The AI craze shows no signs of abating.
Meta, which has just announced its results, exceeded expectations in terms of results and guidance, and also paid an unprecedented dividend, raising the repurchase amount to $50 billion, and the after-hours stock price **15%. At the earnings conference, company executives said that Meta's overall expectation is that more investment will be needed to support the AI business in the next few years, and this year's situation will be reflected.
In terms of AI, Meta really does it. In order to train Llama 3, it is necessary to buy 350,000 Nvidia H100 in one go, and if other GPUs are included, there will be about 600,000 H100 equivalent computing power.
Amazon, which announced its results at the same time, also had better than expected results and guidance, and its stock price rose more than 6% after hours. Among them, Amazon Cloud business revenue increased by 13% year-on-year, dispelling market concerns about the decline in demand for cloud services. Executives also said they would increase their investment in AI to reap greater business value.
As one of the most important items of AI infrastructure, the prosperity of computing power is still very high. Before you know it, Nvidia's stock price has exceeded $600, and its market capitalization has reached a record 1$54 trillion. Similarly, AMD's stock price has also hit record highs, with its market capitalization soaring to a high of $280 billion.
I remembered what Sister Mu said,It (NVIDIA) is very similar to Cisco at the beginning of the Internet wave, in the stage of infrastructure construction, hardware manufacturers tend to show more room for growth, but once this stage is completed, the market's focus will turn to software and applications.
Based on this judgment, Sister Mu sold most of Nvidia** in early 2023. As it turned out later, this operation was very wrong, and it perfectly missed Nvidia's nearly 3-fold increase in the past year.
The misjudgment of AI computing power is not limited to Sister Mu.
Are we misjudging the computing power, or is the computing power so powerful that it subverts the investment philosophy of the past?
What exactly is computing power?
Computing power is indeed the most basic and important infrastructure of AI, just like the roads, bridges and houses in our society, these infrastructures have a stage of large-scale construction, and there are also long-term low demand after saturation, which is the traditional thinking.
The question is, is computing power really to AI what roads, bridges, and houses are to humans?
We might as well think about it from a different perspectiveWhat exactly is AI?
Very simply, AI is equivalent to human intelligence, and it is also human intelligence that needs to be replaced, if it is just a neural network, it may be a software program, but if AI is given an entity, it is similar to a humanoid robot, and the whole person can be replaced theoretically.
This involves a very interesting historical record, the emergence of the human species is about 20-300,000 years ago, but the history of primates, that is, the history of human intelligence, is only 6,500 years ago.
This means,It is often a very long process for human beings to generate intelligence. But once generated, the subsequent growth is exponential and will continue forever. Since AI is going to replace human intelligence, its development should theoretically follow this law. Namely:Before the intelligence is generated, there is a long "training".
What does this "training" depend on?
Obviously,Computing powerIt's an essential part.
Altman, the founder of Open AI, said a lot at the Davos forum, and the most important one was two sentences.
The first sentence is:"The current GPT-4 has too many shortcomings, much worse than the version we will have this year, much worse than the one we will have next year, if GPT-4 can only solve 10% of the human mission at the moment, GPT-5 should be 15% or 20%".
The second sentence is:"The computing infrastructure for large-scale AI preparation is not enough."
Obviously, the capabilities of AI are still quite limited, and it is far from completely replacing humans. And from a practical point of view, the 10% that Ultraman said may be overestimated. For example, in the office field, AI is only used to generate a PPT, a chart, and a copy, of course, it is also developing rapidly, such as short production, film production, game production, etc.
It can only be said that the future is indeed worth looking forward to, but before this future really comes, the time in the middle will not be short, because it cannot escape the laws of physics. Therefore, the demand for computing power is likely to remain high for a long time.
Alternative starting point for commercialization
Speaking of which, the second question arises again:
Can the high demand for computing power really continue indefinitely? Doesn't it have a ceiling?
Let's take a look at the following diagram.
Technological revolutions such as the Internet and AI commercialization are very different. As long as there is a certain point in technological breakthrough, the starting point of Internet commercialization begins, and the curve shows that it will first rise sharply, and then become flat, until the industrial development reaches a saturation period. But AI has the potential to completely disrupt this process, and the starting point of industrialization can be very late, but once it is launched, it will be followed by exponential growth, and it will be faster and faster.
For example, Apple's mobile phone is basically finalized after the iPhone 4, and no matter how it develops, it is only a local change, one camera becomes two, and two then become three, and the result is that the shooting effect is getting better and better, but has the overall architecture of the mobile phone changed?
Not at all. The Internet is essentially the digitization of the physical world, so after the digital infrastructure is completed, there is only one subsequent work, which is to move the offline online.
Therefore, after a few years, as the penetration rate and penetration rate reach saturation, the entire smartphone industry and the entire Internet industry have entered a period of adjustment. The development logic and investment logic of these industries are very simple, as long as the basic architecture is mature, and then it is replicated on a large scale, the cost is reduced, and the experience is improved, you can lie down and make money.
