IBM Builds a 2nm Chip That Runs Two Architectures at Once

IBM used the Hot Chips conference this week to show off a processor that does something no mainstream chip has done before. Each core can run Arm and IBM's own z/Architecture instructions at the same time, on the same silicon. The company unveiled the design on August 24 as the next generation of chip for its IBM Z and LinuxONE mainframes.

The specifications are heavy. The part is built on a 2-nanometre process, carries 11 high-performance cores and runs at a base clock above 5.7 GHz. What matters more than the raw numbers is the idea. A bank could run Arm-native Linux software right next to the z/OS workloads that have quietly handled the world's card payments and account ledgers for decades, without shuttling data between separate machines.

The AI is baked into the transaction

For an AI news site, the interesting piece is what IBM has wired in alongside those cores. The processor includes on-chip AI inference accelerators aimed at spotting fraud while a payment is actually happening, rather than flagging it in a batch job hours later. IBM also detailed a companion AI accelerator with 16 cores, HBM3e memory and up to 4 terabytes per second of bandwidth, tuned for the newer numerical formats that modern models use.

Running a fraud model inside the transaction itself is the sort of workload mainframes are built for. These systems prize reliability and encryption over flashiness, and they process a staggering share of high-value financial traffic. Bringing inference on-chip means the bank can score every transaction against a model without the latency or exposure of sending it elsewhere.

A first result from the IBM and Arm pact

The dual-architecture core is also the first concrete product to come out of the collaboration IBM and Arm struck in April 2026. Arm's instruction set now reaches almost everywhere, from phones to hyperscale data centres, and folding it into the mainframe lets IBM's most conservative customers tap the broad Arm software ecosystem without abandoning the platform they trust.

The chip news arrives during a busy stretch for unconventional silicon. Cerebras recently lashed three wafer-scale chips together to chase Nvidia on inference, and Nvidia itself built a trillion-parameter open model to seed demand for its hardware. IBM is playing a narrower game. It is not trying to win the training race. It is trying to make sure the machines that already run the financial system can think a little on their own.