AI is making the mainframe harder to retire, not easier

Why mainframes rock

Ross Tisnovsky, partner at Everest Group, pointed to the fact that mainframes are now focused on being systems of record rather than systems of engagement or execution, which means they do not need to innovate much. Instead, they just need to provide efficient input/output (I/O) operations at high transaction volumes.

“This is primarily due to architecture,” Tisnovsky said. “Unlike most server platforms, the mainframe offloads I/O to dedicated peripheral processors, which creates efficiency. They also do not hoard data or restrict data usage the way many SaaS [software-as-a-service] systems do. They don’t limit the ability to take massive volumes of data in and out.”

By contrast, Tisnovsky noted that SAP recently rewrote its global application programming interface (API) policy to explicitly restrict third-party, non-SAP agentic AI applications from directly interacting with SAP systems. “In my opinion, that is the main reason why there is often no business case to replace mainframes,” he said.

In fact, IBM has created data integration interfaces to help plug mainframes into modern, API-based systems. “IBM has tried a lot to keep the platform fit. We see enterprises running not just old apps but also new apps on mainframes,” Galimberti said.

Manoj Gupta, vice-president of IT at Restaurant Brands Asia, formerly known as Burger King India, added that the mainframe remains an unsung hero. “Mainframes remain strong where backward and forward compatibility is critical,” he said. “Few platforms can match their ability to run legacy applications while seamlessly supporting modern technologies, APIs and AI-driven services.”

Attempts to write from scratch, or to move massive monolithic applications off the mainframe, have been abandoned due to high cost, high failure rates and loss of quality. The conversation today is less about replacing systems and more about modernising business processes while leveraging decades of trusted infrastructure investment

Another reason for the mainframe’s continued relevance is its quantum-proof design. Galimberti noted that long before the industry was talking about quantum-era security, legacy platforms were already quantum-safe. Tisnovsky added that most modern IBM mainframes are already compliant with post-quantum cryptography (PQC).

In terms of cost, Gupta noted that while mainframes can be expensive, their total cost of ownership is highly competitive for an organisation managing massive transaction volumes. “It also provides zero downtime and operational efficiency, and outweighs infrastructure replacement costs,” he added.

There’s a hardware advantage, too, at a time when graphics processing units (GPUs) account for the bulk of AI spending. Galimberti observed that the mainframe is a purpose-built platform, with IBM owning the supply chain, so price control is better than with GPU vendors that change prices every few days.

Why mainframes struggle

While upbeat about the mainframe’s relevance, Tisnovsky is concerned that Cobol programmers are retiring, creating a massive premium compared with cloud-native engineers.

“Young computer engineers — for some reason — show a lot of resistance to learning a programming language that predates dinosaurs,” he quipped. “A platform that cannot be supported and maintained runs a real danger of becoming unstable, and that would have a direct effect on the business case to replace it.”

Then there is the elephant in the room — almost literally. Mainframes are huge commitments and IBM’s dominance means the risk of lock-in is real.

“An enterprise usually customises pieces of IT and builds them into its own form of a stack,” Galimberti said. “It’s up to that enterprise’s IT team whether the lock-in is way down this stack or somewhere up. You will still have a lock-in, but the margin of manoeuvre in the landscape always exists.”

When it comes to modernising mainframe applications, Gupta said it is important to distinguish between modernising to reduce technical debt and modernising as a strategic investment. Many organisations, he said, have invested decades in developing business-critical applications which, over time, have been modernised, optimised and integrated with modern APIs, and moved to the cloud, with containers and other AI services.

On the other hand, rewriting millions of lines of Cobol code or migrating to another platform for the sake of it provides little business value, “because migration itself brings substantial risks, including service disruptions, compliance challenges and performance uncertainties”, he added.

Rahul Rao, distinguished engineer at IBM India Systems Development Lab, added that while the tech-debt discussion remains valid, it should not be confused with platform relevance.

“Many organisations are modernising monolithic applications, development practices and data architectures to improve agility and innovation within the mainframe. In fact, attempts to write from scratch, or to move massive monolithic applications off the mainframe, have been abandoned due to high cost, high failure rates and loss of quality,” Rao said. “The conversation today is less about replacing systems and more about modernising business processes while leveraging decades of trusted infrastructure investment.”

Will AI kill the mainframe?

Some cloud suppliers such as Amazon Web Services (AWS) have been eyeing mainframe workloads, especially with the advent of AI. But ironically, enterprises tend to use AI to bring more value to the platform, Galimberti said.

“Mainframe exits do not work. You spend anywhere around $20m to $40m on the exit and similarly huge numbers towards the end of the project, but the result is more work, not less. CIOs cannot afford that failure on their résumés.

“A mainframe exit, even if successful, is not a cause for celebration but for another technology upgrade. Plus, if someone brought in mainframes decades back, they can still change the code easily for an upgrade. But in the distributed IT world, every seven years an upgrade-plus-testing cycle is only more work,” he explained.

Rao said that, if anything, AI and quantum have made mainframes more relevant. “As organisations move from AI experimentation to enterprise-scale deployment, the focus is shifting to where and how trusted data is used. For industries such as banking, insurance, healthcare and government, the most critical data and highest-value transactions continue to run on mainframes.

“Rather than moving sensitive data to external environments for insight generation through AI processing, organisations increasingly want to bring those capabilities closer to the data to enable real-time insights, lower latency and stronger governance. Modern mainframes have evolved to support exactly that, with highly performant on-chip AI acceleration,” he said.

Matta noted that the IBM z17’s dedicated AI accelerators, for instance, allow banks and insurance companies to run complex AI models — like real-time fraud detection or instant credit-risk assessment — directly on live transactions with zero latency. “It’s a paradigm shift: bringing the AI to the data, rather than moving the data to the AI,” he said.

On AI workloads, Gupta does not see mainframes competing with GPU clusters for large-scale model training. Instead, their strength lies in AI inference at the point of transaction, bringing intelligence to where the data already resides. This reduces latency, minimises data movement and supports real-time decision-making, while integrating with cloud-based AI platforms when needed.

All in all, Galimberti noted that mainframes have both a big plus and a risk. “The advantages are resilience, stability, backward compatibility and absolute fitness for mission-critical applications. The skills part is what remains a main concern for customers. IBM’s ability to fix this gap will decide the future of the platform.”