But in AI, it will be much more difficult to really cross that commercialization inflection point. Take autonomous driving as an example, it has not been fully realized until now, and the most advanced Tesla FSD cannot be done, and Musk is also constantly bragging and constantly bouncing tickets, so he does not dare to say "FSD will be achieved this year".
However, no one can doubt that the commercialization potential of autonomous driving may be much greater than everyone thinks.
In other words,In the past industrial revolution, as long as the score exceeded 0, the commercial value began to explodeThe front earns more, the back is saturated, and then wait for the next industrial revolution. However, if AI breaks through 0 points, it may have no commercial value, and it will not begin to show commercial value until 60 points.
After humans developed primate intelligence 6,500 years ago, did they need less to improve their intelligence?
No, on the contrary, it increased. Human beings have established schools, established scientific research institutions, and studied various science and technology, and as a result, the improvement of intelligence has become faster and faster, the number of things that can be created has increased, and the level of civilization has become higher and higher, and finally it has entered a virtuous circle.
That is, the higher the level of intelligence, the greater the need to continue to improve intelligence.
From this perspective, it is completely different to understand computing power, because the higher the AI rises, the higher the requirements for improvement, the more difficult it is, and the more urgent the demand for improvement, without the corresponding high computing power support, it is difficult to achieve.
This is the real reason why Ultraman said that "computing power is not enough".
New logic
In the past, the investment in computing power was still based on the traditional analysis framework, such as large manufacturers grabbing orders, good competition pattern, high bargaining power, good profit margins and yields, but rarely talked about long-term demand logic, because this logic is really difficult to say, and it is difficult to see clearly at once.
The most primitive driving factor of the business society is demand, and the greater the demand and the more durable, the business can grow and grow. If you add that the competition pattern is clear and the proportion of faucets is very high, it will be even better.
Are there any businesses that fall into this category?
Moutai counts as one, the operating system counts as one, and the chip counts as one.
In the past, we have already discussed that the demand for computing power will be very large and long-lasting, which is the core reason why computing power companies continue to create new highs, and the competition pattern in the field of artificial intelligence is too good, GPUs are basically Nvidia's one-man show, AMD is just keeping up with the rhythm, and it is difficult to say that it will shake Nvidia's status, as for other big manufacturers that say they want to make chips, they haven't skimmed the eight characters yet, and they will talk about it when they really do something, and now it's still very early to worry about their competitive pressure.
No one can say for sure whether Nvidia will hit $1,000 within the year, but no one can simply deny Nvidia's momentum. When it soared a year ago, many people would think that Nvidia's valuation was too high, but now NVIDIA's dynamic PE has become 80 times, and the static PE is as high as 350 times, and now the valuation in 2024 is only 30 times.
Some companies,A lot up is not a reason for you to sell in a hurry, and a lot down is not a reason to sell.
But it depends on what the core driving force of this ** is, if it is just a false hype, something without substance, it is indeed not worth chasing, but if it is enough to affect an era, there is an industrial revolution that lasts for 10-20 years or even longer, this growth momentum will be endless.
In the industrial revolution, the electrification revolution, and the IT revolution that has affected us to this day, we have all seen this kind of ultra-long-term, ultra-high-speed growth, and some companies have experienced ups and downs in the middle, but they are still shining, such as Apple and Microsoft, which have been listed for about 40 years, and both have become the only two companies with the highest market value in the world.
The AI revolution has just begun, and getting off the bus too early because of some stock price fluctuations may avoid some short-term losses, but it may also lose the possibility of harvesting long-term value.
Conclusion
No one is sure that the market capitalization of Nvidia and AMD will one day reach the level of Apple and Microsoft, and no one is sure that their performance growth rate will remain high after many years, and no one can accurately determine how many GPUs the market will need in 10 years.
But no one can deny the importance of computing power for AI and the long-term demand for computing power for AI. So far, in the field of AI, there is nothing more valuable than computing power.
Of course, stock prices do not simply reflect the growth logic linearly, and as the chart below illustrates, even investments with a definite growth logic can have twists and turns in the middle.
Although the concept of AI has been hyped up for more than a year, from the perspective of the stage of industrial development, it is still at the starting point, in the stage of infrastructure construction, and there is still a huge application stage waiting for you later. If you really want to use the paid penetration rate of the Internet as a reference, it is probably in 1996.
Ironically, Robert Metcalfe, the inventor of Ethernet, said in 1995:"I'm going to be a spectacular supernova in the Internet, and by 1996 it will have a catastrophic crash. ”
If he had been believed at that time, the result would have been a perfect miss on a big era.
Therefore, in the process of investment, if you are lucky enough to encounter a huge industrial revolution dividend period, you can place a heavy bet after understanding the underlying logic of the industry. Although the intermediate process may have ups and downs, and it is not recommended to blindly chase high, as long as it is paired with appropriate high selling and low buying, it can be done to obtain excess returns.
Instead of following the hot spots of the market every day, it is better to bet on this kind of industrial revolution with ultra-long growth logic, and find the most valuable part of this industrial revolution.
Because sometimes, the choice is greater than the effort, and the choice does not need to be much, as long as it is correct.
Even if the demand for computing power comes to an end one day, aren't there still investment opportunities at the application level waiting for you in the future?