icefish| Harbor View ─ Bearing fluctuations is a required course for long-term investment

Investment management of Dongfang Harbor

Us stocks fluctuated greatly in April: since the Nasdaq hit a new high in MarchIcefishThe maximum pullback in April was as high as 8%, and closed down 4% in April.Icefish.41%; the FNGS index, which represents the top 10 AI stars, retreated as much as 10% in April and finally closed down 2%.Icefish.77%.IcefishWe have observed that in addition to the weakening expectations of interest rate cuts caused by rising inflation and the emotional shock caused by the escalation of war in the Middle East, there are also doubts and disagreements about whether the AI market will end due to a lack of fundamental support.

The market doubts about AI investment, the core is that: in addition to Nvidia has considerable profits and growth, other areas, especially AI applications, is still ineffective, business model is unclear, revenue and profit is negligible, now the huge amount of AI capital expenditure, short-term do not receive reasonable returns, will not be sustainable.

In this regard, we have made a preliminary analysis in previous monthly reports. Based on the observation of many facts, we have also made several important inferences about the development of AI. These inferences are the important support for us to firmly invest in the face of doubts and differences. These inferences were further confirmed in the first-quarter results of the industry and listed companies in April.

Corollary one: the "scaling law" has been proved to be still valid, and the iterative evolution of computing chips is also accelerating, so the tech giants have been trapped in the "AI arms race". No one wants to be thrown out of the car in the new technology cycle, and will continue to invest in AI business in the short and medium term.

This is one of the important reasons why we invest in computing, cloud computing and AI applications. In the past first-quarter results, we not only did not see an abrupt stop in capital expenditure, but also continued to improve. In the first quarter of this year, Microsoft's capital expenditure reached a record high of $14 billion, up 79% from the same period last year and 22% from the previous quarter, while forecasting a 50% month-on-month increase in spending in the second quarter. Google's capital expenditure also hit a new high of $12 billion in the first quarter, up 91 per cent from a year earlier and 9 per cent from a month earlier, while full-year capital expenditure rose to $42 billion, up 30 per cent from a year earlier. In its earnings report, Meta raised its 24-year cap on capital expenditure to $40 billion, up 43 per cent from last year. In April, Tesla reiterated that he would invest US $10 billion in AI this year, five times the total of the past eight years.

As a matter of fact, the ratio of the rapid growth of capital expenditure in the AI industry to the current income is still at the historical average. Huatai Securities also reviewed the growth of capital expenditure in the cloud computing business over the past 20 years: historically, in addition to structural adjustments or occasional macroeconomic shocks, with the continuous growth of cloud business revenue, the capital expenditure of technology giants is increasing year by year. According to the latest results, the cloud computing business of the three cloud giants, from Microsoft to Google to Amazon, has stopped the decline in growth and achieved an increase in growth in the past three quarters. Microsoft's cloud Azure growth rate has increased to 31%, Google cloud GCP growth rate has increased to 28.4%, and Amazon cloud AWS growth rate has also increased to 17%, making it the fastest way to realize AI applications. In this business revenue growth momentum, AI capital expenditure continues to grow, it becomes a matter of course.

Even if we take a longer-term perspective, AI capital expenditure still makes sense. According to the forecast of Coatue in April, by 2030, assuming that the number of GPU chips matched for AI business in the world increases from about 4 million at present to 25 million, it is estimated that about US $1.2 trillion will be invested in capital expenditure each year. If you factor in a 25 per cent return on investment and a 50 per cent EBITDA margin, you would need to generate about $3,000bn in AI business revenue a year, which is only 2 per cent of global GDP at that time. From the point of view of AI reducing costs and increasing efficiency, the capital expenditure of $1.2 trillion is equivalent to only 3 per cent of global payroll spending. So, considering that AI technology will sweep every industry in the world and create new industries such as robotics, autopilot and XR, the return on investment of AI capital expenditure is completely within a reasonable range.

Inference 2: with the sharp decline in the reasoning cost of AI in the second half of 24 years, and the upgrading of new models such as GPT5, the penetration of AI applications will be greatly improved.

In fact, AI applications have spread across all walks of life. We can try to divide the current AI applications into the following five categories, we will find that their commercial realization methods are also very clear, but due to the high cost or limited AI performance, the penetration rate is still very low.

Natural language interactive application: the large language model has significantly reduced the difficulty of human-computer interaction. Hardware and software that were difficult to use in the past, such as image modification, video modification, programming, advertising production, search engine, etc., can now be easily operated with the help of Ai dialogue interface, such as Adobe, Github Copilot, Meta's Advantage advertising creative tools and Perplexity's new search engine. The function of natural language interaction can theoretically be used in all computer interactions to achieve commercial value by directly increasing the ARPU value of software and hardware and expanding the user base.

