news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

This article is written based on public information and is only for information exchange purposes. It does not constitute any investment advice

The spiraling $3 trillion club is now ruling the largest pool of wealth in human society:

So far this year, there are three companies whose U.S. stock market value has reached the 3 trillion U.S. dollar threshold. The first is Apple (about 3.5 trillion U.S. dollars), the second is Microsoft (about 3.3 trillion U.S. dollars), and the third is Nvidia (about 3.0 trillion U.S. dollars).

Although their businesses seem to be unrelated, they are actually closely connected. There is a common driving force behind their rise to the top of the 3 trillion US dollar throne: AI.

The development of AI has formed an invisible triangular flywheel, which is computing power, big models and terminal applications. Computing power is the starting point of all miracles. With the support of computing power, the model can turn rotten data into magic. However, no matter how good the big model is, it still needs AI terminal applications that face consumers to pay for it.

• The leader in computing power - NVIDIA, this is the easiest to understand. Relying on the monopoly of AI GPU, NVIDIA has become the undisputed leader in AI computing power and the most authentic water seller. Although Huang talked a lot about smart robots, AI lithography machines and so on, no one on Wall Street believes it. They only care about how many cards it can produce in the next quarter.

• The situation of Microsoft, the master of big models, is relatively complicated. As the largest investor of Open AI, it has a special status. On the other hand, it also has its own large model training, which makes Microsoft the leading player in big models. At the same time, the cloud computing service provided by Microsoft also makes it half a water seller from this perspective. In addition, Microsoft also has an office suite that beats everything else, so Copilot is also the most anticipated AI application, even though it has fallen short of expectations for several consecutive quarters.

• Apple, the disruptor of AI terminals, is the most puzzling. It seems that at the beginning of the year it was still criticized for betting on MR and automobiles and missing out on the entire AI wave. However, in April, relying on the concept of AI terminals, its stock price soared by 30%, staging a story of the elephant dancing and the king's return: it is a case of expectations ahead of the troops moving forward.

The three companies' stock prices have spiraled to new highs around AI, which actually reflects the market's pursuit of different stages of the AI ​​flywheel. Among them, computing power and models have been repeatedly studied in more than a year, and there is not much cognitive gap.

The last piece of the puzzle, the AI ​​terminal, is still unclear, but the industrial and wealth context has gradually become clear - this may be a major direction that we must focus more on when studying AI.

Figure: Stock price trends of three 3 trillion companies in the past quarter Source: Wind

01AI terminal is not a false proposition

The lack of AI applications still stings every technology fan.

Since the AI ​​trend started to take off last year, everyone has been imagining the proliferation of AI applications. AI applications can be divided into two categories: software products with embedded AI functions, such as Microsoft's Copilot and AI voice assistants, and smart hardware terminals with embedded AI functions.

After ChatGPT came out, after a year of development, the highly anticipated AI application software has not been as good as expected. The market ultimately attributed it to the fact that the large model was not powerful enough. It is for this reason that major companies are scrambling to purchase computing power from cloud computing vendors to iterate their own models. Among them, Microsoft has an API interface exclusively authorized by OpenAI, and is clearly ahead of Amazon and Google in AI model services (AI MaaS), becoming an unexpected winner.

However, the explosive sales of CSP computing power cannot cover up the fact that AI applications are not as good as expected. The hot AI applications that the market expects may still need time and wait for new miracles to come out under the scaling law.

Figure: Market share of generative AI MaaS layer (2023) Source: IoT analyst, Huaan Securities Research Institute

But the AI ​​trend has already taken shape, and the wind of innovation will not disappear, but will only shift. With Apple's market value hitting a record high, it seems to indicate that the old bottle of smart hardware terminals can hold the new wine of AI.

Inferring from the laws of industrial development, we believe that AI terminals are not a false proposition, mainly based on the following two considerations.

1) The industrial logic of AI terminals is smooth

At first, people were not optimistic about AI terminals, because smart terminals such as mobile phones and computers are limited by power consumption and size, and the computing power of the main chip is difficult to accumulate infinitely like a server cluster, which also leads to the lack of intelligence of terminal AI. The most representative case is Siri, which has been on the market for more than ten years. After countless versions of iterations, it is still like a fool.

