Microsoft weans itself off Nvidia chips with home-grown alternative - IT Pro

Microsoft's move isn't about "replacing Nvidia"—it's about cracking the CUDA monopoly. Once its in-house ASICs hit volume, the hardware abstraction layer in PyTorch gets rewritten, and the CoWoS advanced-packaging allocation logic at TSMC (Taiwan, China) shifts. Microsoft stops being a queue-waiting customer and becomes a second design pole with independent rules. On compliance, export controls to China have elevated single-vendor dependency from a commercial risk to an existential one. Microsoft's calculus: migrate inference workloads to silicon, keep training on Nvidia, and hedge supply-chain rupture via a workload split. This mirrors Apple's 2020 M1-for-Intel swap—never about raw performance, always about pricing power. In the competitive arena, Nvidia's data-center revenue leans over 60% on top-five clients. Any share erosion hits its valuation anchor directly. AMD will scramble for spillover orders, but the real threat is the AWS Graviton / Google TPU playbook. Once Microsoft validates the ROI, cloud-vendor ASIC penetration jumps from under 15% to 40%+ by 2026. The 12–24-month tail: Nvidia holds the training market, but the GPU premium in inference gets permanently compressed. CUDA's moat downgrades from "impassable" to "expensive to migrate." AI infrastructure pricing power resets.