IBM CEO: Open Weight Models Are Now Irresistibly Good

IBM CEO Arvind Krishna, speaking on the company's Q3 earnings call, told investors something that would have been controversial a year ago: open weight AI models now make brilliant sense for enterprise on-prem use. His argument comes down to simple arithmetic [1].

The collapsing gap

Krishna walked investors through the trajectory. In 2024, closed frontier models were plainly superior. "If you ran some assessment, some benchmark, and it was broad-based, not narrow, the top model might score a 70, and the open model might score a 20 or a 30." By late 2025, that gap had shrunk to 70 versus 50 [2].

Today, Krishna says open weight models deliver "anywhere from 25 percent to 10 percent of the token cost of the frontier models. That is very attractive given that they're now 90 percent as effective" [3].

The math is brutal in its simplicity. Would you rather pay 10 percent of the cost for something 90 percent as good as the best available option, or pay 10 times more for that last 10 percent? For the vast majority of enterprise tasks, summarizing emails, examining contracts, answering customer calls, composing responses, the cheaper option is more than sufficient [4].

On-prem is back

Krishna said those numbers are driving booming interest in on-prem AI infrastructure. His claim is that most businesses can build useful AI infrastructure in just two or three racks of hardware. That is a remarkably small footprint compared to the GPU clusters that have dominated AI infrastructure discussions [5].

But he also acknowledged the catch. Enterprises struggle with the practical reality of running AI on their own metal. "Where are my kernel engineers? Where are the people who worry about the underlying network? Who can make sure that the model is clustered on GPUs that are close to each other in network proximity? Who can make sure that you're doing the right caching of results so you get an incredible response?" [6].

His answer: IBM will produce software to ease AI deployments. And he predicted that on-prem AI demand will spur sales of IBM's Spyre mainframe accelerator cards, betting that the mainframe, of all things, will find new life as an AI inference platform [7].

The mainframe angle

This is the part that makes the story genuinely interesting rather than another "open models are catching up" narrative. IBM is not just observing the trend toward on-prem AI. They are positioning the mainframe as a deployment target for it.

The current z17 mainframe has delivered $1 billion more revenue than any previous model. The new rackable variants had "the strongest midrange launch in IBM Z in nearly two decades." Mainframe revenue still fell 20 percent year over year, which IBM expects given it is 18 months since the z17 launched [8].

But if open weight models really can run effectively in a few racks, and if IBM can package the deployment software so enterprises do not need a team of kernel engineers, then the mainframe's reputation for reliability, security, and predictable performance becomes a genuine selling point for AI inference. The same machine that runs your bank's transaction processing could also run your contract analysis [9].

The numbers behind the quarter

IBM's Q3 results provide context for why Krishna is leaning into this narrative. Total quarterly revenue hit $17.1 billion, up 5 percent. Software was the standout, growing 9 percent to $7.8 billion. Red Hat grew 11 percent with annual recurring revenue up 30 percent. Red Hat's virtualization products have won $800 million in deals with a $750 million pipeline, a solid result for a product line launched in 2024, especially given the possibility that VMware refugees are looking for exits [10].

Infrastructure revenue grew 5 percent to $3.7 billion, much slower than Dell, HPE, or Lenovo. IBM's share price wobbled in after-hours trading and ended down a point or two [11].

Why this matters

The AI industry has spent two years operating on the assumption that frontier model quality requires frontier model pricing, and that open alternatives are a distant second. Krishna is now publicly telling investors that assumption is outdated. Open weight models are not just cheaper. They are close enough that for most enterprise workloads, the quality difference is irrelevant [12].

If he is right, the implications are significant. Cloud AI providers face margin pressure as enterprises discover they can run good-enough models on their own hardware. GPU vendors face questions about whether their premium pricing is justified for inference workloads that do not require the absolute best model. And open source AI, already accelerating, gets a further boost from the explicit endorsement of a major enterprise vendor [13].

The fact that IBM is betting on mainframes as an AI platform is the kind of plot twist that makes this story worth following. The mainframe was supposed to be a relic. Instead, it might turn out to be the ideal on-prem AI appliance: reliable, secure, already installed in the data centers of the exact enterprises that need it most [14].