SemiAnalysis believes OpenAI's new Ultrafast mode, capable of around 300 tokens/sec, is not using Cerebras hardware. The firm says OpenAI is likely serving it at very low batch sizes on Nvidia Blackwell-class GPUs, potentially the GB300.
One free account opens every story complete, the live board, the 30-day calendar, a watchlist and breaking alerts on your phone. No card. The wire stays free to read.
