IBM Machine Learning Puts Watson’s Analytics Power Where the Data Lives

With the ability to process up to 2.5 billion transactions per day, IBM z Systems mainframes are data-generating powerhouses: that’s why they’re favoured by large-scale enterprises. But massive quantities of data without intelligent analysis is the equivalent of speed without control.

Why Mainframes Still Matter

Mainframes can sound like a relic, but they still sit at the centre of a lot of everyday life. When you pay with a credit card, book a flight or file an insurance claim, there’s a good chance the transaction runs on one. IBM’s mainframe line goes back to the System/360, announced in 1964, and the z Systems machines are its modern descendants.

Banks, insurers and retailers keep using them because they’re built for huge volumes of transactions with very high reliability and security. That also makes them a goldmine of data, since a record of every one of those transactions lives on the machine.

Bringing Machine Learning to the Data

Until now, companies that wanted to leverage the powerful analytics of IBM Watson had to first make the decision to move their data off-premise. Now, with IBM’s Machine Learning Platform, that compromise has been eliminated: all the necessary resources reside in the private cloud.

That matters because moving data is harder than it sounds. Copying sensitive financial or customer records out of a company’s own data centre raises security and compliance questions, takes time, and means the analysis is always working on a slightly old copy. Running the models where the data already lives avoids all three problems.

With IBM Machine Learning, data scientists can automate the creation, training, and deployment of operational analytic models that will support:

• Any language (e.g., Scala, Java, Python)
• Popular Machine Learning frameworks such as Apache SparkML, TensorFlow, and H2O
• Any transactional data type

Gone is the cost, latency, and any risk of moving data off-premise.

What Happened Next

The machine learning push turned out to be the start of a bigger shift for IBM’s mainframes. In July 2017, IBM announced a new generation, the z14, and changed the family’s name from z Systems to IBM Z, according to Wikipedia’s history of the line.

The bigger leap came in April 2022, when IBM announced the IBM z16, built around the Telum processor with an AI accelerator on the chip itself. IBM said the z16 could process 300 billion inference requests per day with one millisecond of latency. Its main example was fraud: IBM said fraud detection models were typically run on less than 10% of high-volume transactions because of latency, and the z16 was designed to let banks check for fraud during each transaction instead.

In April 2025, IBM followed with the IBM z17, powered by the Telum II processor. IBM said it could handle more than 450 billion inferencing operations a day and 50 percent more AI inference operations per day than the z16. It also announced the IBM Spyre Accelerator, a card designed to bring generative AI capabilities to the mainframe, and AI assistants for developers and operators built on watsonx.

“The industry is quickly learning that AI will only be as valuable as the infrastructure it runs on,” Ross Mauri, general manager of IBM Z and LinuxONE, said in the z17 announcement.

What It Means for Enterprises

The idea in IBM’s 2017 announcement, bringing the analysis to the data rather than the data to the analysis, has held up. It’s now built into the hardware of IBM’s newest mainframes rather than added as a separate software layer.

For companies sitting on large amounts of sensitive transaction data, the lesson is simple: before moving that data somewhere else to analyze it, ask whether the analysis could come to it instead. That choice affects cost, speed and risk, and it’s often easier to get right at the start than to change later.