Source: Yaya Hong Kong Stock Circle
Author: Kyle

The last two weeks have undoubtedly been the most exciting since ChatGPT initiated the AI wave. First, Trump approved the $500 billion AI infrastructure plan led by OpenAI, followed by the emergence of Deepseek which halted the monolithic rise of the AI theme that lasted over two years. This has left global AI investors anxious, as they grappled with whether the computation power was necessary amid the heated discussions of the past few days.

Until now, $NVIDIA (NVDA.US)$ Returning to the 130 yuan mark, the strong AI stocks are close to erasing the declines caused by the crisis triggered by Deepseek. Do you believe in this AI computing crisis?

1. The grand narrative of AI continues.

Looking back at the AI crash triggered by Deepseek, it can be said that even those who have faith in AI were inevitably shaken by market emotions, not to mention those who questioned whether such a large investment in AI was needed; moreover, large funds could only choose to reduce positions and avoid risks amidst uncertainty.

Currently, the market concludes that the emergence of Deepseek is significant, as it not only prompts a reevaluation of the importance of investment in AI hardware but also opens up a lower-cost technological pathway. At this stage, arguing about whether winning is meaningless; what can be confirmed is that computing power remains important. The AI chain in the USA continues to play music and dance, and as a result, our Chinese concept stocks are receiving attention from foreign capital, starting to rebound. Furthermore, the anticipated domestic increase in AI investment will further narrow the valuation gap between Chinese concept stocks and US stocks.

The rebound of Hong Kong stocks and Chinese concept stocks in the past two weeks is closely tied to Deepseek, with the best example being the USA stocks. $Alibaba (BABA.US)$ Suddenly, a flower of hope for AI blooms again from the pessimistic narrative, somewhat similar to the logic of buying CSP stocks in the USA last year, where the strong model Qwen2.5max has excellent performance scores, combined with Alibaba's computing power reserves and a complete cloud platform, which is also being replicated onto. $XIAOMI-W (01810.HK)$ 、 $Kingsoft Cloud (KC.US)$ It has also replicated onto.

In the past two weeks, some hedge funds and long-only investors have begun to cover some positions in China, and we can even see foreign banks making grand narratives similar to the trend of September 24. For example, statements like "China eats the world"; however, this rebound brought by a non-policy shift is expected to be more sustainable compared to the short-term increase driven directly by a 90-degree slope policy. In simpler terms, the market has finally realized that China has AI and it should not be underestimated.

Over the past year, AI has been lacking proof of whether applications can be profitable, questioning whether heavy investments are necessary for AI applications, or if only large companies with substantial investments can engage in AI. The emergence of DeepSeek this time actually reinforces confidence that 2025 will be a breakout year for AI applications, also indicating that there are lower-cost pathways available for smaller companies and developers.

However, the market is an amplifier of emotions, and regardless of whether it is the US stock market, A-shares, Japanese stocks, or Korean stocks, related symbols experienced a massive correction during the Spring Festival due to the panic caused by DeepSeek.

Specifically, in the market, Nvidia fell by 17%, $Broadcom (AVGO.US)$ fell by 17%, $Taiwan Semiconductor (TSM.US)$ fell by 13%, the sharp decline among the big players has caused even more exaggerated drops in the smaller stocks within the AI subsector, for example, those that have seen a rise exceeding three times since last year. $Credo Technology (CRDO.US)$ On that day, it dropped 30%, and the insider stock guru Pelosi just bought AI power generation equipment VST, which also dropped 28% that day.

To be honest, the volatility on that day was indeed quite exaggerated, and this is considered one of NVIDIA's largest single-day declines in the past two years. However, if viewed over a longer period, these stocks have increased two to three times in the past year or two, and there hasn't been a substantial adjustment along the way. The emergence of deepseek has shaken the narrative of the computing power arms race, leading to a full release of AI themes.

And $Apple (AAPL.US)$ and $Meta Platforms (META.US)$ On that day, it not only did not fall but actually rose by 2%, and the reason is quite simple: the certainty of AI applications is becoming stronger, and even edge AI may have reached a turning point.

