Microsoft’s new AI chip makes it less reliant on Nvidia’s GPUs - The Verge

Microsoft's in-house AI accelerator isn't about replacing Nvidia—it's about killing the compute-rental model. Stripping inference workloads out of the CUDA ecosystem attacks Nvidia's thickest software moat: the lock-in premium justifying 3-5x hardware markups. Upstream, advanced packaging allocation at TSMC (Taiwan, China) tilts toward Microsoft's custom ASICs, reshaping wafer priority for H100/B200. On compliance: under tightening export controls, single-vendor dependency has escalated from commercial risk to geopolitical exposure. Microsoft, AWS, and Google now form a hyperscaler trinity that collectively reclaims pricing power from chipmakers. Market dynamics: Nvidia's existential threat isn't silicon—it's CUDA de-standardization. Once the Big Three capture 40%+ of inference compute, Nvidia retreats into training workloads and sovereign-AI contracts. AMD gains a structural second-source window. 12-24 month outlook: the GPU tax on inference compresses below 30%. Nvidia's data-center revenue growth decelerates from 80%+ to the 30-40% band. Microsoft's Azure AI-inference gross margin improves 5-8 points. This isn't substitution—it's repricing.