“Early benchmarks from Microsoft suggest its custom Maya 200 AI chips deliver competitive inference performance at a lower cost per token compared to equivalent NVIDIA GPU configurations.”