Less than a year would also match what DeepSeek was reporting, where compute pays off in just 10 months. I think they also said they had an 80% margin on inference.