IBM CEO says open weight models are now good and cheap on-prem
Points and comments are a snapshot, not live.
IBM CEO says open weight models now deliver 90% of performance at 10% cost.
IBM CEO Arvind Krishna told investors that open weight models now deliver 90% of frontier model performance at 10-25% of the token cost, making them irresistible for on-prem enterprise use. He noted the gap has shrunk from 70 vs 20-30 in 2024 to 70 vs 50 in late 2025. Krishna predicted booming on-prem AI interest, but acknowledged enterprises lack kernel and network engineers to deploy it. IBM plans software to ease deployments. Big Blue's Q3 revenue hit $17.1B, up 5%. Red Hat grew 11%, and z17 mainframes generated $1B more revenue than any prior model.
What commenters are saying
Commenters split on whether Krishna is simply talking his book. Some dismissed his claims as self-serving "CEO said a thing" journalism, while others agreed open models are genuinely competitive for many tasks. A handful of users shared real on-prem setups: one runs a $25K 8-GPU rig, another uses 4xV100S and a MacBook Pro for local coding and agents. Several noted hardware cost and complexity remain barriers, with IBM's own gear seen as expensive. A few argued independence from providers like OpenAI and Anthropic is worth the surcharge, citing concerns over censorship and anti-competitive behavior.