For example, Meta's glasses performed quite well last year. According to Wellsenn data, the sales of Rayban Meta smart glasses have exceeded 2 million units in 2024, reaching 2.25 million units. Meta's success has also attracted more Technology stocks to bet on AI glasses, such as Samsung, XIAOMI-W, $Baidu (BIDU.US)$ , and even ByteDance has made moves in this area.

Meanwhile, Apple’s approach to embedding AI is actually more extensive. Although Apple's Intelligence is currently not very impactful, as large models continue to develop, the possibility of AI voice assisting mobile phone usage is indeed viable. For instance, using ChatGPT, deepseek, and Doubao, the experience of voice functions in large models is quite good. Integrating this into Apple’s products isn’t particularly difficult; the problem is that SIRI currently lacks capability, resulting in a very poor personal user experience. In fact, SIRI often struggles to understand what you are saying and doesn’t have enough permissions to assist you with using the phone.

One point worth mentioning is that everyone should consider in daily life what AI can do for them. When do they need AI assistance? Would it be more convenient if AI is there at that moment? Approaching daily life with this question, one can actually discover that there are many tasks AI can help with, and the difficulty is not high. In fact, AI voice assistance is one of the easiest applications. Therefore, in edge applications, AI voice will certainly bring a significant market.

So, how can these things be achieved?

Ultimately, it still revolves around more, stronger, and more efficient computing power. When the application hasn’t exploded, it is meaningless to judge hardware investment; this is too emotional.

The emergence of deepseek has only resulted in a structural change in the training narrative, but the conclusion is that the scale has strengthened the inference narrative. In the words of foreign-funded entities, this relates to the Jevons Paradox, which states that increasing resource use efficiency may actually increase total consumption.

For example, in the past two weeks many companies have started to integrate Deepseek's API or use Deepseek. When the Deep Thinking and online search functions are used together, it shows that the server is busy, which indicates a shortage of computing power. When comparing the 910B and H200 running the Deepseek R1 model in the market, there are two orders of magnitude difference, not to mention that the GB200 running R1 is ten times faster than the H200.

There is also a noteworthy example. In the past two years, there has been significant debate in the AI industry regarding ASIC inference and GPU training. However, the reality is that NVIDIA and Broadcom have both reached new heights. The shift of CSP customers to ASIC does not imply that it is Bearish for NVIDIA, as the increasing overall demand is far more significant than competition.

In the USA stock market, when new AI technology routes emerge, both tend to rise together. However, in the A-share market, it may be characterized by the concept of 'light out, copper in,' where one moment it is light going in and copper going out, and vice versa. The speculation around AI in A-shares is much more complex than in the USA stock market.

II. Technology stocks and individual investors support NVIDIA.

Returning to the discussion about computing power in the market, the only ones who can rescue NVIDIA from crisis are its CSP clients. Fortunately, the Deepseek crisis is not far from the earnings report period of the American stock market's seven sisters, otherwise, NVIDIA may not be able to rebound so quickly in the short term. After all, even if foreign banks insist that computing power demand will not be affected, it is still less direct than the several hundred billion dollars of Capex.

Specifically looking at the Capex situation in the fourth quarter;

$Meta Platforms (META.US)$ : 14.8 billion dollars, year-on-year +88%, quarter-on-quarter +61% (market expectation 12.8 billion dollars);

$Alphabet-A (GOOGL.US)$ 14.3 billion USD, +29% year-on-year, +9% quarter-on-quarter (market expectation 13.2 billion USD);

$Microsoft (MSFT.US)$ 22.6 billion USD, +96% year-on-year, +13% quarter-on-quarter (market expectation 21 billion USD);

$Amazon (AMZN.US)$ 27.6 billion USD, +95% year-on-year, +22% quarter-on-quarter (market expectation 22.3 billion USD).

2025 Guidance:

Meta: 60-65 billion USD, +53%-66% year-on-year, market expectation over 50 billion USD;

Alphabet: $75 billion, year-on-year +43%, market expectation over $60 billion;

Microsoft: $85-90 billion, year-on-year +44%, market expectation over $80 billion.

