Why we believe Nvidia can reach $5 trillion
Just how big AGI is is the biggest non-consensus today.
This research is a summary of the thinking of the Shixiang team on AI investment in the past six months. It answers the question of "how to participate in AI investment today" from the perspective of primary and secondary integration. AI is not only the core of technology investment in the primary market, but also has become a driving force for the growth of technology stocks in the secondary market. In the early stages of each round of technological revolution, start-up teams pursue the most original innovation under new technologies, and large companies are the first to capture the value.
Shixiang launched the AGIX Index to more quantitatively reflect the development trend of AI by tracking the progress of these large companies in benefiting from AI. The AGIX Index first screened from thousands of US-listed technology companies to about 50 companies, and adjusted the number and weight as the industry developed, helping the market feel the flow of AI value between different fields and sectors.
When the iPhone was launched in 2008, Apple's stock price rose from $8 to $10. Today, Apple's stock price is between $230 and $240. In 16 years, Apple's stock price has risen 15 to 16 times. When a major technological revolution comes, the long-term return on investment is very considerable. Considering that we are still in the early stages of the AI revolution, now is the best time to participate in AI investment. This is why we firmly believe that Nvidia's valuation is far from saturated and will reach a market value of at least $5 trillion. In addition to Nvidia, in this study, we also conducted fundamental analysis on important companies covered by the AGIX Index, such as Amazon, Apple, Broadcom, and Servicenow.
01
Why you should bet firmly on AI
Investment logic 1:
Believe in Power Law
In the past 10 years, 1% of the top technology companies in the US stock market have contributed 99% of the returns of US technology stocks. In 2013, there were about 1,700 technology companies in the US stock market, with a total market value of 4.2 trillion US dollars. Now the total market value of US technology companies has risen to 20 trillion US dollars, but the top 1% is contributed by about 10 companies, from 2 trillion to 18 trillion US dollars, which is basically equivalent to the top dozen companies contributing 99% of the increase. These companies have undoubtedly seized the two key technological changes of mobile Internet and cloud computing. This also determines the strategy of Shixiang: only focus on betting on the top companies in the big track, and bet heavily. Shixiang has also invested in a number of the best top unicorn companies in the primary market, and their total valuation adds up to more than 600 billion US dollars.
There is no doubt that the next generation of leading technology companies will be born in the field of AI, and power law has also been verified in the field of LLM.
Investment logic 2:
AI-native is the technology investment paradigm for the next 10 years
Looking back at the development of the LLM field over the past two years, AI will definitely be the core driving force for future technology investments.
Since the release of ChatGPT, it is clear that AI has driven the growth of US technology stocks. Mag 7's growth in 2023 actually exceeds that of the S&P 500: each company in Mag 7 has its own AI layout and has extremely high potential to fully benefit from this wave of AI, especially Nvidia.
We also compared several indices, and since the release of ChatGPT, the indices with higher AI content have grown better. Even if QQQ and the IT sector of the Nasdaq 100 were separated out, these indices would perform better.
We recently had a strong feeling: today it is difficult for us to judge whether AI startups can subvert giants. For example, mobile phones are still our most core devices in the next 3-4 years, and many AI consumer devices today cannot replace mobile phones. Secondly, AI startups are still relatively expensive, and the business model is not as good as that of previous advertising platforms, and the giants are also well positioned.
Therefore, we believe that the first wave of narratives will benefit the giants, and mature technology companies in the secondary market have great opportunities. From today's AI investment narrative, the leading companies under the logic of large infrastructure are the most worthy of investment. The prelude to the new generation of AI-promoting companies has not yet officially begun. In order to better quantify this trend, we have made a tracking index AGIX Index for AGI. Since the release of ChatGPT, it has performed the best compared to several other technology indexes. We also hope that AGIX can become a tool for everyone to surf the future golden decade of AGI.
Investment logic 3:
AI will lead to a reconstruction of business models
In the past two years, the Shixiang team has focused on doing a lot of research on large models, and we have invested at least 10,000 hours in large model research. In recent months, we have also been conducting research on large model application scenarios. We have a strong feeling that many first-tier AI companies in Silicon Valley are innovating in AI features, but few companies have achieved a complete product commercial closed loop, and they may be proofing for old companies.
