SemiAnalysis: OpenAI's "consideration of adopting Google TPU" leads to NVIDIA's concession, reducing its computing power cluster costs by about 30%

Research institution SemiAnalysis points out that since the release of GPT-4o in May 2024, OpenAI researchers have not successfully completed any large-scale deployment of new cutting-edge pre-trained models. In addition, OpenAI has not even truly deployed TPU, only "possibly preparing to adopt TPU", which prompted NVIDIA to make concessions, resulting in a cost savings of about 30% in overall computing power clusters.