Amazon: approximately $100-105 billion, year-on-year +26%, exceeding market expectation over $90 billion.

Meta's Zuckerberg stated that significant investments will enable the company to adjust technology in line with AI development. Meanwhile, Amazon's CEO described AI as an once-in-a-lifetime business opportunity, stating that such large-scale expenditures are due to observed demand. This year, North American CSP's capex will exceed $300 billion, with a year-on-year growth rate of 30%. With such scale, having a 30% increase is already quite alarming.

With backing from four major clients, NVIDIA can be said to have no immediate risks. Not only has the demand for computing power not decreased, but the decrease in inference costs has actually strengthened the confidence of various CEOs in increasing investments. The simplest and most straightforward reason for such significant investment is that these tech stocks are incredibly profitable. Even with annual AI expenditures in the hundreds of billions or thousands of billions, free cash flow is still increasing.

In addition to support from major clients, on the day NVIDIA fell by 17%, individual investors who believe in AI once again acted contrary to Wall Street.

According to Vanda Research, following the deepseek-induced decline in US stocks, individual investors saw an inflow of $4.25 billion into the US stock market. On January 27, during the most severe sell-off in tech stocks and Semiconductors, $1.85 billion from individual investors was used to capitalize on AI. The sharp increase in net purchases raised the average daily inflow of retail funds over the past five days to approximately $1.3 billion—similar to levels seen just before and after the US elections.

So, after the crisis of computing power demand has been resolved, how should NVIDIA be viewed next? This is a question worth thinking about.

1. After China's AI breakthroughs, it will definitely be unacceptable for someone as sensitive as Trump, especially since I just led the signing of a new Infrastructure deal worth $500 billion last week, and now I am being smacked in the face by this news. The impact on NVIDIA is likely to lead to stricter restrictions on chip sales to China; there have already been recent rumors about this, and even the stripped-down version of the H200 might not be allowed for sale. Additionally, sales through channels in Singapore may also face tighter regulations.

This means that NVIDIA cannot choose a path; the initial biggest beneficiary of the $500 billion Infrastructure project is NVIDIA's GPUs. If Trump makes demands, compromises will have to be made. As of Q3 2024, NVIDIA's total revenue is $35.08 billion, with approximately $5.4 billion coming from the Chinese market, accounting for about 15.4%. So if we consider the extreme case where the stripped version cannot be sold, it will indeed have a significant impact on NVIDIA, which will be reflected in its valuation.

2. Looking at the latest target prices from foreign investment banks on Wall Street, while target prices often fluctuate widely, they can offer insights into the market's expectations for NVIDIA's potential upside.

Citigroup: Downgraded the target price from $175 to $163, mainly considering the risks brought by restrictions on China's Datacenter.

Morgan Stanley: Maintains a Buy rating and raised the target price from $152 to $172.

Bank of America: Maintains a "Buy" rating with a target price of $190, emphasizing NVIDIA's leading position and growth opportunities in Emerging Markets.

Deutsche Bank: Maintains a "Buy" rating with a target price of $220, believing that NVIDIA's leadership in AI and Datacenter will drive its long-term growth.

Goldman Sachs: Maintains a "Buy" rating with a target price of $200, expecting NVIDIA to benefit from the growth trends in AI and Cloud Computing in the coming years.

Conclusion

Currently, NVIDIA is at $130, with a market cap of $3.15 trillion. The odds of it doubling like in the past two years might not be as high, but investing in AI at least cannot do without NVDA. Therefore, NVDA will at least be at the core of the infrastructure investment in AI in the USA.$BETA (0263.MY)$Stock.

In that case, it may be necessary to look for opportunities in more segmented areas. The $500 billion Stargate program has a high demand for other server components, such as optical fibers, the newly added AEC core wire in the GB200 server, server power supplies, server electrical utilities, etc., so there are greater opportunities in smaller stocks in more segmented fields.

Overall, the grand narrative of AI investment is not over yet; AGI is still far from us. This year will mark a new phase in AI investment, gradually shifting from GPUs to segmented fields, then to applications and edge computing.

Editor/Rocky