AI vendors broaden silicon options as Windows 11 restores controls

AI infrastructure providers are increasingly pairing or replacing general-purpose GPUs with specialized silicon for inference. Cerebras is moving to a rack-scale CS-4 architecture with modular compute and improved power delivery, positioning its wafer-scale systems as a high-speed option for token generation alongside partners including AWS and AMD.

Google is bringing Marvell into work around custom XPUs and TPUs, adding another chip IP supplier alongside Broadcom. Waymo has also developed a custom machine learning accelerator for autonomous vehicles, aiming to handle large sensor streams with low latency while moving beyond FPGA-heavy designs that can be difficult to program and less compute-dense.

Microsoft is meanwhile restoring Windows 11 features that users had in earlier releases. The company has added taskbar repositioning to the Release Preview channel, with options for left, right, top, or bottom placement, and is testing more customizable context menus. Task Manager is also gaining more visibility into AI workloads, including process-level use of NPUs.