Even big companies like OpenAI are under the shadow of giants: GPU is limited by Nvidia, infra infrastructure is limited by Microsoft, GPT enterprise market also depends on Microsoft's sales, and to C end is actually a feature on iPhone. Therefore, we believe that in the AGI chain, many fruits of ToC will be picked by Apple, while many fruits of ToB will be picked by Microsoft.
A few examples we particularly like:
• Adobe was just a traditional software company worth several billion dollars before it switched to the cloud, but after switching to the cloud, its business model, market size, and growth rate all changed, and it became a company worth nearly 300 billion dollars;
• In the last wave of computer vision transformation, Hikvision shifted from selling equipment to selling systems, and its business model became more recurring, with higher profit margins and valuation multiples;
Therefore, we believe that AI will definitely bring about similar stories, but the curtain has not really been opened yet. We are looking forward to AI changing the business models of more companies in traditional industries.
02
Key judgments on AI investment
Judgment 1:
We don’t know how big AGI is.
At this stage, there is no consensus on the definition and understanding of AGI. The best definition we have heard is that AGI can surpass 90% of experts in 90% of industries and complete 90% of economically valuable work. In fact, these three 90% are still very radical, but they represent a vision.
In addition, AGI is not entirely a business issue, it also has the attributes of scientific discovery. Behind it is the endless spirit of research and discovery, constantly exploring new capabilities. In more abstract terms, how does AGI use energy and chips to produce intelligence? In the future, the competition among major companies will be about how efficient and capable the output of intelligence is, and whether it can produce hundreds of millions of new labor forces indefinitely?
Historically, technological progress is the most critical factor in creating incremental value. An interesting phenomenon is that the market value of the leading companies in each wave of technological change in history will be increased by one zero compared to the market value of the leading companies in the previous wave. 10 years ago, it was very impressive to invest in a unicorn with a valuation of 1 billion US dollars, but today, many companies have become unicorns within half a year, or even have a valuation of 1 billion US dollars at birth. Around 2010, Apple's market value was only more than 200 billion US dollars. At that time, we would never imagine that there would be a trillion-dollar company in the world, just as we would never think that there might be a 10 trillion-dollar company in the future.
During the AI investment boom in 2016, Nvidia’s market value grew from $20 billion to $100 billion, but we could never have imagined that Nvidia would reach $3 trillion today.
Therefore, the biggest non-consensus in the AI field today is that we actually don’t know how big AGI is. Previously, software products were priced based on the number of people, such as $20 per person, but in the future, they may be priced based on the output. For example, if the incremental value created by AI for my work is $10,000, then it is reasonable for AI technology to take 5-10% of it. In fact, e-commerce and advertising platforms are already paying based on results. We believe that future work tasks are likely to be paid based on results, which is equivalent to taxing the incremental GDP.
Judgment 2:
AI is the key force driving global growth in the next 15 years
AI and AGI will definitely be the strongest driving force for GDP in the next 10-15 years. Combining Sam Altman’s previous views, we also have a prediction: AI has the opportunity to double the global GDP in 10-15 years, from today’s $100 trillion to $200 trillion, but the direct correlation between the AGI wave and today’s $100 trillion GDP may be less than 0.1%, which is still a long way from unlocking 90%.
Another way to calculate is: the big model empowers or subverts the group of "knowledge workers", which is simply the 1 billion white-collar workers who work in offices around the world every day.
Historically, agricultural mechanization has eliminated the need for 90% of farmers to farm. We believe that this wave of AI technology can basically automate 90% of the daily work of these white-collar workers. Frankly speaking, the work of most knowledge workers is relatively simple and mechanically repetitive. So, just as the safety level of autonomous driving has exceeded the average level of drivers in many cases, if 300 million white-collar workers can be "created", the cost of electricity and chips in the future can replace the annual salary of more than 30,000 US dollars for each white-collar worker, which is an income of 10 trillion US dollars. If it is to correspond to the market value, this number must be multiplied by 10.
Subjectively, we are more looking forward to AI creating Li Bai and Du Fu. On average, there may be one genius like Li Bai and Du Fu in every 10 million people, and AI is also a probabilistic model. As the model capabilities are unlocked, if the aesthetic and creative level of AI can be fundamentally improved, Li Bai and Du Fu can be created in larger quantities, and this will become an era of new culture and creativity.
