To The Moon

AsianFin -- Apple is in talks with tech giant Tencent and ByteDance, the Chinese parent company of TikTok, flirting with the idea of integrating a Chinese firm’ artificial intelligence models into iPhones sold in China, according to three sources familiar with the matter.

This month, Apple began rolling out OpenAI's ChatGPT on its devices as part of its Apple Intelligence product, allowing Siri to access the chatbot’s capabilities for user queries related to photos, documents, and presentations.

However, ChatGPT is not available in China due to the country's strict regulatory requirements, which mandate government approval for generative AI services before they can be publicly released. As a result, Apple is looking to localize its AI features by partnering with local companies at a time when its market share in China is shrinking.

Sources indicated that talks with Tencent and ByteDance about using their AI models are still in the early stages, and all parties involved are keeping the discussions private. Tencent has declined to make comments on the matter. ByteDance’s Volcano Engine Cloud President Tan Dai refused to comment on the issue, saying only that Android phones have a higher market share than iPhones in China.  

The successful partner for Apple’s AI services in China could become a significant player in the country’s competitive AI landscape, where numerous large language models are being launched by major tech firms and startups. These include ByteDance’s Doubao, Tencent’s Hunyuan, and Baidu’s Ernie.

Apple had reportedly been in talks with Baidu about utilizing its AI model in China. However, discussions reportedly hit obstacles due to some issues, including disagreements over using iPhone user data to train AI models.

Following the news report on Thursday, Baidu's Hong Kong-listed shares dropped by 4.2% at close, while Tencent's shares rose by 2.3%. ByteDance is still a private company although it is planning to get listed in Hong Kong or the Chinese mainland.

The lack of AI capabilities in the latest iPhones sold in China has put Apple at a disadvantage, especially as it faces intensifying competition from Chinese brands, including Huawei.

Huawei, which re-entered the high-end smartphone market in August with a device powered by a Chinese-made chip, launched its Mate 70 series last month featuring AI capabilities driven by its own large language model.

Apple briefly fell out of the top five smartphone vendors in China during the second quarter, before recovering in the third quarter. However, Apple’s sales in China still saw a slight year-over-year decline of 0.3% in the third quarter, while Huawei's sales surged by 42%, according to research firm IDC.

However, the Apple Intelligence feature is expected to serve as a major catalyst for the multi-year upgrade cycle of Apple devices, according to a report by Morgan Stanley. Over the next two years, iPhone shipments are projected to exceed 500 million units, with shipments in fiscal years 2025 and 2026 estimated at 235 million and 262 million units, respectively.

ByteDance and Tencent have emerged as leading purchasers of Nvidia Hopper chips in 2024, trailing just behind Microsoft, according to Omedia.

Each company reportedly procured approximately 230,000 units of NVIDIA’s H100 and H200 GPU compute cards, based on the Hopper architecture, surpassing major U.S. tech players like Meta, Amazon, and Google.

Microsoft topped the list with around 485,000 GPUs purchased this year, more than triple its 2022 procurement volume of the same generation NVIDIA processors. ByteDance and Tencent tied for second place, highlighting the growing demand for advanced AI hardware among Chinese tech firms.

The substantial GPU investments underline ByteDance's and Tencent's intensifying focus on artificial intelligence applications and cloud computing infrastructure, as well as their strategic emphasis on staying competitive in an AI-driven digital economy.

In media interviews before the launch of new Doubao model on Thursday, Tan was quoted as saying that ByteDance was aiming to become the top AI player in China.

“The new technology revolution is … AI and large language models. Our goal is to become No 1 in the area and we’re marching towards that direction,” Tan told a local media outlet.

Tan told AsianFin that the emergence of AI has shifted the focus of all infrastructure from CPU-centric to GPU-centric, making the concept of “cloud-native” increasingly significant. “We believe that over the next decade, AI cloud-native will become even more critical. This represents a major transformation in infrastructure, moving from cloud-native to AI cloud-native. Volcano Engine aims to become a leading company in this area,” he said.

Tan also stressed the importance of application scenarios for AI large models. He pointed out that it’s not just about platforms and algorithms but also about providing services to help enterprises identify use cases for AI large models, thereby ensuring the successful implementation of AI technologies.

ByteDance's Doubao and another short video platform Kuaishou in China have two key advantages in the realm of large AI models: First, the language models require continuous training through a "data flywheel," and major tech companies naturally possess significant traffic advantages, which enhances their model capabilities, particularly in subjective evaluations; second, in the domains of text-to-image and text-to-video models, ByteDance and Kuaishou hold a clear edge due to their access to high-quality short video data, giving them a distinct advantage over non-internet companies, Lin Yonghua, Deputy Director and Chief Engineer of the Beijing Academy of Artificial Intelligence (BAAI), a Beijing-based research institution engaged in AI research and development, told AsianFin.

BAAI on Friday released the latest results from its large-scale model evaluation platform, FlagEval, based on over 800 open-source and closed-source models worldwide. The evaluation covered more than 20 types of tasks, over 90 datasets, and over two million test questions. The results included comprehensive and specialized evaluations in nine areas, such as language, visual-language, text-to-image, text-to-video, and speech-language large models.

For language models, ByteDance's Doubao-pro-32k-preview and Baidu's ERNIE 4.0 Turbo ranked first and second, respectively. In the objective evaluation of language models, OpenAI's o1-mini-2024-09-12 and Google's Gemini-1.5-pro-latest secured the first and second spots, followed by Alibaba's Qwen-max-0919 and ByteDance's Doubao-pro-32k-preview in third and fourth places. Meta's Llama-3.3-70B-Instruct rounded out the top five.

For vision-language multimodal models, OpenAI's GPT-4o-2024-11-20 and ByteDance's Doubao-Pro-Vision-32k-241028 took the lead, ahead of Anthropic's Claude-3-5-sonnet-20241022. They were followed closely by Alibaba's Qwen2-VL-72B-Instruct and Google's Gemini-1.5-Pro.

Lin pointed out that there is no such thing as running out of data for AI large models. A decade ago, internet data accounted for nearly 5% of global data volume, but in 2021-2024, this figure dropped to 1.3%. Despite this, the share of Internet users in Chinese has remained steady at about 19%. This has led to an "islands of isolation" formed by vast amount of internet data. Breaking these data silos in AI model training will become critically important. Meanwhile, "synthetic data" addresses more complex problems and directions, offering a more efficient way to generate data.