Often seen as the mammoths of IT shops well past their time, mainframes injected with AI are finding new opportunities for the massive data residing in the belly of the beast.
"The report of my death was an exaggeration," quipped Mark Twain in 1897 upon reading a newspaper story erroneously reporting his death. If the wry humorist was working in a large IT shop today, he might say the same thing about the mainframes he was tending to.
IBM's zSeries family of mainframes, including the latest member, the z17 introduced in April 2025, still handles the bulk of the world's transactional workloads. IBM mainframes are used by more than 70% of the Fortune 500 companies, including 45 of the largest 50 banks. The z17 is purposely optimized for AI applications containing the 5 nanometer Telum II processor and Spyre AI accelerators, capable of processing 450 billion inferencing operations a day with 1 millisecond latency.
In addition to its new AI-exploitive chips, IBM in the past year debuted a handful of AI-based software designed to make its mainframes more competitive in the new age:
The Watsonx Code Assistant for Z, an embedded AI development cycle tool designed to buttress the shrinking pool of mainframe-skilled professionals.
AI capabilities integrated into its z/OS 3.1 AI System Services, designed to optimize system performance and lessen the specialized skills typically needed to better manage mainframes.
New more intelligent observability tools such as anomaly detection that use machine learning to correlate metrics across CPUs, storage and databases that can issue warnings about possible bottlenecks.
IBM's last couple of mainframe announcements keep pushing the company further into the world of AI-based inferencing.
Frank DzubeckPresident, Communications Network Architects
With the latest on-chip accelerators, the zSeries mainframes can now move beyond legacy transaction processing, which has been the foundation of its success in the Fortune 500 accounts for decades. New markets for the mainframes include real-time fraud detection, healthcare diagnostics and the secure enterprise generative AI (GenAI) businesses. The zSeries on-chip inferencing is capable of scoring 100% of transactions a day, instead of just sampling a small fraction and processing up to 300 billion requests a day. Also debuting are specialized Synthetic Data Sets that can be used to train AI for tasks such as money laundering detection without exposing customers' data.
"IBM's last couple of mainframe announcements keep pushing the company further into the world of AI-based inferencing," said Frank Dzubeck, president of consultancy Communications Network Architects. "Transaction processing will also gain significantly more intelligence that will be driven by agents. Going forward into the agentic age, [IBM's] mainframes will be further upgraded to do that [transaction processing] internally and not out in the cloud where those transactions are more vulnerable."
Separately, in late April of this year, IBM signed a deal with Arm Technologies that will result in Arm-based software along with cloud-native AI frameworks to run natively on IBM zSeries and LinuxOne mainframes. The deal will close the gap for developers who need to apply cloud-based AI models directly to the transactional data residing on a mainframe.
Distributed platforms outside of the mainframe, right down to smart phones and desktop PCs, have or will have inference processing capabilities, but those devices largely do this processing in the cloud. As developers increasingly integrate their database structures with their large language models (LLMs), however, that effort must be conducted internally where it has more layers of security. This capability should make the mainframe a more attractive AI platform for developers.
Real-time fraud detection and inferencing are among new avenues for AI-infused mainframes.
Mainframes own the data AI needs
"Here's what's in IBM's favor in these AI wars: You can't do much with AI without data. Right now, most of the data in the world isn't in LLMs; it's in the databases residing on mainframes," Dzubeck said. "They still serve as the repository for corporate information throughout the world."
Financial institutions worldwide will spend nearly $2.3 billion relying on ensuring existing classical infrastructure works well with existing and upcoming AI offerings.
After several years of decline, mainframe usage has stabilized as some corporate users increased MIPS usage to support mission-critical workloads, according to a 2025 IDC MarketScape mainframe modernization infrastructure report. The survey revealed that 88% of mainframe users said they expect to rely on mainframes for current workloads for the next five years with AI singled out as the top strategic workload driving plans for upcoming mainframe refreshes along with security enhancements. The survey also reported that 82% of respondents are planning to significantly improve data integration for AI support over the next two years.
IDC said the AI-capable z17 is responsible for IBM's zSystem division racking up its highest revenues in 20 years, noting that adoption of mainframes grew by 51% through the first quarter of 2026. Corporate users are largely using GenAI to accelerate legacy code modernization and on-premises cloud infrastructure -- not to replace mainframes with alternative platforms.
