I don't know when but I'm going off:

- "OpenAI is partnering with Cerebras to add 750MW of ultra low-latency AI compute"

- Sam Altman saying that users want faster inference more than lower cost in his interview.

- My understanding that many tasks are serial in nature.