If a foundation model provider allows fine-tuning on confidential data but the result is locked into the platform, wouldn't that be extremely sticky?