At Hot Chips 2026, IBM unveiled the successor to its z17 mainframe processor — the first commercial chip to execute Arm and IBM Z instructions on the same cores. It answers a gap that for years cut mainframes off from modern AI software.
Key takeaways
- Each of the 11 cores switches between Arm and Z modes in nanoseconds, without emulation.
- The chip is built on a 2nm process, with a base clock above 5.7 GHz and 36MB of L2 cache per core.
- A next-gen Spyre card packs 16 active AI cores and 96GB of HBM at about 4TB/s for LLM inference.
- Arm Linux binaries are designed to run 100% binary compatible, alongside z/OS in separate logical partitions.
- Launch is expected around 2027–2028, in the next system after the z17 (which shipped in Q2 2025).
Two architectures on one core
The trick is that a single IBM core natively understands both instruction sets and switches between them as the KVM hypervisor dispatches each virtual machine. A core serves Arm and mainframe modes over time rather than both at once. The switching cost, the company says, amortizes to zero across millisecond windows, while isolation between z/OS and Linux partitions is preserved.
AI comes to the mainframe
The processor itself has built-in AI inference accelerators, for example to detect fraud during transactions. The heavy lifting, though, falls to a new generation of the Spyre card: 16 active AI cores plus a spare, 96GB of HBM?HBM: high-bandwidth memory placed close to the compute die for very fast data access at about 4TB/s and PCIe Gen6. That setup is meant to run large language models (LLM) for agentic tasks, document understanding and insurance adjudication — right next to mission-critical transactions.
As technology enthusiasts on both sides, we're really excited about being one of the most powerful commercially available dual-architecture processors.
Tina Tarquinio, Chief Product Officer, IBM Z and LinuxONE.
Why pair Arm with Z
The reason is the ecosystem. Roughly 22 million developers now build for Arm, and Arm-native AI frameworks like PyTorch and ONNX Runtime dominate enterprise deployments. Running that code on a mainframe previously required porting or emulation. The new chip removes the barrier: the same server, scaling to hundreds of cores and tens of terabytes of memory, handles legacy transaction apps and modern AI tooling at once.
Why it matters
Mainframes still run the transaction core of banking and insurance, but they were losing new AI workloads to the cloud. Running Arm code and LLM inference on the same resilient, regulated hardware keeps sensitive data close to the model instead of shipping it out. For IBM, it is a way to keep the mainframe relevant in the age of AI agents rather than a relic for overnight batch jobs.
What's next
- The chip is set to ship in the next IBM Z system, expected around 2027–2028 on a three-year cadence after the z17.
- The new Spyre card with PCIe Gen6 targets larger models — real performance will only be confirmed by production deployments.
- The claimed 100% Arm Linux binary compatibility is the key test — it determines whether developers actually move AI tooling onto the mainframe.
Sources
- VentureBeat — IBM's next-gen mainframe chip is the first to run Arm and Z workloads on the same cores
- Moor Insights & Strategy — IBM Puts Arm Inside The Mainframe Core
