Well actually that's not true, they have "some idea" but only to the extent that the cloud providers themselves "have an idea."
Remember, these algos are constantly being tuned and tweaked. So if you build an accelerator for an algo that isn't fully "baked" then you've basically wasted a ton of money building an ASIC to accelerate a workload you no longer give a hoot about.
The FPGA solution is nice if you're running a bunch of different algorithms that need to be accelerated and/or your algorithms themselves are constantly in flux and so too is your accelerator.