2025-08-23XQuote

Source item

As inference exploded & compute remained constrained, specialized inference clouds (w varying degrees of price / speed advantages) are quickly becoming at-scale compliments to huge NVDA clusters. @cerebras is helping lead the way. 💥🚀

Quoted context

These are the quoted author’s words.

Earlier this week @NVIDIA shared its performance on @OpenAI 's gpt-oss-120B model running on DGX B200 ( 8 two die Blackwell chips). They asked @ArtificialAnlys analysis to do a special set up for them. They showed performance of 900 tokens per second for one user, and dropped to 580 tokens per second for 10 users. I have pasted their link in the comments below. We asked Artificial Analysis to run the same setup for us. The result is ~2700 token/s for 1 and the same for 10 users. Not only is @cerebras Inference the fastest, we sustain performance as you scale up your workload. Cerebras Inference is in full production today, serving billions of GPT 120B tokens per week on @openrouter , @huggingface , and the Cerebras Cloud.
andrewdfeldman ↗