However, with the advent of the big model era, the problem of intelligent terminals being unable to accumulate computing power has finally been solved: the big model on the cloud side acts as the intelligent brain, and the big model on the terminal side of the intelligent terminal acts as an intelligent interaction method, with the cloud and the terminal collaborating with each other. The big model on the cloud side solves the problem of deep intelligence, and the big model on the terminal side solves the problem of timely intelligence.

In addition to timeliness and portability, why do we need to buy a smart terminal equipped with an end-side model when there is a large model with huge parameters and powerful functions, instead of allowing people to communicate directly with the large model?

Because this also involves the issue of privacy. Every individual has different needs for AI. Our ultimate goal is to make AI act as a personal secretary, which naturally requires intensive interaction under high trust, and privacy issues are particularly prominent. Imagine a scenario like this: processing a private photo and randomly sending it to a large model in the cloud for processing. How great is the risk of privacy leakage you face?

Therefore, Qualcomm released a report titled "Hybrid AI is the Future of AI", in which it points out:

• Just as traditional computing has evolved from mainframes and thin clients to the current combination of cloud and edge terminals, AI processing must be distributed across the cloud and terminals to achieve the scale expansion of AI and realize its maximum potential.

• Instead of processing only in the cloud, hybrid AI architecture distributes and co-processes AI workloads between the cloud and edge terminals. Working together, the cloud and edge terminals (such as smartphones, cars, PCs, and IoT terminals) enable more powerful, efficient, and highly optimized AI.

  
  

Based on the law of industry development, our smart terminals such as mobile phones and computers, with the support of cloud big models, have become natural carriers of AI, which is a complementary relationship rather than an either-or relationship. As the functions become more powerful, we will hopefully see AI terminals become users' smart assistants and personal secretaries, automatically providing personalized functions and suggestions.

Although Qualcomm issued a report in 2023 explaining this logic, it was not until June 2024 that the industry trend of AI terminals gradually became clear and began to be widely accepted when Apple, which has a full range of products and its own operating system, entered the market and gave up on making cars.

2) Supply-side logic: manufacturers are more urgent than you and me

No matter how good the industry trend is, it still needs manufacturers to promote it. Especially for the technology industry, we believe in the golden rule that "supply creates demand." For example, although photovoltaic power generation is good, it took 20 years for the supply side of the industry to reach parity, and then there was a huge explosion in demand.

Another reason for being optimistic about the ultimate realization of AI terminals is that the supply side is in the process of self-revolutionary innovation. It can even be said that smart terminal manufacturers are even hungrier than AI practitioners.

How do we know? From IDC's statistics, it is not difficult to find that in 2022 and 2023, global shipments of smartphones, computers and tablets have declined year-on-year for two consecutive years. The global shipments of the largest single product, mobile phones, have dropped from a high of 1.5 billion units to 1.1 billion units. Even a strong company like Apple can only rely on micro-innovation to squeeze out toothpaste and raise prices every year to barely keep its profits from falling.

The entire smart terminal industry is in great need of a "gimmick" to boost its sales. According to the forecasts of third-party organizations such as Counterpoint, this is definitely an opportunity that the giants cannot miss: the number of AI mobile phones worldwide is expected to reach 520 million in 2027, with a penetration rate of 40%; the number of AI PCs worldwide will reach 290 million in 2028, with a penetration rate of over 70%.

03AI terminals will flourish, and AI phones and PCs are worth paying attention to

The highly anticipated AI PIN and AI smart speakers last year were ultimately short-lived; after the concept faded and returned to its roots, we believe that in the innovation of AI terminals, perhaps the most noteworthy ones are the traditional smart terminals, such as mobile phones and computers.

1) The AI ​​phone that has already rolled up before it even starts

According to Canalys' definition, AI mobile phones must at least meet the following standards: the main control chip SoC should include a dedicated unit that can accelerate AI tasks, commonly known as the NPU, and be able to run LLM and other generative AI models on the terminal side; the reasoning performance of the terminal side model should be faster than the human reading speed, equivalent to 10 token/s; the time for the terminal side AI to generate an image is less than 2 seconds.

On January 18, 2024, Samsung launched the Galaxy S24, which is known as the first AI mobile phone. In addition to the large model on the mobile phone side, the large model on the cloud side of Samsung's AI mobile phone uses Google Gemini overseas and Wenxin Yiyan in China. It can provide "novel" functions such as AI cutout and real-time translation, which once impacted Apple's high-end mobile phone market.

