Yeah indeed, if the ML units are integtated within the CUs (as in just accelerated instructions similar to the rapid packed math FP16x2), then yes, any ML workload is going to encroach on FP32 throughput, reducing it substantially, especially if we account for things like register pressure and memory bandwidth.
This would be the same situation as RT units in current consoles and current AMD GPUs, they are integtated within the Texture units, which is why RT is weak on these GPUs, texturing reduce RT throughput and vice versa, ML units being structured in similar way would lead to the same result.
If the PS5 Pro is working like this, then they won't afford to do anything ML other than upscaling, there is simply no hardware throughput to do anything else. The PS5 Pro would also be limited in the amount of upscaling it can do. DLSS for example can afford to do competent 8X upscaling (from 1080p to 8K), while other upscalers can't. DLSS can also handle competent ML antialiasing at native resolution (DLAA) without too much of a performance impact, and it can do ML downsampling from any resolution. These things would be outside the scope of PS5 Pro ML capabilities.
It would also mean Sony calculated how much ML it exactly needed for upscaling, without an ounce more.
I guess we should be on the lookout for upcoming implementations, if we see PS5 Pro games with nothing more than typical 4x PSSR upscaling, then that would be a very powerful indication of the limited integrated ML capabilities of the PS5 Pro.
However, if we see games with with further ML capabilities like PSSR 8x upscaling to 8K or games doing multiple ML algorithms at the same time like, PSSR frame generation or PSSR AA, or HDR conversion, then that would be evidence that PS5 Pro possess more advanced discrete ML capabilities.