Calling it the highest performance chip of any custom cloud accelerator, the company says Maia is optimized for AI inference on multiple models.