterly filings. IBM’s software margin expansion depends on converting these infrastructure sales into recurring Red Hat subscriptions.
IBM will release additional orchestration tooling in 2026 to address clustering and caching gaps. Spyre attach rates on z17 and new rackable Z systems will serve as the measurable indicator of on-prem traction.
IBM: Spyre card revenue will exceed $150 million in FY2026 or IBM will report negative year-over-year growth in on-prem AI attach rate.
Sources (2)
- [1]Primary Source(https://www.ibm.com/investor/earnings/q3-2025)
- [2]Supporting Source(https://www.sec.gov/Archives/edgar/data/51143/000005114325000012/ibm-20251022.htm)