IBM CEO says open weight models are now irresistibly good and cheap on-prem

IBM reported Q3 revenue of $17.1 billion, up 5% year over year; software grew 9% to $7.8 billion, while infrastructure rose 5% to $3.7 billion and mainframe revenue fell 20%. CEO Arvind Krishna said open-weight models deliver roughly 90% of frontier-model effectiveness at 10%–25% of the token cost, driving interest in on-prem AI; Red Hat grew 11%, with $800 million in virtualization deals and a further $750 million pipeline. IBM shares ended down a point or two in after-hours trading. Separately, an IBM Cloud Classic emergency-maintenance incident blocked new provisioning for 90 minutes, with IBM saying restoration of provisioning and resource management would take at least 16 hours.

Analysis

The key shift is not simply cheaper inference; it is a redistribution of AI economics from scarce frontier-model access toward deployment, integration and workload control. If open models are adequate for routine enterprise tasks, token pricing loses some pricing power, while value accrues to whoever can make secure, reliable systems work inside customer environments. That is a plausible opening for IBM’s Red Hat stack and services, but also for Dell and HPE hardware; it does not establish that IBM captures the infrastructure spend. The bottleneck Krishna identifies—specialist engineering, networking and efficient GPU utilization—could make implementation labor and total cost of ownership more decisive than model token cost. Verify customer deployments and realized economics before treating his cost/effectiveness comparisons as market-wide facts.

Near term, the infrastructure growth comparison favors monitoring Dell and HPE over assuming IBM’s on-prem thesis translates into faster IBM hardware growth. IBM’s software and Red Hat results offer a counterweight, particularly if VMware displacement converts from pipeline into booked revenue. The Classic Cloud provisioning incident is a contained service-reliability warning, not evidence of a broad IBM Cloud outage; nonetheless, it undercuts the reliability proposition that can support enterprise infrastructure decisions. Over 6–18 months, persistent open-model quality gains would pressure premium inference economics but could expand deployment demand. The contrarian risk is that cheaper tokens do not mean cheaper production AI: engineering, security, power, hardware utilization and model maintenance may absorb the savings.

Key Decisions for Investors

• No high-conviction directional trade on this report alone. Treat IBM’s open-model/on-prem thesis as a demand signal, not proof of incremental IBM revenue or margin capture.
• Watch a conditional relative-value setup: favor DELL/HPE over IBM only if subsequent results confirm stronger infrastructure growth and IBM’s hardware momentum remains subdued. Reassess if IBM infrastructure growth accelerates or Red Hat converts VMware-related pipeline into material bookings.
• For IBM, track Red Hat virtualization bookings versus its stated pipeline, software growth, and evidence that Spyre/on-prem deployments generate repeatable revenue. Failure to convert pipeline or deceleration in software growth would weaken the offset to infrastructure softness.
• Do not extrapolate the Classic Cloud incident to IBM’s whole cloud business. Revisit the reliability risk if IBM discloses broader customer impact, repeated provisioning failures, or consequences for renewals; absent that, treat it as a watch item.