Deepseek is a conspiracy: AI equal rights movement

Conclusion:

1. The impact is very large and all-round, not a conspiracy theory

2. DS provides a new method, which will also have a great impact on traditional models. There is an explanation at the end of the article

3. Huawei Ascend service is launched, which is of great significance. Domestic large models can truly be implemented

4. In my personal judgment, it will intensify confrontation, and the Artificial Intelligence Act is very interesting

5. Monopoly is the beginning of backwardness, and competition is the source of innovation

Why did the launch of DeepSeek have such a great impact? During the Spring Festival holiday, most of the articles that were flooding the screen were about deepseek, but most of the discussions were limited to technology. With the actions of companies such as Microsoft, NV, and AMD, I think the doubts about Deepseek (technology & cost) can almost stop. I guess no one expected it to happen so soon... I still remember writing an article about deepseek on Monday afternoon, and someone was saying that DS had hoarded a lot of cards. In a few days, the American AI giant basically proved the credibility of DS technology and cost.

In fact, I realized this when I was studying DS last weekend. DS did not come out in one day. This time, he opened the code, and there is also a paper telling everyone how this thing is implemented, and everyone can use it to reproduce it - and indeed a team has reproduced this thing. In addition to the exaggeration in the media dissemination process, I believe that the content in the original paper is rigorous - the question now is: What kind of impact will DS have?

Conclusion: This is not a simple technical dispute, but also an open conspiracy. Its essence is about: Should AI technology be used as a tool to obtain monopoly profits, or should it be a public good that benefits global development?

1. Not just cost

The emergence of DS this time is a disagreement about the development direction of AI. Can great strength produce miracles? Scaling Law has been insisted by some people in the past two years, but now it seems that this view is basically a wrong view. In fact, in an industry that is changing and developing rapidly, where is the law that people can abide by?

New methods can achieve better results at a lower cost. It is nothing more than where the limit of this methodology is. What I believe is that this matter has just begun.......

Americans develop AI, especially from the export control of NV chips and the change of OpenAI to CloseAI. This is not a simple technology or product. Americans regard AI as a base and product that affects the country's destiny. Its most fundamental purpose is only one: monopolize the global AI technology and continue to obtain monopoly profits from the world.

In the past, when China's manufacturing industry was not doing well, didn't the United States make money in this way? Isn't the chip profitable in this way? Now it is nothing more than wanting to replicate it in AI again. The reason why this time has a big impact is that once AI can really be industrialized, its output value imagination space will be huge, which may be a volume of trillions or even tens of trillions of US dollars.

The core of the problem is two levels of problems: 1. AI chip companies may face very large risks; 2. AI is unlikely to achieve monopoly - AI equality will greatly promote the development of the global economy.

So the matter has reached this point today, and it is not a dispute between China and the United States at all. This is a naked conspiracy: Telling Americans that you have taken the wrong direction in AI. American companies will certainly follow suit in the future. After all, the capabilities and resources are there, but the final result may be that if you don’t follow, you will die, and if you follow, you won’t make any money... The high valuations of companies like NV may also be a big problem.

The logic of OpenAI may also change greatly. OpenAI has changed from open to closed. It originally wanted to make a big profit by monopolizing, but in the end it chose to close between open and closed. This may prove to be a major mistake. This thing is completely different from the operating system of MSFT. This ecosystem is very short and is provided more in the form of services. In this case, the logic of choosing to close and build your own ecosystem is wrong.

2. The underlying logic of the chip is broken

I think, when everyone is arguing about cuda or ptx, there is a voice that has been ignored: Can DeepSeek run on Ascend?

While writing this article, the official news came out: https://mp.weixin.qq.com/s/sl_N-kjouq8NRK3kcdsaaQ.

This is the wealth code after the holiday. If you can understand it, you can understand it~If you don’t understand it, there is nothing you can do~

In the article linked above, AI services based on Huawei Ascend servers are provided-this proves a lot of things, a lot of things. You should know that the 910 launched 5 years ago has been upgraded to 910C. If there is no major accident, the chip should be basically enough this time. The yield rate of chips is also expected to increase - these are all estimates, because the real yield rate data is absolutely confidential, and everyone is guessing. No one can get the real data of the production line. Those who can get the data will never dare to speak out.

The demand for chips comes from two aspects, one is training and the other is reasoning.

The emergence of DS first broke the demand for training chips. In fact, objectively speaking, the training computing power is already in excess. Now, DS may aggravate this excess. Training computing power has been the core of AI in the past two years, but it is no longer difficult to get a machine. You can look at the hourly rental of AI servers. Those who do this business know how much this thing has declined compared to the high point. Moreover, in the United States, the threshold for AI is too high, and basically only a few giants are playing with it. Now if I tell you that the computing power demand for training a model must be reduced by 90%, I don’t know how to solve the problem of excess supply of existing computing power.