Content generation applications: generative models that can be used to generate code, music, text, pictures, videos and even virtual characters, such as Canva, Dall-E, Notion AI, Sora, Teams Copilot and CharacterAI. At present, the logic, planning and accuracy of generated content need to be improved, but in theory, all human content generation products and services can use this AI tool to greatly improve output efficiency, while AI itself can be paid by subscription or usage, and the model is relatively clear.

Recommendation engine application: accurate matching of people and goods has always been an important part of business activities, while the large language model in content customization, user portrait cognition, accurate content recommendation and advertising precision, all make the original advertising recommendation more intelligent. This involves the core business of the Internet, such as social networks, search engines, content communities and so on. By improving the user base, advertising efficiency, advertising volume and advertising unit price, AI recommendation engine can realize the value well.

Intelligent agent application: liberating human beings from some tedious and repetitive tasks is one of the ultimate goals of AI applications. To this end, according to Professor Wu Enda of Stanford, AI not only has the ability of multimodal generation, but also needs the ability of planning and reflection, the ability of using more tools, memory and multi-agent cooperation. At present, the AI application that can really be called an intelligent agent is Tesla's FSD (supervised) autopilot software. It is believed that such applications will continue to emerge in the future and gradually become the mainstream application mode of AI applications. Similar to FSD, this kind of application can be realized by subscribing, authorizing and charging agent service fee.

MaaS cloud service applications: the above four types of AI applications need to be deployed and applied through cloud computing. Model as a Service (MaaS) is becoming the latest growth pole of cloud computing after IaaS, PaaS and SaaS. Cloud vendors such as Azure, GCP and AWS have also become "distributors" of AI capabilities, realizing fees according to the distributed AI capabilities and the hardware resources consumed.

Among the above five types of applications, only MaaS cloud services and recommendation engines have shown initial success in reducing costs and increasing revenue. Other types of applications, even Microsoft's Copilot business, are still negligible in terms of penetration and revenue implied in this quarter's results. This is true, and that's one of the reasons why we think that the AI investment and application cycle is just beginning, not coming to an end.

However, with the off-line of GB200 products and the popularity of model distillation, the reasoning cost of the above applications will be greatly reduced in the next six months. In our previous report, we have explained how GB200 achieves a 15-30-fold performance improvement over H100 on the inference side. As for model distillation, Llama3 released by Meta in April shows us this very well: Llama3 uses 1.5 trillion tokens for training, uses 48000 H100 chips for six months, and obtains a model with 400 billion parameters, whose performance is expected to be the same as or even higher than the level of GPT-4 with 1.8 trillion parameters, and the low number of parameters greatly reduces the unit cost and power consumption level of the model. This shows us the remarkable effect of model distillation on the compression reasoning cost in addition to the scaling law to improve the model performance.

In addition, with the release of the new model GPT-5, we boldly predict that the evolution direction of GPT-4 is "multimodal" and that the evolution direction of GPT5 is likely to be the improvement of "agent ability". At the Meta earnings conference in April, Zuckerberg said bluntly when asked the last question about the evolution of Llama: "I think the next stage of these things is to deal with more complex tasks and become more like agents, not just chat robots." For a chatbot, you send it a message and it will reply to your message. It is almost an one-round communication. What the agent has to do is that you give it an intention or goal, and then it will do it, and it may execute a lot of queries in the background to help achieve your goal, no matter what the goal is to study online. or finally find the right thing you want to buy. " In a previous lecture at Stanford University, Wu Enda also showed a picture (below) that in programming tasks, a large model of using tools, planning and reflection, and multi-agent cooperation has been improved by leaps and bounds on the basis of GPT3.5 and GPT4, and the ability of GPT3.5-based models does not significantly lag behind GPT4. Just like a Tsinghua graduate and an ordinary college student, although there are differences in ability, if you teach them the work flow of the company, give them useful tools, and work closely with other colleagues in the department, the results of their work will be significantly improved, and the results may be very good. GPT4 launched the "Primary memory function" in April, and we expect more agent capability modules to be available next, significantly improving the performance of the large model.

