Apple uses Google, TPU trains AI models, Nvidia's dominant position faces risks

According to the financial news app Zhitong Finance, on Monday, Apple (AAPL.US) stated that its artificial intelligence system Apple Intelligence relies on artificial intelligence models pre-trained using processors designed by Google (GOOG.US, GOOGL.US), indicating that large tech companies are seeking alternatives to Nvidia (NVDA.US) when training cutting-edge AI.

Apple detailed in a recently released technical report the use of Google's self-developed Tensor Processing Unit (TPU) for training. Additionally, Apple also released a preview version of Apple Intelligence for some devices on Monday.

Nvidia's expensive Graphics Processing Units (GPUs) dominate the high-end AI training chip market, with demand being so high in recent years that it has been difficult to purchase them in the required quantities. OpenAI, Microsoft (MSFT.US), and Anthropic are all using Nvidia's GPUs to train their models, while other tech companies, including Google, Meta (META.US), Oracle (ORCL.US), and Tesla (TSLA.US), are also competing to buy these GPUs to build their AI systems and products.

Meta CEO Mark Zuckerberg and Alphabet CEO Sundar Pichai both stated last week that their companies and other companies in the industry may be over-investing in AI infrastructure, but also acknowledged the high business risks of not doing so.

In its 47-page report, Apple did not mention Google or Nvidia, but noted that its Apple Foundation Model (AFM) and AFM servers are trained on a "cloud TPU cluster." This means that Apple is renting servers from cloud service providers for computation. Apple stated in the report: "This system allows us to efficiently and scalably train the AFM model, including AFM devices, AFM servers, and larger models."

Compared to its peers who quickly embraced generative AI after OpenAI launched ChatGPT at the end of 2022, Apple revealed its AI plans relatively late. On Monday, Apple introduced Apple Intelligence. The system includes several new features such as a new look for Siri, improved natural language processing, and AI-generated summaries in text fields.

Over the next year, Apple plans to introduce features based on generative AI, including image generation, emoji generation, and an enhanced version of Siri that can access user personal information and perform actions in apps.

In the report on Monday, Apple stated that the device-side AFM was trained on a "slice" consisting of 2048 TPU v5p chips. This is the most advanced TPU, first introduced in December of last year. The AFM servers were trained on 8192 TPU v4 chips, which are configured to work together in eight slices through the data center network.

According to Google's website, the latest TPU costs less than $2 per hour when reserved for three years. Google first introduced the TPU for internal workloads in 2015 and made it available to the public in 2017 They are now one of the most mature custom chips designed for artificial intelligence.

Nevertheless, Google remains one of NVIDIA's main customers. It uses NVIDIA's GPUs and its own TPUs for training AI systems, while also selling the usage rights of NVIDIA's technology on its cloud.

Apple previously stated that the inference process (i.e., using pre-trained AI models to generate content or make predictions) will be partially done on its own chips in its data centers.

This is Apple's second technical paper on its AI system, following a more general version released in June. Apple stated at that time that TPUs were used in developing AI models.

Apple plans to announce its quarterly performance after the close of trading on Thursday