IBM CEO says open-weight AI models now make sense on-premises

IBM's top executive is telling business customers that they may not need the most expensive AI models for most of their work. Arvind Krishna, IBM's CEO, used the company's third-quarter earnings call to argue that open-weight models are now a sensible pick for companies that want to run AI on their own premises, The Register reported.

His pitch is about price as much as quality. Krishna said open-weight models are 90 percent as effective as frontier models while their token cost runs somewhere between 25 percent and 10 percent of what frontier models charge, according to The Register.

Key facts

• Krishna told investors that open-weight models now make sense for on-premises use by enterprises.
• He put open-weight models at 90 percent of frontier-model effectiveness for 10 to 25 percent of the token cost.
• He said most businesses can build useful AI infrastructure in two or three racks, though many will struggle to get it working.
• IBM posted third-quarter 2026 revenue of $17.1 billion, a 5 percent increase.
• CNBC reported that IBM beat consensus estimates for adjusted earnings and revenue, yet its shares slipped in after-hours trading.

Background

Krishna described how quickly the gap between open and closed AI models has narrowed. He said that in 2024 a leading closed model might post a score of 70 on a broad benchmark, while an open model would land at 20 or 30, The Register reported.

By late 2025, he said, that spread had started to tighten to about 70 for closed models against 50 for open ones, according to The Register. Today, in his telling, open-weight models sit at 90 percent of frontier performance.

He framed the trade-off as a simple question for buyers. Krishna asked, "Would you like something that's 10 percent of the cost" when the lower-priced option can do most of the same work.

The CEO's comments came alongside a quarter of mixed financial signals. IBM said third-quarter revenue grew 5 percent, or 6 percent at constant currency, to $17.1 billion.

Software revenue reached $7.8 billion, up 8 percent, and the Hybrid Cloud business that includes Red Hat grew 11 percent. Infrastructure revenue rose 4 percent to $3.7 billion, but IBM Z revenue dropped 20 percent. Consulting revenue held flat at $5.3 billion compared with a year earlier.

IBM also said it now expects full-year free cash flow to grow by more than $1 billion over last year.

Results against analyst forecasts were mixed. CNBC reported that adjusted earnings of $2.95 per share topped the $2.86 consensus from LSEG, and revenue of $17.10 billion beat the $16.82 billion analysts expected. Even so, the adjusted gross margin of 58.5% came in below the 59.1% StreetAccount consensus, CNBC reported.

IBM shares slipped in extended trading after the results, and CNBC said the stock was down 15% as of that day's close, while the S&P 500 had gained 13%.

AI is already a large part of IBM's services work. CFO Jim Kavanaugh said AI accounted for about 30% of consulting revenue in the quarter and 35% of the consulting backlog, CNBC reported.

Red Hat is another piece of the on-premises story. The Register reported that Red Hat's virtualization products have landed $800 million in deals, with another $750 million in the pipeline.

What it means for you

For companies paying for AI by the token, Krishna's argument is that the most powerful model is often more than the job requires. He said routine business work, such as summarizing an email, reviewing a contract or handling a customer call, can be done just as well by the cheaper models, according to The Register.

That matters most for firms that run large volumes of simple requests. If Krishna's cost estimate holds, moving routine work to open-weight models could cut the token bill sharply while keeping results close to what frontier models deliver.

The hardware bar may also be lower than many expect. Krishna said most businesses can stand up useful AI infrastructure in two or three racks. He also warned that many companies will have trouble getting such systems up and running, The Register reported.

There are limits to what is known. IBM has not said which specific open-weight models it has in mind.

Practical steps

• Sort your AI workload. List the tasks your team sends to AI models and flag the routine ones, such as email summaries, contract reviews and customer calls, that Krishna said cheaper models can handle.
• Compare token costs. Price the same workload on a frontier model and an open-weight model, keeping in mind Krishna's estimate that open-weight models cost 10 to 25 percent as much per token.
• Budget for setup. If you intend to run models on-premises, set aside staff time and money for outside help. Krishna warned that many firms will struggle to get the infrastructure running, The Register reported.
• Review the numbers. IBM's full results are posted in its third-quarter earnings release.

What happens next

Krishna said IBM will build software to make AI deployments easier and predicted that on-premises AI demand will drive sales of its Spyre mainframe accelerator cards, The Register reported.

IBM has not said how much on-premises AI revenue it expects. For the full year, the company is guiding to free cash flow growth of more than $1 billion.

FAQ

What did IBM's CEO say about open-weight models? Arvind Krishna said they now make sense for companies that want to run AI on-premises.

How do open-weight models compare with frontier models? Krishna said they are 90 percent as effective at 10 to 25 percent of the token cost. No independent benchmark backing that figure is cited.

Which open-weight models does IBM mean? IBM has not named the specific models it has in mind.

How much hardware does on-premises AI need? Krishna said two or three racks are enough for most businesses, but many will struggle to get them running.

How did IBM do in the third quarter? Revenue rose 5 percent to $17.1 billion, and adjusted earnings of $2.95 per share beat estimates, CNBC reported.