That’s the claim, it has apparently been trained using a fraction of the compute power of the GPT models and achieves similar results.

If you need far less computing power to train the models, far less gpus are needed, and that hurts nvidia