CITIC Securities: How to view Apple's AI future upgrade direction?

According to the financial news app Zhitong Finance, CITIC Securities released a research report stating that, based on the continuous optimization of AI experience, Apple's new round of upgrades will be driven by AI. The research team at CITIC Securities is optimistic about the iPhone upgrade cycle with a huge user base. It is expected that in 2025, the iPhone will have a strong upgrade momentum, with estimated shipments of 2.25+/2.4+/2.5+ billion units in 2024/2025/2026, entering an upward trend in shipments. Looking ahead, the curtain is rising on Apple's innovative year, firmly bullish on Apple's software and hardware innovation cycle.

The following is a summary of the research report:

We believe that in order to further seize the distribution entrance of data and traffic to maximize user dependence on Apple Intelligence itself, it is necessary to continue to strengthen Apple's proprietary model capabilities. This will be manifested in the enhancement of computing power and operating memory capacity at the end, increasing the model parameter volume; and there is potential to increase computing power in the cloud. Based on the continuous optimization of AI experience, we believe that Apple's new round of upgrades will be driven by AI. We are optimistic about the iPhone upgrade cycle with a huge user base, and we believe that the iPhone is expected to have a strong upgrade momentum in 2025, with estimated shipments of 2.25+/2.4+/2.5+ billion units in 2024/2025/2026, entering an upward trend in shipments. Looking ahead, the curtain is rising on Apple's innovative year, firmly bullish on Apple's software and hardware innovation cycle.

How is Apple Intelligence doing? - The end-side 3B model rivals mainstream 7B models, and the cloud side is on par with GPT-3.5-Turbo.

Apple's AI landing is completed by Apple Intelligence's proprietary models (a large end-side 3 billion parameter model + a private cloud large model) + ChatGPT 4o. Depending on the availability of computing power, calculations rely on the terminal, while more complex scenarios use private cloud computing, and ChatGPT is called relatively later in the sequence. Apple's Senior Vice President Craig Federighi stated at an event after the WWDC keynote speech that Apple will also collaborate with third-party models outside of OpenAI (such as Google's Gemini model). We believe that Apple considers third-party models as alternative solutions, as the confidence comes from the relatively excellent capabilities of its proprietary models. According to Apple's official website, based on human evaluation scores of large models: 1) Compared to mainstream 7B models (Mistral-7B, Gemma-7B) and Phi-3-mini (3.8B), Apple wins + draws probability >65%. 2) Compared to cloud-side private cloud large models and DBRX-Instruct, GPT-3.5-Turbo, Mixtral-8x22B, Apple wins + draws probability >70%. Furthermore, considering that Apple's proprietary models have a higher priority in calling than third-party models, we believe that in order to further seize the distribution entrance of data and traffic to maximize user dependence on Apple Intelligence itself, It is necessary to continue to strengthen Apple's in-house model capabilities in the future.

What is Apple's AI development strategy in the future? - Enhancing edge-side computing power + operating memory, increasing model parameters; expanding cloud-side computing power layout.

1. Edge-side: The enhancement of computing power + operating memory is the main logic, driving the iteration of SLP, heat dissipation, and battery. Currently, Apple supports a 3 billion parameter large model on the edge side (considering that the Qualcomm Snapdragon 8Gen3 released in 2023 is aimed at a 7 billion parameter large language model that can generate 20 tokens per second, we expect that the parameter volume of Android flagship smartphones on the edge side may exceed 100 billion this year). Under the driving force of the Scaling law, the parameter volume still needs to be further increased, driving the upgrade of computing power (we expect it to increase from last year's A17 Pro's 35TOPS to over 40TOPS this year, and it is expected to reach 45-50TOPS next year), operating memory upgrade (starting from 6GB for the iPhone 15 series, we expect new models this year to start from 8GB), as well as SLP, heat dissipation (VC thermal plate), charging power, battery, and other supporting upgrades.

2. Cloud-side: Apple is expected to increase server procurement in the future and accelerate the development of in-house inference chips. According to Trendforce citing Mac Rumors, Foxconn is currently assembling Apple AI servers including the M2 Ultra (the M2 Ultra uses Apple's UltraFusion custom packaging technology, connecting two M2 Max chips through an interposer to provide over 2.5TB/s of low-latency processor-to-processor bandwidth, previously used only in high-end Mac Studio and Mac Pro products), and plans to assemble Apple AI servers driven by the M4 chip by the end of 2025 (we expect Apple's in-house chips to be used for the inference end, while the training end will still rely on NVIDIA). According to C114 citing industry chain sources, Apple is expected to purchase around 2000-3000/18,000-20,000 servers in 2023/2024 respectively.

What is the logic of transmission from product innovation to demand realization?

1. Review of the last replacement cycle: Apple's last replacement cycle occurred during the 5G replacement in 2020-2021. On October 13, 2020, Apple launched its first 5G phone, the iPhone 12 series. The shipment volume in the fourth quarter of 2020 began to significantly increase year-on-year, and this trend continued until the third quarter of 2021 (according to IDC, the year-on-year shipment growth rates from 2020Q4 to 2021Q3 were 18.6%/50.5%/18.2%/22.7%). On an annual basis, in 2020, iPhone achieved shipments of 203 million units (an increase of 12 million units year-on-year), and in 2021, iPhone shipments reached 236 million units (an increase of 33 million units year-on-year). Subsequently, the iPhone entered a stock cycle again, with shipments of 225 million and 231 million units in 2022 and 2023 respectively

Focus on this round of replacement cycle: We believe that Apple's new round of replacements will come from AI-driven (affected by the threshold of 8GB operating memory capacity, only the iPhone 15 Pro and iPhone 15 Pro Max models in existing phones can support Apple Intelligence). According to statistics, as of 2023, there are 1.46 billion active iPhone devices globally, even excluding the 230 million new iPhones shipped in 2023, the number of active devices still reaches 1.23 billion. Therefore, the replacement momentum brought by "AI experience + hardware exclusivity" is worth looking forward to. It is worth noting that we believe that the replacement trend brought by AI in this round will be stronger than the replacement trend brought by 5G upgrades, because Apple's differentiated advantage in AI implementation is stronger (specifically, in the 5G replacement cycle, there is no fundamental difference between Android and Apple, both achieve higher network transmission rates; in the AI replacement cycle, Apple's system-level AI advantage is significant, while the Android side has not yet opened up cross-application data calls, resulting in a clear short-term gap), in addition, AI is expected to bring users a better user experience. Therefore, we are optimistic about the iPhone replacement cycle with a huge user base, and we believe that the iPhone is expected to usher in a strong replacement momentum in 2025, with estimated shipments of 225 million+/240 million+/250 million+ units in 2024/2025/2026, entering an upward trend in shipments.

Risk factors:

Consumer replacement willingness is lower than expected; the rise in terminal hardware costs affects demand; changes in the international industrial environment and escalating trade frictions; AI commercialization progress is lower than expected; hardware-side computing power upgrade progress is lower than expected; large model iteration speed is lower than expected.

Investment strategy:

Considering that only a few models of Apple's AI functions are implemented, coupled with the expectation that future iPhone new products will continue to innovate in computing power, large model implementation, heat dissipation, optics, and other directions, we believe that the iPhone is expected to usher in a strong replacement momentum and enter an upward trend in shipments