Separation of data and model weight in machine learning inference

Is there a combination of AWS services that would allow the following: B1 trains a model that transforms input data into valuable insight. B1 wants to make this model available to other businesses at a cost. B2 is such a business, but B2 has strict data ownership requirements that does not allow it to expose the data to B1. So B1 does not want to expose model weights to B2 and B2 does not want to expose data to B1. What kind of architecture would allow this to happen?

I was thinking of a contract that forces Cloudtrail with log file validation turned on in an account managed by B2, B1 shares the AMI with B2 and has access to a role within that account to monitor the fair use of the AMI? What are your thoughts?