ZK is currently stuck with arithmetic circuit representations which are predictably very expensive to use as a representation for tensor data.
The matrix based formulations are still limited and don't play nicely with the parts of ML models which go beyond simple matrix multiplication.
I suspect someone will unify the two threads of research eventually, but it doesn't seem like it's there yet.
(FHE ML is even further away)
I can see some niche applications for this system, but I am very skeptical it's ability to handle larger models (100M+) and the ability to and it's scalability when there are increased demand.