Judgment 3:
LLM is in the early stages of major infrastructure construction
Many people are concerned about what stage AI is at now. The Shixiang team has a deep understanding that the global AI big model is still in the early stages of large-scale infrastructure construction, and is also the early stage of human AGI infrastructure construction.
Why is infrastructure important? In the past, China's 4G, 5G and smartphones were all a kind of infrastructure. Only after the completion of telecommunications infrastructure did the explosion of short video and mobile payment applications occur; highway and railway infrastructure brought about the explosion of e-commerce and express delivery; and urbanization infrastructure brought about the explosion of takeout and local life consumption.
Some people may ask why AGI applications have not exploded yet. Our answer is also very clear: because the computing power infrastructure is not enough, algorithms and data have been waiting for computing power infrastructure. The global GPU consumption time in 2023, if calculated based on a base of 200 million people, is only 1 minute per person per day on average. We believe that it is very likely to penetrate 3-5 billion people in the future like short videos, consuming 1 hour of computing power per person per day, so we are still in the early stages of AGI infrastructure, and AI is a supply-driven market.
This year, the first-tier large-scale model companies in Silicon Valley already have 32,000 H100 fully interconnected GPU clusters. Together with the small clusters they have deployed, the total number of GPUs at the end of the year has exceeded 100,000. In the competition landscape next year, the first-tier entry ticket is a single cluster with 100,000 cards fully interconnected. This number is not exaggerated at all, and it is a single fully interconnected cluster. It is very likely that in 2027-2028, 2-3 companies will build a $100 billion supercomputer, which will have the same impact as the Manhattan Project and the Moon Landing Project.
Many people are concerned about what the input and output of this computing cluster will be? Its output is actually unlimited intelligence and ability improvement or higher-level productivity, which will also have a great impact on global geopolitics. The essence of the Opium War is that the industrial society crushed the low-dimensional agricultural society from a high dimension, and the Gulf War at the turn of the old and new centuries was also the information society crushing the traditional industrialized forces. AGI has reached such a new technological revolution today, and its impact may exceed all previous technological revolutions. AI is advancing very fast. In the past year, the speed of AI's progress has exceeded the development of human history for thousands of years.
Many people are discussing whether the scaling law will continue and whether the model will continue to grow. Our view is that the scaling law has not yet failed and the marginal benefits have not diminished, so the model will continue to grow in the next few years. Today, the largest model has 1-2 trillion parameters, and in the future it may reach 100 trillion parameters, which is equivalent to the number of neurons in the human brain.
As models get bigger, people have underestimated a trend: models are getting smaller very quickly. OpenAI has just released the GPT-4o mini model, and the knowledge density of small models per unit parameter is also increasing rapidly. Maybe within a year, we will be able to run a model with the capabilities of GPT-4 on mobile phones and computers.
The trend of small models on the end also leads to another key judgment: in addition to the data center infrastructure represented by NVIDIA, the consumer-side infrastructure represented by mobile phone manufacturers such as Apple and the mobile phone industry chain is also a major theme of large-scale infrastructure and is as important as data centers.
Whether it is data center infrastructure or Apple mobile phone infrastructure, almost 100% of high-end chips are produced by TSMC, so TSMC will continue to benefit from these two waves of infrastructure.
Judgment 4:
AGI infrastructure is an engineering problem.
It can be solved by investing money and time
AGI infrastructure is an engineering problem. The artificial intelligence industry has more unsolved problems than solved ones, because scientific problems are often not immediately clear, but AGI infrastructure can be solved by investing money and time.
Take the training of GPT-4 as an example. Currently, training a GPT-4 requires at least 8,000 H100s of effective computing power, which is close to a 10,000-card cluster. A 10,000-card cluster is a standard configuration. If you buy the cards yourself, the price of each H100 is close to 30,000 US dollars. Add the peripheral equipment, and the hardware cost alone will be 300 million US dollars. Of course, you can also rent it. If you rent H100 for a year, it will cost about 150 million US dollars.
Why is the 10,000-card cluster project so difficult? This is related to the characteristics of GPUs. Traditional CPUs perform serial computing, and a broken CPU will not affect other CPUs' tasks. However, GPUs perform parallel computing. If one of the 10,000 cards breaks down or the processing speed of one card slows down, the task progress of the remaining cards will also slow down. This requires us to quickly locate the faulty card, which is difficult. In addition, the 10,000-card cluster basically breaks down every day, and there is a failure rate every day, because the failure rate of the GPU itself is very high.
