Cloud Providers Target NVIDIA Dependence with Custom Silicon and PyTorch Portability

Microsoft says Maia 200 serves Copilot and plans Maia 300 for 2027 inference, claiming improved rack output and efficiency; shifting selected workloads could trim NVIDIA purchases, but execution and impact are unproven. Amazon’s TorchNeuron and Google’s TorchTPU also target portability to rival chips, with broad adoption and compelling switching economics not yet established.