They could potentially quantize the model and run it at lower quality taking less VRAM.