The second is the energy problem. The power consumption of each H100 is very high, about 1,000 watts. Assuming that GPT-4 uses 8,000 H100s for 100 days of training, it will require 26 million kWh of electricity, which means the power generation of the Three Gorges in one day, or 5% of Shanghai's daily electricity consumption. In addition, the peak and trough demand for electricity in model training is obvious. Some tasks are very computing-intensive and naturally consume a lot of electricity. If they are not well planned, it may affect residents' electricity consumption. No one expected that AI would develop so fast before, and it often takes about 2 years to rebuild a power station. In addition, many regions have energy and environmental protection policies, so building a power station will be slower than we imagined.
We have been talking about the power consumption of 10,000 cards. The annual power consumption of 100,000 H100 clusters is about 1.24 billion kWh, which is about 0.8-1% of the annual power consumption of Shanghai. Now AGI data centers are mainly in the United States. The total annual power consumption in the United States has been relatively stable in the past 20 years, about 4 trillion kWh, of which data centers may use 200 billion kWh, which is about 5%. However, the media predicts that by 2028, the power consumption of data centers in the United States will increase to 670 billion kWh, which means a 3-fold increase in 5 years, from 5% to 16%. This is actually a big challenge for the United States, whose manufacturing and infrastructure capabilities have obviously deteriorated.
Judgment 5:
In the next 2-3 years, we will see AGI in the coding field
Another point that the Shixiang team often says is that "AGI is not achieved overnight." Its key word is "progressive unlocking." As mentioned earlier, draw a mountain climbing route map. Every time the model capability increases a little, some new scenarios will be unlocked and some new applications will be generated.
At present, search is the biggest killer app in the early stage of LLM. GPT-4 has been out for more than a year, but AI applications have not exploded yet, and the results are relatively boring. But if we put aside the applications of large model companies such as ChatGPT, the only company I can think of that has been invested by mainstream VCs in Silicon Valley, has already run out and developed to a certain valuation is Perplexity.
But if we look at the time dimension, we think we will have a chance to see AGI in the field of coding in the next 2-3 years, that is, a programmer who is very good at doing various tasks. Because the logic of the software is clearer, it is easier to get feedback from 0 to 1. There is a good analogy: in the past, the Internet was used to search for information on web pages, which was called a search engine, but in the future, coding may be called a task engine, which is used to solve tasks with economic value.
The software we use now is developed and executed by a large development team after the product manager defines the head requirements. Just like making a movie, the director needs to issue a definition first, and then the whole team works together to shoot. In the future, long-tail or personalized needs will not need to be developed by the team, and AI combined with various agents can solve them. This scenario has been repeatedly realized many times in history. For example, the cost of making movies is very high, but after the outbreak of short videos, everyone can make their own movies.
In the summary released at the beginning of this year, we actually proposed that "big models are the Moore's Law of a new era", which is divided into two main lines: the first main line is the evolution of intelligent capabilities, and the model level will be improved by one generation every 1-2 years; the second main line is that the cost of the model will drop very quickly, and it will drop by more than 10 times every 18 months. Now many developers use the best models to do PMF when starting a business, but soon they can use cheaper models to reduce costs. One or two more generations of development will likely bring more AGI applications.
03
How to trade in the secondary market
Capturing AI Alpha?
When a big technological revolution comes, the long-term investment returns are very impressive: the iPhone was launched in 2008, and Apple's stock price rose from $8 to $10 that year. 16 years later, Apple's current stock price is between $230 and $240. In 16 years, Apple's stock price has risen 15 to 16 times. Considering that we are still in the early stages of the AI revolution, now is the best time to participate in AI investment.
In recent weeks, the hardware sector has experienced some pullbacks, and the market has begun to discuss how to plan AI investments in the future. As for the reasons behind this pullback, we generally believe that there are two aspects:
• In terms of logical trends, a considerable number of opinions in the market have begun to worry about the monetization of AI;
• Another very important point is that the United States has entered a stage of style rotation.
From the perspective of fundamental trends, we believe that there are actually no particularly big changes. Take Meta as an example. In its advertising model based on GPU algorithms, every $1 invested by customers can generate $3-4 in revenue, so the revenue gain brought by this GPU recommendation advertising algorithm is actually very good. We have also calculated Microsoft's GPU payback period before, which is about 3 years, which means that it can bring about a 30% cash return each year. From the perspective of large Internet companies, this return on investment is not bad, and there is no doubt about it. This is why these companies are constantly raising their capital expenditure guidance this year. In our research and exchanges with ServiceNow and other companies, we can also deeply appreciate that the ROI of AI investment is very high.