Since then, the competition for AI mobile phones has officially begun. For mobile phone manufacturers, since large models are currently mostly external to OpenAI, Google, and Baidu, etc., in terms of hardware parameters, what everyone is competing for is nothing more than the number of parameters of the LLM large language model on the end side, and the strength of the NPU on the end side that supports the operation of these models.

Different from the hundreds of models, the edge model needs to find a balance between power consumption and software capabilities. Therefore, Apple, which has the advantages of self-developed chips, self-developed operating systems, and the largest single customer base, is a latecomer and has basically locked in a ticket after launching Apple Intelligence.

As for Android, with the support of Google and Qualcomm, some companies should be able to win, so that Apple will not dominate the market alone. However, it is still unclear who will do the best, but what is certain is that the success or failure of AI phones will directly change the future of Android.

Figure: Comparison of terminal model parameters of different AI mobile phone manufacturers Source: Everbright Securities

2) AIPC with the most variables

According to the definition of IDC's "AI PC Industry White Paper", the six core elements of AIPC are: personal intelligent agent with natural language interaction, embedded personal big model, embedded personal knowledge base, local heterogeneous computing power of CPU+GPU+NPU, system ecology connected to open AI applications, and personal privacy and data security.

2024 can be called the first year of AIPC. As a new hardware hybrid carrier of the AI ​​application ecosystem, AIPC is another arena of fierce competition and also the scene with the most variables.

In the changing situation, we believe that the core of AIPC is heterogeneous chips, because this is the main source of differentiation. Therefore, by tracking chips, we can quickly understand the direction of the industry. In this article, we make a preliminary discussion:

• The last glory of X86? Intel, which has been abandoned by the new era, claims that it will support more than 80 new AIPC designs by the end of this year based on the immature Lunar Lake processor architecture, and the NPU computing power will reach 45TOPS. AMD plans to launch AIPC equipped with the RyzenAI300 series, and the NPU performance is claimed to reach 50TOPS.

• ARM's counterattack? Apple's upcoming M4 is undoubtedly the absolute king of AIPC chips under the ARM architecture, but it is not sold to the public, but it is enough to ensure that Apple's Mac has a place in AIPC. The biggest threat to X86 comes from Qualcomm, which is the king of low power consumption and has become even more powerful after acquiring Nuvia. The computing power of Qualcomm's X Elite NPU is 45TOPS, which is the highest computing power currently on the market. Microsoft seems to have voted for Qualcomm in AIPC.

• A spoiler? Nvidia does not seem to be willing to give up the AI ​​terminal market. Since Nvidia does not have low-power capabilities, it has begun to seek cooperation with MediaTek and is understood to be building an ARM architecture chip.

This is just the chip. There are many other brands in the PC market, such as Lenovo and HP, which will add more variables to the industry. It is very likely that we will witness that the computer industry, which is considered to be extremely mature and a sunset industry, will return to growth, and the traditional pattern will collapse.

Figure: Important AIPC chip players Source: CITIC Securities

Unlike the AI ​​computing chip competition, where Nvidia is the dominant player due to its significant lead in GPU ten years ago, the situation in AI terminals will eventually be prosperous but changeable. On the AI ​​mobile phone side, except for Apple getting a certain ticket, other variables are still large; and on the PC side, the changes may even be disruptive. In addition, the AI ​​terminal industry chain is longer and has more participants, so there are still many topics worth discussing in the future.

Putting aside the overall situation, while the snipe and the clam fight, the fisherman benefits. As AI spreads to smart terminals, chip manufacturers and terminal brand manufacturers that have been lost for several years are working hard to try to make a groundbreaking product, which will ultimately be welcomed by consumers.

Of course, now major manufacturers are competing with each other in terms of the tone. According to the current product strength, it is likely that many spectators will be disappointed in the second half of this year. The real first year of AI terminal products may not come until 2025. But no matter what, the wind of AI terminals will eventually blow to you, and you can take advantage of it or take advantage of it.

Reprinting is allowed | Business cooperation | Content exchange

Please add WeChat: jinduan008

Please add WeChat and indicate your name, company and purpose of visit

Previous recommendations