When the cost of model reasoning is greatly reduced, the performance of the model is also greatly improved, and the permeability of the application should become wider. More human-computer interaction will be replaced by natural language, the content generated by the model will become more controllable and accurate, the model recommendation effect will be more accurate, more important is that more complex tasks can be "automated", MaaS cloud AI and EdgeAI will also develop synchronously. FSD is a good example. When FSD entered the V12 version, the reasoning cost and running power were also greatly reduced, and the model ability was greatly improved. After a 20-day trial of a comprehensive North American promotion in April, Tesla reported that 900000 car owners had generated 300 million miles of driving data in 20 days using FSD V12, equivalent to 30 per cent of the total in the past eight years. We expect that FSD penetration may also exceed 50 per cent in North America after the trial period.

Corollary 3: with the gradual increase in application penetration, unlike the Freemium model of the Internet, the AI business model based on subscription fee / charge by usage will increase revenue or profit margin at the same time.

The essence of the business model in the Internet era is the "flow realization model", which attracts and accumulates a sufficient number of users through a variety of free content and services, and then realises users' attention as advertising resources. or from a small number of users to earn revenue from virtual products called "value-added services". Under this model, the accumulated traffic comes before the revenue and profit, so even today, a small number of well-known Internet companies are still on the brink of profit and loss.

But the AI business model may be different. Whether it is 2B or 2C Magi AI "enabling" business nature, it is more like the business model of software, mainly through "subscription, authorization fee, charge by usage" these three models for commercial realization. There may be a free trial phase in the early stage, but the overall charge is more direct and higher. We can see that almost all AI applications, such as Gpt4, Copilot, Perplexity, CharacterAI, Midjourney, Tesla FSD, Github, Adobe Firefly and so on, are paid for, but the financial income is not significant because of the low penetration rate. Even Google, the world's largest advertising company, is considering switching to paid use of its free Ai product, Germini, in an attempt to explore the transformation of the business model in the AI era.

Therefore, we believe that the market's worried "AI business model is not clear, revenue and profit is not significant", there is a good chance that it will disappear quickly with the increase of AI application penetration, rather than a time lag like in the Internet era.

Inference 4: the current AI valuation is above average, the memory of the Internet bubble makes the market do not see the rabbit does not cast an eagle, volatility will always be relatively large, but can not change the direction of the times.

At present, several AI technology companies are valued at above-average valuations, which are not cheap but not particularly expensive. Nvidia's dynamic valuation is 30 times, the average valuation of the past five years (2050); Meta's dynamic valuation is 20 times, which is above the average valuation of the past five years (1025); and Microsoft's dynamic valuation is 34 times, which is at the top of the valuation of the past five years (15535), but Microsoft is about to enter fiscal year 25 and will switch to an average of 28 times at a 20 per cent growth rate.

Under this valuation state, every earnings season will start with worry and suspicion, and end with a surge and a slump. The market is highly short-term and long-term games are fierce, and options trading is active. However, this surging market volatility has not changed the direction of the trend of the times. Most of those companies that have plummeted after the earnings report will prove themselves again and gradually recover lost ground. Google is a good example. After the financial results of the past two quarters plummeted, Google was finally able to recover its lost ground; and this quarter, Google Cloud continued the upward growth trend of Q4 growth and further accelerated growth, with a year-on-year increase of 28%. At the same time, Google Cloud's operating profit increased from $190 million in the same period last year more than quadrupled to $900 million, which surprised the market. The CFO said: "The benefits of artificial intelligence to customers support our growth across the cloud computing space." The stock price jumped after this earnings report.

This kind of valuation and market performance are the common manifestations of market rationality and irrationality: rationality lies in being guided by facts and certainty, and not casting an eagle until you see a rabbit; irrationality lies in the fact that the market is too eager to see business results and investment returns, resulting in the game. The components are too large. Long-term investors should return to the long-term fundamentals of the company itself and rationally bear market fluctuations.

The above four inferences are our judgments on the current fundamental risks. As for the risks of war and interest rates, we have also analyzed them in the previous report, so we will not repeat them here.

The bull market often falls sharply. Compared with joy and joy, doubts and differences are signals that the market is moving forward healthily. The long-term investment competition is "who can see far, see accurately, dare to take heavy positions, and persist". In this context of the times, it must also be accompanied by doubts and differences; and learning to distinguish between primary and secondary contradictions and avoid permanent losses, endure temporary fluctuations is the only way to accompany the long-term growth of great enterprises and harvest the fruits of the times.

disclaimer

The market is risky and investment needs to be cautious. Under any circumstances, the information in the article is for readers 'reference only. The companies mentioned in the article are only to explain the logic of the industry. All content does not contain any investment advice. Readers should not rely solely on this article to replace personal independent judgment. Dongfang Harbor assumes no responsibility for investment losses, risks and disputes caused by the use, citation and reference of the content of this article.