Customers are also competitors, which has always been a common question raised when the outside world pays attention to Nvidia.
Microsoft, Google, Amazon and other cloud computing giants have purchased a large number of NVIDIA GPUs in the past year, and these major clients are also laying out their own chip research plans. Huang Renxun previously mentioned this situation, stating that Nvidia faces more competition than 'anyone else on Earth'.
Even at this time when Nvidia is in the limelight, this situation is still happening. On July 29th, according to Reuters, Apple mentioned in a research paper that the AI big model being developed by the company uses Google TPU (Tensor Processor) as the underlying layer, rather than the commonly used Nvidia GPU (Graphics Processing Unit) in the industry.
As the last tech giant to enter the battle, Apple has not publicly reported any record of large-scale purchases of Nvidia's GPUs, and the market's focus has always been on who will be the winner of Apple's AI. According to reports, Apple used 2048 TPUv5p and 8192 TPUv4 to support the training and inference of AI models running on iPhones and other devices.
At present, Nvidia has not provided any evaluation or response to this news.
TPU was originally a specialized chip designed by Google for its TensorFlow machine learning framework. Since its launch in 2015, TPU has evolved to its sixth generation and has maintained an update rhythm of approximately one iteration per year. Prior to this, TPU was mainly used for internal use by Google, and although it was later linked to Google Cloud services for external use, there have been no major external expansion actions.
Last year, the explosion of AI big models led to a frenzy of market competition for AI chips, with high-end GPUs in high demand. Nvidia has won over 80% of the market share in this field and is undoubtedly the dominant player. But at the same time, Google is also secretly working on TPU. According to Techinsights, Google's estimated self use of TPU chips last year exceeded 2 million, with a market share second only to Nvidia and Intel, making it the world's third-largest data center chip design manufacturer.
Despite having its own TPU chip, Google remains one of the world's largest buyers of Nvidia GPUs. In a report written by market research firm Omdia, the list of major buyers of Nvidia H100 GPUs that were snapped up last year was compiled. Meta and Microsoft tied for first place with a purchase volume of 150000 H100 GPUs, while Google, Amazon, Oracle, and Tencent each purchased 50000 H100 GPUs, tied for second place.
Gu Geyun also had close cooperation with Nvidia last year. Google not only uses Nvidia GPUs internally, but also provides Nvidia GPU based services on its cloud service platform to meet customers' demands for high-performance computing and AI applications.
In addition to Google, cloud giants such as Amazon Web Services and Microsoft are developing their own chips based on the Arm architecture. The chip manufacturing of cloud computing giants has always been seen as a threat to Nvidia by the outside world. But Nvidia has always insisted that it has an exclusive advantage in the face of competition.
As early as 2017, when Google launched its second generation TPU, Huang Renxun stated in an interview with CNBC that he was "not worried about competition from Google TPU". In his view, although some major cloud computing clients may develop their own AI server chips to reduce their dependence on NVIDIA chips, NVIDIA can still maintain its leading position in the AI field with the outstanding performance of its GPUs.
Cost has always been Nvidia's trump card for its own products. According to Huang Renxun's proposition of "the more you buy, the more you save", due to economies of scale, the average cost will decrease. Therefore, when enterprises purchase a large number of NVIDIA GPUs, although the initial investment may be large, in the long run, high-performance GPUs have a longer service life, lower maintenance costs, and lower overall operating costs (TCO) for customers. If competitors want to compete directly with them, even if they are free, they are still not cheap enough in the end.