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

The thread treated the CEO’s claim less as a neutral market observation than as a pitch from a company positioned to sell the infrastructure behind on-prem AI. Several readers dismissed the headline’s evidentiary value outright: IBM benefits if enterprises conclude that local open-weight deployment is ready, so its chief executive is a particularly interested messenger. Others nevertheless argued that the underlying shift is real. One practitioner reported replacing paid hosted models with local multi-GPU deployments for months, accepting somewhat weaker output and extra correction in exchange for privacy, control, and freedom from provider refusals. Readers also pointed to capable small IBM models and fast, inexpensive open-model options as evidence that self-hosting is becoming a serious enterprise calculation. The catch, in the discussion, is that owning racks is not the same as operating an AI service well. Inference needs GPU networking, caching, monitoring, and continual tuning; commenters argued that skilled inference engineers can unlock major utilization gains, while careless agentic automation creates security risks. Skeptics further questioned whether merely adequate open models are worth the human review burden when frontier APIs let workers concentrate on the task itself. The broad view was that open weights are increasingly viable for privacy-sensitive, good-enough workloads, but not automatically cheaper, simpler, or competitive at the capability frontier.