I think they mean serving inference workloads

How does that work? Isn't most bitcoin mining done on custom ASICs? I didn't think that the ASIC could be repurposed for inference.

Training ASICs (like Google’s TPUs) can generally run inference too, since inference is a subset of training computations. TPUs are widely used for both.

Mining ASICs (Bitcoin, etc.) cannot be repurposed…they’re hardwired for a single hash algorithm and lack matrix math needed for neural networks.

The biggest cost is the power which is often on multi year contracts. The hardware is comparatively cheap

That's wildly inaccurate. The cost in enormous both on the inference side and the mining side and has short lifetimes if you want SOTA.