Floating compute for AI training and inference, in months instead of years.