IBM and Arm have introduced a plan to develop {hardware} that may run each IBM and Arm-based workloads, to let Arm software program run on IBM mainframes.
The 2 firms plan to work on three issues: constructing virtualization instruments so Arm software program can run on IBM platforms; ensuring Arm purposes meet the safety and information residency guidelines that regulated industries should comply with; and creating widespread know-how layers so enterprises have extra software program choices throughout each platforms, IBM stated in a statement.
IBM has not stated whether or not the virtualization work will occur on the hypervisor degree, by way of its current PR/SM partitioning know-how, or through containers — a query enterprise architects will want answered earlier than they’ll assess the collaboration’s sensible worth.
IBM described the hassle as serving enterprises that run regulated workloads and can’t merely transfer them to the cloud, the assertion stated.
The place Arm stands right this moment
IBM’s mainframe prospects have thus far been unable to faucet into the Arm effectivity positive aspects that cloud customers already take pleasure in, the assertion stated. Arm says near half of all compute shipped to high hyperscalers in 2025 runs on Arm chips, with AWS, Google, and Microsoft deploying their very own Arm silicon by way of Graviton, Axion, and Cobalt, respectively.
AWS says greater than half of all new CPU capability it has added for the third consecutive 12 months runs on Graviton. Independent benchmarking by Signal65 discovered Graviton4 delivered as much as 168% higher LLM inference efficiency and 220% greater price-performance than comparable AMD x86 chips.
That hole is exactly what the collaboration is meant to handle, stated Rachita Rao, senior analyst at Everest Group. “This can be a mainframe adjacency play,” she stated. “The intent is to increase IBM Z and LinuxONE environments by enabling Arm-compatible workloads to run nearer to methods of report. Whereas hyperscalers use Arm to decrease their very own inside energy prices and go financial savings to cloud-native tenants, IBM is concentrating on the sovereign and air-gapped market.”
For banks and insurers particularly, the collaboration is as a lot about folks as it’s about know-how, Rao stated. “These organizations are hesitant to vary architectures as a result of threat of breaking the ledger, however they face a shrinking pool of legacy specialists,” she stated. “It doesn’t change the procurement cycle right this moment, but it surely de-risks the long-term viability of the LinuxONE or the Z platform as a contemporary inside cloud.”
The chips behind it
The collaboration includes two {hardware} platforms constructed by IBM to deal with AI workloads at mainframe scale: the Telum II processor and the Spyre Accelerator.
The Telum II processor, introduced at Scorching Chips in August 2024, has eight cores operating at 5.5GHz, a 40% bigger on-chip cache at 360MB, a built-in AI accelerator for real-time transaction inferencing, and a brand new information processing unit for IO duties, IBM stated.
The Spyre Accelerator is now delivery as a part of the IBM z17 and LinuxONE 5 platforms. It connects through PCIe, has 32 compute cores, as much as 1TB of reminiscence per IO drawer, and attracts not more than 75 watts per card, IBM stated. The 2 chips work collectively to run ensemble AI, the place a number of AI fashions are mixed to provide extra correct outcomes, in line with IBM.
No timeline but
IBM gave no delivery date and no technical specs for the deliberate dual-architecture methods. Statements on future course “characterize targets and aims solely” and are topic to vary, the announcement stated.
Enterprises ought to plan for a three-year improvement horizon primarily based on how lengthy IBM’s earlier {hardware} cycles have taken, Rao stated. Whereas IBM revealed the Telum II and Spyre at Scorching Chips in August 2024, Spyre is barely now reaching normal availability, roughly 12 to 18 months later.
IBM can also be pursuing different AI infrastructure partnerships on the identical time. In March 2026, the corporate introduced an expanded collaboration with Nvidia at GTC 2026 masking GPU-based information analytics, doc processing, on-premises and controlled infrastructure, and consulting. IBM plans to supply Nvidia Blackwell Extremely GPUs on IBM Cloud in early Q2 2026 for large-scale AI coaching, inferencing, and reasoning, the corporate stated.
That parallel transfer alerts the place IBM sees its major AI bets, Rao stated. “IBM’s Nvidia enlargement is clearly about GPU-native analytics and AI deployments throughout cloud and on-prem regulated infrastructure,” she stated. “That tells you IBM itself will not be treating Arm compatibility as the first reply to enterprise AI scale. Massive-scale AI continues to be shifting to GPU-heavy environments. Arm compatibility may assist convey newer software stacks or companies nearer to the mainframe, but it surely is not going to redefine the first AI infrastructure technique.”