However, from the perspective of enterprise users, the bottleneck for them to adopt LLM is not the ROI issue, but the need to ensure that the model meets the requirements in terms of infrastructure, network, customer privacy and accuracy. Enterprise users often consider three issues when adopting LLM:
• Is the accuracy high?
• Can data security issues be resolved?
• Will there be compliance requirements in various aspects in the application scenarios?
In fact, from a practical perspective, many companies in the United States have begun to deploy LLM internally. For example, for an internal IT question-answering system, if a large company has an IT problem, in the past it might take five engineers to ask all the questions clearly, then find the problem in the system and fix it. But with the use of LLM, the manpower of these five engineers is completely replaced.
With the further development of the model capabilities in the second half of the year, especially in logical reasoning, multimodal capabilities, and agent capabilities, the improvement of model capabilities will also unlock more application scenarios. We believe that companies are also constantly trying various scenarios, and we expect to see more application terminals appear next year. So from a fundamental perspective, we believe that AI is still developing, and we are also very much looking forward to the next two years.
In the short term, there may be a phenomenon of sector rotation, and uncertainties brought about by external events such as the US election will become the cause of fluctuations. Therefore, our view is that the fundamentals are still developing, and these short-term factors are more market factors, which are non-fundamental factors.
In secondary investment, Shixiang chooses to focus on five main investment lines, namely computing power, cloud infrastructure, end-side, interconnection and software. Combined with our surveys and research, we will also dynamically adjust our layout in these five areas according to the progress in the field of AI.
At present, we believe that computing power and cloud infrastructure are still the bottlenecks that restrict the development of AI. The demand for computing power is still very strong, so we have a relatively heavier layout on computing power and cloud infrastructure. In contrast, the impact of AI on software is not very clear, so our layout on software will be relatively light for the time being. However, it is expected that in the next two years, we will see more and more obvious progress in terminal applications and AI software, so we may further increase the layout on the terminal and software sides in the next two years.
Investment Theme 1:
The world needs more computing power.
Cheap computing power is always a scarce commodity
As mentioned earlier, there are currently an average of 200 million people in the world who use GPU time for one minute per day. If we follow this line of thought, assuming that in the future 2 billion people use GPU time for one hour per day, the demand for computing power may increase 100 times or even 1,000 times.
There is a mainstream view that computing power is waiting for the improvement of the model, and the development of the market requires an explosion in the application side. However, we believe that this view is putting the cart before the horse. From the perspective of the Shixiang team, the model has been waiting for cheaper computing power, and cheap computing power is always a scarce commodity.
The lower the computing power cost, the more economically feasible algorithms will be. Taking the development of GPT as an example, GPT-5 requires 50,000 H100 clusters to train for 6 months, so if each generation of models is to be successfully trained, it takes 6-9 months. The iteration speed of the model is very slow, and the main bottleneck restricting the development of the model is actually computing power. But if we use the next generation of NVIDIA B100 cards to train GPT-5, its training time may be greatly shortened to only 1-2 months, and the iteration speed of the model can be greatly improved.
In the medium and long term, we see three factors that will further drive the demand for computing power.
1. We are still in the early stages of the scaling law, and the capabilities of the model will continue to improve as the model parameters increase. Senior executives of leading companies such as OpenAI, Microsoft, and Anthropic have repeatedly mentioned in interviews that the scaling law is far from reaching its boundaries, and their products are still in the process of catching up. The computing power arms race is still in full swing.
2. Multimodal models, especially the release of GPT-4o mini last week, will further unlock more application scenarios, such as carrying cameras with you, providing real-time AI assistant functions, etc. Multimodal models can also further replace positions with high labor costs, such as doctor visits, lawyer consultations, sales, etc. These positions will likely be replaced by AI in the future, so the economic value brought by AI will become higher and higher.
3. We can make some bold assumptions. The current global defense spending is about 2.4 trillion US dollars per year. If the big model can further unlock new capabilities in the field of images and videos, assuming that 3% to 5% of defense spending is used for AI or information-related investments each year, there will be a marginal increase of 100 billion US dollars.