Everyone basically agrees that the demand for training has peaked, but the greater space in the future comes from the demand for inference. I think this problem is not big, but the debate on the GPU and ASIC route has been going on for some time. The big manufacturers’ self-developed and Broadcom all follow this logic. But the problem is that Huawei’s 910 has also come out to participate in the battle. You have to know that the user’s logic is cheap and easy to use, and they don’t care about the underlying technology behind it. If DS’s model capabilities and Huawei’s chips can support this application, people will definitely ask: How much is NV worth?

For inference needs, for big manufacturers, taking the route of self-developed ASICs may be a better choice. In this case, inference chips are difficult to be monopolized. Several domestic companies also have this capability. Now it is nothing more than a problem of production capacity - the solution to this problem may only be a matter of time.

If the threshold for private deployment of large models is greatly reduced, the cost of cloud services facing the C-end will be greatly reduced. Obviously, it will accelerate the implementation of AI. This is beyond doubt. However, while demand is exploding, there are also many diversified supplies, and many products are more cost-effective. You have to know that the United States has just divided different countries into different levels for AI computing power... This is simply nonsense. If this is a matter at the level of national defense and military industry, it can be understood. This is obviously just a business. When this business itself cannot build barriers and needs to build barriers through administrative orders, then this barrier will inevitably be extremely fragile.

3. Without deepseek, does the logic of AI hold?

We talked about a problem earlier, that is, the success of the United States is based on the collection of seigniorage from the world-simply put, the world needs to use US dollars, so the United States can pass on the crisis through the US dollar. This is the most efficient colonial model... Does the United States rely on high technology to impose taxes on the world? There are a group of companies in the United States, such as Meta, Google, MSFT, and Amazon, which impose taxes on the world by providing services on the Internet, but this taxation right is not based on strict business standards. In other words, these American companies do not rely on their own technological strength to gain global competitiveness. For example, let's talk about a very typical example. What is the barrier of Meta? If you compare Meta with domestic tools such as WeChat, you will find that Meta may not be necessary. However, these corresponding services in China have become like this in the Chinese market, but it is difficult for us to promote such software and services in Europe and the United States, or to say that if we create a Tiktok in the United States, it will also be messed up by others.

If we look at it this way, we have reason to doubt: Even if the AI ​​revolution in the United States succeeds and is monopolized by the United States, can the United States impose a heavier "AI tax" on the world through the AI ​​revolution?

It is obvious that they can rob, but now they have to do business with you, why bother? What's more, the global taxation tools represented by AI and semiconductor chips are now being dismantled step by step. What industrial foundation do you want the United States to use to impose global taxes next? The US dollar can no longer anchor resources. In fact, the US dollar is anchored on China's manufacturing industry. A large proportion of the US dollar's global output (the United States forms a deficit) relies on the trade deficit with China, that is, manufacturing products - this is an important means for the US dollar to flow out of the United States to other countries; then the profit repatriation created by the service industry and the surplus in the capital account are important ways for the US dollar to flow back.

The United States cannot effectively reduce the trade deficit in goods. If the path of the dollar flowing to the world is cut off, other countries will easily find alternatives after losing the way to obtain the dollar. Another way for the dollar to flow back is technology (the high-tech service industry in the United States has created huge overseas profits) + finance (US stocks + US bonds). If the technology path is cut off, it will be interesting... In this way, you can understand that in this global circulation of the dollar, Chinese-made goods are actually the real foundation of the dollar. It is impossible for China and the United States to break up. An industrial producer will not suffer in the process of separation. On the contrary, technological hegemony is indeed the key means for the United States to achieve a healthy return of the dollar, but this leg may soon be cut off, and the United States, which only has finance left, is estimated to not last too long. This dollar cycle can be digitally confirmed on the US balance of payments, and can be continuously tracked. Moreover, this is an accounting identity. After the technology path is killed, it will be the most interesting. The United States has only one way to return the dollar, which is US bonds, and will eventually lose its currency credit.

There has been a lot of confrontation between China and the United States in the past two years, but now it seems that the trade war in the past five years has nearly doubled China's surplus. Now it is close to a trillion US dollars, which is the largest surplus of a single country in human history. As the technology exhibition unfolds, the challenges to the high-tech industry in the United States are basically based on defense (chips and AI are similar), but China is offensive - although the offensive process is a bit tragic, a bit like Shangganling. But now it seems that it is only a matter of time before it is attacked. The United States did not show aggressiveness in the technology war, and the offensive edge was much less!

4. Is there no one who understands in the United States?

We must first emphasize that AI is a powerful tool, but it may not be a tool to make money. The next such project should be Musk's Starlink. I have done some calculations. After China's Starlink is built, it is highly likely to imitate the current GPS and Beidou, and be directly free, or charge a very low fee, or even charge different fees in different countries. In the end, making money depends on the production of terminal equipment and some B-to-B services, while the C-end is directly free-because the leverage of this industry is too high.

Let's take a look at the American games and we will understand some of the problems here.