Some analysts attribute the financial success of IBM's mainframe and accompanying AI initiative more to a captive and technologically conservative user base. One reason is there's no guarantee users will achieve the same speed and reliability by moving massive amounts of mission-critical data to lower-end distributed platforms. Another reason is the risk of wasting the heavy investments made in existing mainframes and software over the decades.
"It's not about [IBM] pitching users that its hardware is better than anyone else's, that's irrelevant," said Jack Gold, president of J. Gold and Associates. "If I'm the Bank of America, for instance, and I've been running a certain application for 30 years, I'm not going anywhere else because what I have in place works. It's how I make money. Also, many users don't know how much better the new AI will make their existing app run."
Even after a successful mainframe-based AI sale, Gold added, IBM must face the pressure of whether its offering will stand up against a wide range of AI suppliers that could change users' minds. "IBM must do AI well in order to keep their installed base happy," he explained. "If they screw it up, their users are going to have to go somewhere else whether they want to or not because that's just the way the new IT operations are now and will be going into the future."
Opportunities for mainframe AI integrators
One of the richest market opportunities now open to IBM in the AI market is the integration of its mainframe-based AI with mid-level non-IBM on-premises and cloud-based platforms. Not only is the number of veteran IT professionals managing mainframes dwindling, but there are also fewer of them trained in AI to do this integration work.
It's going to take integrators with AI experience under their belts to hook users up so the AI can communicate across platforms.
Jack GoldPresident, J. Gold and Associates
To address this problem, IBM is in close alliance with Kyndryl, formerly part of IBM's services organization, and Ensono. Both service providers focus on hybrid strategies for securely connecting mainframes to public cloud AI models.
"Existing users can't hope to take a 30-year-old mainframe out of the closet and put it to work with all this new software," Gold said. "It's going to take integrators with AI experience under their belts to hook users up so the AI can communicate across platforms. I think this is where the fattest marketing opportunity lies."
IDC research on how corporate IT should work with service providers integrating AI recommended starting with a three-to-five-year managed service contract that includes staff augmentation and remote support. Users should also request in the contract that a third-party host a dedicated mainframe platform. The report also suggests using a shared mainframe as a service to achieve more cost-effective flexibility.
Beware the mainframe exit strategy
Another favorable sign for mainframe longevity was found in a June 2026 Gartner press release that raised a warning flag for users moving too quickly away from mainframes to primarily lower-end AI-enabled platforms -- otherwise known as a mainframe exit strategy. Gartner noted that GenAI has risks, even if it can offer a technically solid platform, reduce the skill barrier in software development by aiding code understanding and enable support teams to be more efficient in executing routine daily tasks.
Many mainframe exit projects can lead to costly overruns and significantly interrupt business and operational continuity, according to a Gartner. Such an approach, the study noted, is a far cry from being the magic tool that enables a clean, seamless break from the venerable host platform. The study suggested that users are better off carefully assessing their individual workloads and picking the best platform that meets those specific needs -- an approach that typically achieves a greater return across their operations.
Yet IBM unexpectedly released disappointing preliminary revenues and earnings figures for the second quarter of 2026, casting a pall over its growing mainframe and AI initiatives. IBM's preliminary revenue number of $17.2 billion fell short of analysts' expectations of $17.85 billion.
A letter to investors from IBM President and CEO Arvind Krishna stated the company's infrastructure revenue slid 7%, driven by shifts in user spending to purchase more servers, storage and memory as larger companies felt the need to acquire more compute power amid a component supply shortage.
Another contributing factor was shortfalls in the IBM zSeries segment compared to the stellar performance of that segment in last year's second quarter. "While we anticipated some supply chain- related impact in our expectations, we did not anticipate the magnitude of the capex reprioritization," Krishna wrote. He added that the company's anticipated drop in infrastructure revenue "was worse than our expectations."
While the zSeries and AI software products integrated into the mainframes underperformed, Krishna also blamed the sales teams for their failure to close several large deals before the close of the second quarter -- another contributor to the subpar financial performance. But just days later, the executive said about 30% of those deals that failed to close have now been signed and he is confident the company will close as many as 75% of them by the end of the year.
Ed Scannell is a freelance writer and journalist based in Needham, Mass. He reports on a wide range of technologies and issues related to corporate IT. He can be reached at [email protected].