Taking all factors into consideration, we believe that the world still needs more computing power.
Top Pick: Nvidia (NVDA)
Nvidia is still the stock we value most in our investment logic of computing power. Although its stock price has risen a lot, we believe that the valuation is far from saturated. Nvidia’s current market value has reached 3 trillion US dollars, but our psychological price for Nvidia is at least around 5 trillion US dollars.
There may be a view in the market that due to the rapid catch-up of competitors, Nvidia's moat is also narrowing at the same time, but Shixiang's judgment is: reasoning has higher system requirements, and Nvidia's system moat is actually getting wider and wider. This is the prerequisite for Nvidia to achieve 5 trillion.
We are bullish on Nvidia for three reasons:
• In terms of market size, we believe that the entire acceleration chip market will further expand and has the opportunity to reach a scale of US$300-400 billion;
• Nvidia has a trinity of competitive advantages from chips, interconnection to software. In the short term, we do not see any competitors for Nvidia;
• NVIDIA has a very complete product matrix. At this year's exhibition, NVIDIA clearly announced its product roadmap for the second half of 2024, 2025, and 2026. We can see that it is gradually shifting from selling chips to developing mature systems and software. Not only is the roadmap clear, but the product strength is also improving year by year.
Based on our previous bullish view on Nvidia, our logic on “why Nvidia can reach a market value of at least 5 trillion” is:
In the next five years, the data center infrastructure will reach a market size of 2 trillion US dollars. The 2 trillion here refers to the scale of the entire IT infrastructure, which covers both CPU and GPU. The current scale of existing data centers is about 1 trillion, with an annual increase of about 200-300 billion. In five years, it will reach a scale of about 2 trillion US dollars.
When the data center scale reaches 2 trillion, it will appear: the current hardware life cycle is generally about seven years. If 7 years is used as a replacement cycle, then the annual replacement rate is actually 15%. Under normal circumstances, the annual growth is probably high single digits, that is, about 7-8%. Therefore, with a 15% replacement cycle plus a high single digit natural growth, there will be a market space of nearly 20% replacement and growth every year, which will also bring a Capex space of 400 billion.
In every era in the past, the top companies accounted for 70-80% of the value, and today Nvidia's moat is very wide and worthy of the market's trust, so we believe that Nvidia can still account for 75% of the value, that is, it can reach a revenue volume of 300 billion US dollars.
Currently, Nvidia's net profit margin is 45-46%. As its scale expands, its operating leverage will be further reflected, so we believe it has the opportunity to reach a profit margin of 50%, which means data center profits of 150 billion.
In addition to data centers, we believe that Nvidia can definitely reach 5 trillion by adding autonomous driving, gaming, and the data center business it has helped many companies develop. It is worth emphasizing Nvidia's layout in autonomous driving. Nvidia may be the second largest company in autonomous driving investment after Tesla. We believe that this point is actually overlooked by the market.
Investment Theme 2:
AI is beginning to boost cloud vendors’ profits
Cloud vendors can benefit most directly from the progress and application of large models. We can already see that AI is gradually increasing its revenue for cloud vendors: in the past four quarters, AI's marginal contribution to Microsoft was 1%, 3%, 6% and 7% respectively. In many CIO surveys, we can also clearly feel that the investment willingness of global leading companies in the cloud is increasing. The investment willingness in the cloud in 2025 and 2026 is significantly higher than that in 2024. Therefore, we are very optimistic about the revenue of several major cloud vendors in the next few years.
Top Pick: Amazon (AMZN)
Amazon is our current favorite of the four cloud vendors, and we believe that the market actually underestimates or even ignores the extent to which AWS will benefit from GenAI, and therefore we are optimistic about AWS's long-term growth potential.
The Shixiang team's investment logic for Amazon is mainly based on the following two points.
• First, as mentioned earlier, AWS as a cloud vendor is also an important distribution channel for LLM. Although the cooperation between AWS and Anthropic cannot be 100% regarded as Azure and OpenAI, it is also a fairly deep cooperation. The performance data of Anthropic's latest Claude 3.5 series model is amazing. Judging from market feedback, Claude Sonnet 3.5 is a model that is comparable to OpenAI GPT-4o;
• Second, a very important part of Amazon’s revenue comes from retail and advertising, and this year and next year, the retail and advertising businesses will enter a period of rapid recovery and growth.
