Buyers of training data often prefer a perpetual license because it matches how models work. A model trained on data keeps what it learned for as long as it exists, and later models are often built from earlier ones. A license that never expires removes the question of what happens to those models at the end of a term. It also lets the buyer retrain on the same files without returning to the seller.
For a seller, a perpetual license is a one-time decision. There is no renewal at which to raise the price or withdraw. It is usually priced higher than a term license for the same data, and it is commonly paired with a narrower scope. Perpetual and irrevocable are different words. A perpetual license has no fixed end but may still be terminated for a stated cause, such as non-payment or misuse. An irrevocable license cannot be ended by the seller. Read which one the draft says, and take advice on the difference.
A seller that grants a perpetual license should keep a permanent record of it. The grant continues to bind the company through later changes of ownership, and a future acquirer of the company will ask to see it.