Inference, don't need gpu's for inference. Frontier labs are eking out progress by scaling up inference-time compute. Pre-training scaling has kind of stalled / giving diminishing returns (for now).