OpenAI was bragging the other day about how HLE on Cerebras only took 11 hours. But you're saying if they just spun up a separate cheapo VM for each task, it would have finished faster and cheaper?