Source: Global Times

[Global Times Finance Comprehensive Report] Saturday's information shows that one of the world's largest AI chip clients, OpenAI, has recently started renting Google TPU chips to provide computing power support for products such as ChatGPT. This is its first large-scale use of non-NVIDIA chips.

Eastern IC

OpenAI's move aims to alleviate reliance on Microsoft data centers and reduce the cost of inference computing. With the number of paid subscribers for ChatGPT increasing from 15 million at the beginning of the year to over 25 million, plus hundreds of millions of free users each week, its computing demand has rapidly risen. Previously, OpenAI mainly rented NVIDIA server chips through Microsoft and Oracle, with expenditures exceeding $4 billion last year, and it is expected to approach $14 billion in 2025. Early this year, the popularity of ChatGPT's image generation tool placed significant pressure on Microsoft's inference servers, prompting OpenAI to seek support from Google Cloud.

Although Google has opened up TPU chips to OpenAI, it will retain more powerful TPUs for its own AI teams and Gemini model. Google began developing TPUs about 10 years ago and started providing them to cloud customers in 2017. In addition to OpenAI, companies such as Apple, Safe Superintelligence, and Cohere are also renting them, partly because these companies have former Google employees who are familiar with TPU operations. It is reported that Meta is also considering using TPUs.

Notably, an increasing number of companies are developing inference chips to reduce reliance on NVIDIA and cut costs. Amazon, Microsoft, OpenAI, and Meta have all launched self-developed plans. However, Microsoft's chip-making plan has encountered setbacks; Maia 100 is only used for internal testing, and Braga's AI chips have been delayed by at least six months, with performance expected to be significantly lower than NVIDIA's Blackwell chips.

Currently, Google Cloud still rents out servers supported by NVIDIA chips, as NVIDIA chips are the industry standard, generating higher profits, and developers are more familiar with their dedicated software. Google has already ordered over $10 billion worth of the latest Blackwell server chips from NVIDIA and has begun providing them to some customers this February. (Chen Shiyi)

Original article: https://www.toutiao.com/article/7521216318373560868/

Statement: The article represents the views of the author and welcomes your opinion by clicking on the [Up/Down] buttons below.