In summary, we are very optimistic about Amazon, whether in the recent recovery of retail and advertising, or in the medium and long term layout of large models and cloud services. Therefore, among the four clouds, our first choice is Amazon.
Going further, we can compare AWS and Azure to get a feel for why AWS's actual performance may be stronger than the market perceives.
Compared with AWS, Microsoft is better at serving large enterprises and governments, and it is more closely tied to OpenAI. However, we can also see that Microsoft and OpenAI are wary of each other. For example, OpenAI does not open certain interactive API permissions to Microsoft Azure, but AWS is more friendly to small and medium-sized enterprises, especially start-ups.
Judging from the difference in model capabilities, the two can be said to be moving forward in parallel. To some extent, although each generation of Anthropic's model may be a few months later than OpenAI, its performance is actually slightly better than OpenAI in many aspects, especially in the ability to read long texts and draw graphs. Therefore, we believe that there is no absolute difference in capabilities between Amazon and Microsoft.
Even compared to AWS, Microsoft is more closed because its goal is to promote sales of other products through AI, such as Office 365 and Windows, so Microsoft is more subjective and has a stronger tendency to subjective control. Amazon is more open. If we believe that AI will produce a new unicorn or a new disruptive company, then this company is likely to be an AWS customer at first.
Because the market has such a strong consensus on Microsoft, its PE multiple is very high, perhaps around 35 times. However, because the expectations for Amazon are not very high, and some even think that it will be damaged by AI, the multiple is not high, which is also a very big difference.
Investment Theme 3:
End-side applications begin to progress
For the consumer side, this year is the first year of AI phones. In 2008, we saw that smart phones began to replace feature phones. With the launch of Apple Intelligence at this year's WWDC, we believe that this year has entered the era of AI phones. On the enterprise side, many surveys show that corporate executives attach great importance to AI on the edge and are very willing to further invest in AI applications on the edge.
Top Pick: Apple (AAPL)
In terms of terminal layout, our current first choice is Apple. We believe that the most critical value of Apple lies in that because its layout in software and hardware is comprehensive enough, it can control the entrance of the entire traffic. Apple is the undisputed biggest beneficiary of the ToC side in the AI era.
First, from the perspective of the entry, in addition to OpenAI's GPT series of models, Apple is also in contact with many LLM companies. It has chosen a so-called hot-swap strategy. There are also reports that it is discussing new cooperation with Google. At WWDC, we can also see that Apple has clearly stated that it wants to take control of the traffic entrance.
AI Chatbot is definitely a relatively large application for Apple, but at the system level, it must and will firmly control all traffic entrances, distribution rights, and APP calling rights. This is why we like Apple very much: in the AI era, it is still the most capable of occupying traffic entrances and mind positions.
Our biggest feeling about this WWDC is that Apple has organically integrated AI into the iOS system. We can even say that on the one hand, Apple is deeply embracing AI, and on the other hand, AI is also reshaping Apple to some extent. You can imagine that in the future AI era, the mobile phone entrance may no longer require clicking on the APPs on the screen. Siri has become an all-powerful agent because it has great authority and can call up various applications to complete complex tasks and interactions.
For example, we can ask Siri to help us check a flight from Beijing to Shanghai tomorrow. According to my current calendar schedule, perhaps a flight from Beijing Airport to Pudong Airport at 10 a.m. would be more suitable. At present, we may still need to search and compare prices in several apps ourselves, but in the foreseeable future, this may become an interaction between us and the mobile phone, allowing the mobile phone to help us perform this operation.
If we look at Apple's thinking on LLM modularization, that is, using three models of different sizes in the cloud to implement tasks of different complexity, we can also feel Apple's future potential.
In addition, we believe that iCloud, as Apple's service end, has a growth potential that is far underestimated by the market. Apple's current mobile phone sales have slowly entered a relatively stable period of 200 million. LLM will first improve this situation, but more importantly, the resulting privacy needs will also drive the revenue of iCloud, the "personal cloud". The penetration rate of Apple's service end is currently only 10%. If the proportion of users subscribing to iCloud increases with the addition of AI functions in the future, then every point of penetration rate increase will contribute 3-4 points of EPS (Earnings Per Share) growth to Apple. Therefore, we believe that Apple will enter a super replacement cycle in the next 3-4 years, and its service end will also have a great predictability of revenue growth, so on the end side, our first choice now is Apple.
