As NVIDIA dominates AI chips, AWS and Google's in-house silicon is edging toward a real alternative, from both the software stack and cloud distribution sides.
AWS and Google Cloud open-sourced PyTorch builds for AWS Trainium and Google Cloud TPU (Publickey, Oct 26). The next day, Oracle Cloud announced it will offer Google TPUs starting in 2027 (Publickey, DatacenterDynamics).
Open tooling and TPU availability on a rival cloud landed in the same week, both lowering the barrier to moving off NVIDIA. Still, these are announcements; whether workloads actually migrate is not confirmed. The same week also saw a reported $63bn Broadcom-led financing for Anthropic's chip purchases and a Google data center hearing in Minnesota shut down by protests, so strong demand coexists with siting headwinds.
Real-world performance and adoption of the new PyTorch builds, the terms of Oracle's 2027 TPU offering, and how chip supply and data center siting constraints evolve.