IBM CEO says open-weight models deliver 90% of frontier performance at a tenth of the cost

Open weights have closed the gap, says IBM chief

On IBM's third-quarter earnings call, CEO Arvind Krishna told investors that open-weight models have become the sensible default for enterprise AI running on company hardware.

According to The Register, Krishna walked through how quickly the quality gap has narrowed. In 2024, he said, broad-based benchmarks told a lopsided story: a top closed model might score around 70 while an open alternative scored 20 or 30. By late 2025, the open contenders were reaching around 50 on the same scale.

The economics have moved even faster. Krishna said open-weight models currently cost roughly 10 to 25 percent of what frontier models charge per token, while being about 90 percent as effective. His summary of the pitch: "90 percent of the performance at ten percent of the price," delivered in just two or three racks of equipment.

Most enterprise tasks don't need the frontier

Krishna framed the decision for buyers in blunt terms: pay ten times more for the absolute best model, or accept something nearly as good for a fraction of the cost.

He acknowledged that some tasks genuinely require frontier-level capability. But he argued that the bulk of everyday enterprise work — summarising an email, reviewing a contract, handling a customer call, drafting a reply — performs just as well on the cheaper option, which is why he expects most businesses to decline the premium.

The bottleneck is deployment, not model quality

The Register reports that Krishna sees booming interest in on-prem AI infrastructure built on these economics. Most organisations, he believes, need only two or three racks to run useful AI workloads.

The hard part is operating them. He asked where enterprises would find kernel engineers, people to manage the underlying network, specialists to keep GPUs clustered in close network proximity, and engineers to cache results so responses stay fast. Beyond promising that IBM will ship software to ease AI deployments, he offered no detailed answer — but he did predict that on-prem demand will lift sales of IBM's Spyre mainframe accelerator cards.

The quarter behind the remarks

The comments came alongside mixed financial results. Total quarterly revenue was $17.1 billion, up five percent year over year. Infrastructure revenue grew five percent to $3.7 billion — growth The Register notes is slower than that recorded by Dell, HPE and Lenovo.

Software did better, rising nine percent to $7.8 billion. Red Hat grew eleven percent and increased annual recurring revenue by 30 percent. Its virtualization products, launched in 2024, have now won $800 million in deals with a further $750 million pipeline — a result IBM links partly to the possibility that VMware customers are looking to move.

On the mainframe side, Krishna said current-generation z17 systems have generated $1 billion more revenue than any previous model, and that the new rackable machines delivered the strongest midrange launch for IBM Z in nearly two decades. Mainframe revenue still fell 20 percent, which IBM expects given the z17 launched around 18 months ago. The company's share price ended down a point or two in after-hours trading.

Separately, The Register reports that IBM Cloud suffered a severity-one incident: "emergency maintenance" left users unable to provision new resources or virtual private clouds for 90 minutes. Classic Cloud users — the service IBM acquired with SoftLayer in 2013 — were told provisioning would be restored by 08:00 UTC Thursday, implying an incident lasting at least 16 hours. The work began hours after critical PHP bugs were disclosed; The Register asked IBM whether the two are connected and had not received confirmation.

Why it matters

When the CEO of a major enterprise technology vendor declares that open-weight models handle the vast majority of business tasks at a tenth of the cost of frontier APIs, it is a significant market signal. It puts pricing pressure on closed-model providers, accelerates the shift of enterprise AI workloads onto on-premises hardware, and relocates the industry's bottleneck from model quality to operational talent — the kernel, networking and clustering skills needed to run GPU racks well.

That bottleneck is exactly where IBM intends to sell: deployment software and hardware such as the Spyre accelerator cards. If Krishna's numbers hold up across the market, the next phase of enterprise AI competition may be fought over who makes those two or three racks easy to run, not over who has the single best model.