IBM has launched its homegrown Spyre XPU accelerators and strengthened its AI strategy through a new partnership with Anthropic, marking a significant step in its enterprise AI ambitions. Though IBM missed the initial wave of large model development, the company is leveraging decades of expertise in high-performance computing, system architecture, and enterprise software to position itself as a key player in the generative AI transformation of businesses. At its techExchange2025 developer conference, IBM unveiled the Spyre accelerators—custom-designed AI chips developed by IBM Research. These XPU cards will ship in eight-card bundles and are designed to integrate with IBM’s Power Systems and System z mainframes. The first deployment begins on October 28 for System z mainframes, with Power Systems support launching on December 12. Each bundle includes Red Hat Enterprise Linux, the RHEL.AI Inference Server, and future integration with OpenShift.AI and Watsonx.data governance tools by Q1 2026. The Spyre XPU delivers over 2.4 petaops of performance at FP16 precision, with up to 1 TB of shared memory and 1.6 TB/sec memory bandwidth when eight cards are combined into a virtual unit. It supports multiple data types including INT4, INT8, FP8, and FP16, with performance scaling as precision decreases. A key technical advantage is live migration of inference workloads between CPUs and Spyre accelerators on Power Systems—a capability not available with Nvidia or AMD GPUs. IBM also introduced Project Bob, a new AI-powered integrated development environment (IDE) designed to modernize legacy enterprise applications. Built on a fusion of Anthropic’s Claude Sonnet 4.5, IBM’s Granite models, Meta’s Llama 3 70B, and Mistral models, Project Bob enables developers to generate, analyze, and refactor code across platforms, including COBOL and RPG on mainframes. Internal testing at IBM showed that 6,000 developers used the tool over a four-month period, with half using it daily and 75% using it at least every two days. IBM claims a 45% average productivity gain among its own developers. The partnership with Anthropic, announced by IBM’s Dinesh Nirmal and Anthropic’s Dario Amodei, appears to be a strategic alignment rather than a financial investment. IBM is likely using its enterprise reach to distribute Anthropic’s models—especially Claude Sonnet 4.5, widely regarded as superior for code generation—through its Watsonx platform and tools. This allows IBM to offer pre-integrated, trusted AI solutions without building models from scratch. IBM is also rolling out a suite of AI agents and services that run on Spyre accelerators. These include cross-industry assistants for IT operations, security, supply chain forecasting, and code modernization, as well as industry-specific agents for finance, healthcare, insurance, and government. Prebuilt AI services handle vector database management, model serving, document summarization, natural language to SQL translation, and data tagging. A major differentiator is Spyre’s ability to migrate inference tasks live between CPUs and accelerators—critical for maintaining uptime in mission-critical mainframe environments. This capability gives IBM a unique edge over GPU-based systems. While pricing details remain under wraps, IBM says the bundles will be competitively priced and appeals to early adopters. As enterprises seek to modernize legacy systems without starting from scratch, IBM’s blend of hardware, software, and AI services positions it to capture value in the enterprise AI transition—especially in sectors where mainframes and Power Systems remain dominant.