Investment Theme 4:
The super cycle of storage and interconnection is still unfolding
The training of the next generation of large models is based on the interconnection of 100,000 cards. According to the current training requirements of GPT-5, the cluster size we can currently use is at the level of 30,000 to 40,000 cards, and the entry threshold for the next generation of models is to reach 100,000 cards. The entire scale needs to be further expanded. Moreover, the scaling law is still advancing rapidly, and there will be further requirements for model parameters and data bandwidth. In addition, with the further popularization of LLM, we will also create more data. For example, this year's Sora model will continue to accelerate the speed of video generation. It is conceivable that video data will explode in the next few years, which will put forward higher new requirements for data storage, interconnection, and interaction.
Top Pick: Broadcom (AVGO)
In the field of interconnection and storage, our first choice is Broadcom, which can be said to be the most strategic company in the interconnection field.
The reason why we value the interconnection link is that it is the second largest expenditure item in the data center after computing. Take Meta as an example, among its 24,000 card clusters, the reuse rate accounts for 70%, and the interconnection accounts for about 25%, so the interconnection plays a very important role in the entire data center infrastructure.
Another reason we like Broadcom is that it is the main and largest manufacturer of customized chips. Broadcom has cooperated with Google on six generations of customized chips, TPU, and will start developing the seventh generation of TPU this year. We can see that companies like Meta and Byte also make customized chips from Broadcom. Therefore, in addition to Nvidia, another indispensable hardware company for many large Internet companies is Broadcom.
Investment Theme 5:
The SaaS model based on seats has shifted to a compute-based model.
Token-as-a-Service Pattern
In our initial investment decision, we also mentioned that LLM will revolutionize the business model. We believe that this trend is bound to happen in the SaaS field. The previous Seats-based software model may be transformed into a Token-as-a-Service model.
In the past, software was charged per person, because each unit of productivity was an employee, so the main way for software to grow was to serve more employees. However, in the AI era, the growth point of business or productivity may not be the number of employees, but the number of LLMs. In other words, the productivity of enterprises in the future will no longer be determined by the number of employees, but by how many LLMs the enterprise has, how many tokens the LLM can generate, and how much computing power it has. These factors will become the main model for future corporate profit growth.
So the question we are thinking about now is: Will the entire software model undergo a radical change? Will the model based on the number of people and seats be replaced by software companies that use computing as a service model?
A recent study on Salesforce said that AI will reduce the number of Salesforce seats by 10%. We believe this news confirms our conjecture: the future software model will gradually change from the number of seats to the number of tokens, but there may be more expensive tokens or cheaper tokens. For example, companies with good models may have higher token prices. We will follow this idea to make corresponding investment arrangements.
Top Pick: ServiceNow (NOW)
Our favorite company in software is ServiceNow. We believe that ServiceNow is the clearest AI revenue target among overseas SaaS companies. Currently, there are very few companies that help enterprises deploy AI and LLM, and ServiceNow is the best one among the few companies that can help enterprises deploy AI.
ServiceNow's main business is to provide various services for enterprise IT automation. After so many years of development, it has become a long-term partner of many large companies and is very familiar with the internal IT systems and data systems of many companies. Therefore, in the AI era, if many companies want to deploy LLM, the first partner they find is ServiceNow. Therefore, since last year, ServiceNow has mentioned on many occasions that its orders are very strong. Whether it is the Now Assist service based on LLM capabilities, or the services provided to solve enterprise data and internal enterprise search, they have been well received by many corporate customers and there are many related demands.
In the implementation process of LLM, a very important point is internal enterprise search. ServiceNow does the best in this regard. For example, many times people criticize LLM in the application process that the model will produce some illusions during the application process. There are two main solutions to the illusion: RAG and graph search within the enterprise. ServiceNow has launched corresponding services in both aspects, so ServiceNow is a very important partner for enterprises when deploying LLM.
Now Assist is the fastest growing product of ServiceNow in the past 20 years. This actually shows two things: first, the demand for enterprises to deploy LLM is very strong. Second, enterprises recognize and trust ServiceNow services.
Therefore, we believe that ServiceNow is one of the first targets that will benefit the most from AI in the AI era, whether in terms of orders, cash flow, location, or horizontal expansion capabilities, and is also one of our favorite targets in the software field.
京公网安备 11010802033920号