Did you see the note on the btcrecover docs page:
Just be aware that the JTR kernel doesn't scale as well once you get past ~2x GPUs...
So, it would seem that simply cranking up the number of GPUs may not necessarily result in vastly superior hashing rates and it's possible that you're running into this issue where the added GPUs are just not being utilised properly due to the lack of scaling.
Have you tried a machine that just has 2x GPUs and tested to see what sort of performance you get there?
In the examples further down the page, they seem to use a setup that uses five instances of 2x 2080ti to achieve a "10 GPU" rig.
So, you might want to do as ETFbitcoin has suggested and benchmark a few different setups to find the best $/hashrate setup.