IBM Is Playing a Long AI Game with Its Customers (www.nextplatform.com)

IBM is quietly executing a long-term, enterprise-first AI strategy: rather than chasing the biggest consumer-facing model builders, it’s hardening its existing Power and z platforms for measured, private AI adoption across 100,000+ customers. The company has added matrix-math units to Power/z CPUs, developed a homegrown Spyre inference accelerator, and layered RHEL.AI/OpenShift.AI on its Linux/Kubernetes stacks. It’s also shipping Project Bob — a code assistant built on Anthropic’s Sonnet 4.5 — to run in Linux partitions using Spyre, replacing internally developed Granite models. Internally, IBM’s “Client Zero” deployments (sales, supply chain, HR) are projected to generate $4.5B in annualized productivity savings by 2025, while GenAI bookings since Q3 2023 exceed $9.5B. Why it matters: IBM’s play emphasizes secure, integrated on-prem and hybrid AI for enterprises that don’t want public cloud or to replatform around GPUs. Technical implications include tighter hardware-software co-design (matrix units + Spyre), middleware and orchestration tuned for enterprise workflows, and a path for inference offload that could compete with NVIDIA/AMD if Spyre proves cost-effective and easy to adopt. Near-term financial signals are positive — Q3 sales rose 9.1% to $16.33B and infrastructure revenue jumped on Power11/z17 ramps — but IBM’s success hinges on demonstrating Spyre’s real-world inference performance and pricing to convince conservative customers to favor its stack over cloud GPU alternatives.