首次大规模使用“非英伟达”芯片，OpenAI租用谷歌TPU，降低推理计算成本

Group 1 - OpenAI has begun renting Google's TPU chips for the first time on a large scale, reducing its reliance on NVIDIA's GPUs and alleviating pressure on Microsoft's data centers [1][2] - OpenAI's demand for computing power has surged, with paid subscribers for ChatGPT increasing from 15 million at the beginning of the year to over 25 million, alongside hundreds of millions of free users [1] - Companies like Amazon, Microsoft, OpenAI, and Meta are developing their own inference chips to decrease dependence on NVIDIA and lower long-term costs [1][2] Group 2 - OpenAI spent over $4 billion on NVIDIA server chips last year, with training and inference costs each accounting for half, and is projected to spend nearly $14 billion on AI chip servers by 2025 [2] - The shift to Google's TPU was driven by the explosive popularity of ChatGPT's image generation tool, which increased pressure on OpenAI's inference servers [2] - Google has been developing TPU chips for about a decade and has provided this service to cloud customers since 2017, with other companies like Apple and Cohere also renting Google's TPU [2][4] Group 3 - Meta is also considering using TPU chips, indicating a broader trend among major AI chip customers [3] - Google Cloud continues to rent out NVIDIA-supported servers, as they remain the industry standard, generating more revenue than renting TPUs [4] - Google has ordered over $10 billion worth of the latest Blackwell server chips from NVIDIA, starting to provide them to select customers since February [4]