Better than 50% discounts for reasonable FPGA volume, true. Still expensive when compared to a V100.

Microsoft is seeing ~90Tops for their MS-FP8 datatypes: https://www.top500.org/news/microsoft-takes-fpga-powered-dee...

It is certainly hard to compete against an NVidia V100 with its tensor ops.

It can get interesting though. There are quite a few graphs that tensorflow cannot place on a GPU, especially with RNNs. An FPGA does give you more flexibility in architecture which was, and would remain one of the motivations for MS to choose them as a platform.