We all know that Microsoft is the largest shareholder of OpenAI, with a shareholding ratio of about 50%. Let's assume that if we want to do AI research, if Microsoft does it within the company, the additional losses may be as high as 10 billion US dollars each year.

In fact, OpenAI is not as NB as everyone thinks. Since its establishment, it has raised less than 1 billion US dollars until 2023. Before the arrival of gpt3 around the Spring Festival in 2023, not much money was invested. But the AI ​​boom in 2023 has come, raising $10 billion, and then $6.6 billion last year, and now plans to raise $40 billion. So, if you really look at the data, OpenAI has entered a turning point in the past two years, and then the narrative given to Wall Street is: the core barriers of AI are capital and hardware - but this sounds a bit awkward.

For Microsoft, if it invests $10 billion in OpenAI, it can get a lot of money from other investors. Then a considerable proportion of OpenAI's money is used to lease Microsoft's cloud computing resources - this forms Microsoft's revenue and profit in accounting. Moreover, OpenAI's model of burning computing power is good for Microsoft in the short term, but in the long run, it is not good for Microsoft - the cost of providing services has never come down. This is why Microsoft and OpenAI have a good relationship in the technology development stage, but when it comes to the service landing stage, the two have some conflicts.

OpenAI must tell the story of capital expenditure, otherwise your financing and valuation will not go up, so a narrative of scaling law emerged. This narrative actually showed signs of collapse before, but no one paid attention to it. I once wrote an article on the official account. After the threshold of large models was rapidly raised, many small teams withdrew from the large model training link, because the capital expenditure required was simply not something that a startup team could support. The big models in the United States are basically supported by industry giants. This leads to a problem. The potential demand for NVIDIA cards in North America may be capped at a scale of a few million cards, which is about the same. On average, each giant purchases 500,000 to 1 million cards each year (these cards are already a lot). The singleness of the ecosystem and customers has caused NV to actually have problems.

Now is the second impact - if algorithm improvement can greatly improve efficiency and reduce costs, and the key is that China's AI chips have been proven today (in the physical sense today) and can run deepseek very well, what is NVIDIA's core competitiveness at this time? At present, the training computing power is basically in excess (even based on the deepseek method, it may be seriously in excess), and the inference chip is destined to be a time for a hundred flowers to bloom...

As for the wise man here, NV, he will never stand up and say that the Scaling Law does not hold, this is the foundation of their survival.

So you will find that this capital game is very interesting: NV has made real money; OpenAI has obtained valuation and a steady stream of financing narratives; Microsoft has not only increased its revenue and accounting profits, but also obtained 50% of the equity of a company valued at 100 billion US dollars and obtained AI technology. So the question is, who is the one who is hurt here? When you don't know who is the idiot at the poker table, you'd better doubt yourself...

In fact, you can easily find that after DeepSeek appeared, no matter how the media argued, what the US government wanted to investigate, threats to national security, etc.... During our Spring Festival holiday, large American companies spent a few days and almost collectively launched the DeepSeek service - this is just a model, and it is an open source model. There is no so-called national security issue at all, because you can use it... So I said this is an open conspiracy, and it is an unsolvable open conspiracy. Don't these large technology companies that deploy DeepSeek services know how to reduce costs?

5. Summary

Let's make a summary of the impact of the DeepSeek incident:

a. Bad news for Nvidia - For those who are currently holding Nvidia shares, don't argue here. The really cool thing is that you made money from going long on Nvidia in the past, and you will also make money from going short on Nvidia in the future. They're all in US dollars anyway, so why should you distinguish between the pros and cons?

b. DeepSeek provides a set of methods, not a specific product. Looking at this path in the future, there may be a lot of room for improvement (a technical detail that I am not very clear about. In fact, due to the limitations of communication and memory, the idle computing power of NV cards is actually very obvious. The utilization rate of computing power may be less than 20%, and most of the time period is basically used for data storage and transmission (AI has three parts, data storage, computing, and transmission). This time DS is targeting this. When H100 was castrated to 800, the main idleness was actually the memory access speed. As a result, it was an accidental achievement... That is to say, if this method is promoted, the current computing power utilization rate will be greatly improved, which will accelerate the disintegration of the demand for cards - after all, for NV, wasting computing power can also make money).

c. It proves that low-end hardware can also implement high-performance models. In the future, there will be a large number of competing products in the market, such as Huawei Ascend 910c, Cambrian, Suiyuan, etc. The landing of these products will be accelerated, so the good news... I won’t say more, I will start a live broadcast to talk about this matter in two days

d. There may be two responses from the United States: 1. Increase the intensity of technological blockade, but it is basically useless - technically speaking, the success of ds this time is due to US restrictions, and the details of the algorithm solve the problem of data storage and transmission; 2. The beginning of easing between China and the United States, the United States must really sit down and have a good talk from the perspective of strength.

e. The current investment opportunities in the supply chain will undergo very big changes... The logic of many things will change. I won’t say more here