Comment by dkdcdev

at their scale they could also just run a large on-premise or rented (basically still cloud, but cheaper) GPU cluster and run through that. fixed costs, even license a SOTA model’s weights if you’d like