IBM chief executive Arvind Krishna has told investors that open-weight artificial-intelligence models are becoming attractive for mainstream enterprise workloads because their performance gap with closed frontier systems has narrowed while their operating cost remains substantially lower. His comments frame on-premises AI as a growing infrastructure opportunity for IBM.
During IBM's third-quarter earnings call, Krishna described a rapid change since 2024, when he said broad evaluations showed closed models decisively ahead of open alternatives. By late 2025, that gap had contracted, and he now estimates that open-weight systems can deliver about 90 percent of frontier-model effectiveness at between 10 percent and 25 percent of the token cost. Those figures are Krishna's characterization rather than an independent benchmark.
The economic trade-off matters because many business tasks do not demand the strongest model available. Krishna cited email summarization and composition, contract examination and customer-call assistance as examples where a less expensive model may be sufficient. Under that view, enterprises would reserve costly frontier services for the smaller group of jobs where marginal capability has clear value.
IBM expects that calculation to support local AI deployments. Krishna said useful systems can often fit within two or three equipment racks, keeping company data and inference infrastructure on site. Building such a system is not automatic, however. Organizations still need expertise in GPU clustering, networking, kernels and caching to deliver dependable response times. IBM's commercial opening is to provide software that reduces that operational burden.
The chief executive also linked on-premises demand to IBM's Spyre accelerator cards for mainframes. The company's current z17 generation has produced $1 billion more revenue than any prior model, according to the earnings-call account, while new rack-mounted systems recorded IBM Z's strongest midrange launch in almost two decades. Mainframe revenue nevertheless fell 20 percent as the product cycle moved 18 months beyond the z17 launch.
Across the quarter, IBM reported total revenue of $17.1 billion, a 5 percent year-over-year increase. Infrastructure revenue also grew 5 percent to $3.7 billion, while software revenue advanced 9 percent to $7.8 billion. Red Hat grew 11 percent and increased annual recurring revenue by 30 percent. IBM said its Red Hat virtualization products had secured $800 million in deals with another $750 million in the pipeline.
Krishna's argument is therefore both a technology assessment and a sales thesis. Cheaper capable models may expand enterprise AI usage, but the complexity of operating them creates demand for packaged infrastructure and management software. Whether the quoted cost-performance relationship holds across particular workloads will depend on each organization's models, hardware and quality requirements.
