When Microsoft first invested in OpenAI (in 2019), it did not anticipate the ferocity of the current artificial intelligence (AI) revolution. Thus, Bill Gates remarked that we tend to overestimate the changes in the next two years but underestimate the transformations over the next decade.

In 2023, AI welcomed its "iPhone moment." Nvidia is undoubtedly the "shovel seller" in the AI revolution gold rush, and in recent years, it has been hailed as the "god of global stock markets." This "first AI stock" has often faced skepticism about its stock price, but each earnings report has proven the bears wrong, with the stock price rising higher and its valuation becoming cheaper (profit realization exceeding expectations).

After a sharp drop in April due to expectations of interest rate cuts, Nvidia's stock price ($898) has recently approached its previous high ($910), with the company set to release its Q1 2024 earnings report on May 22nd.

Recently, Wall Street has once again raised its stock price forecasts. $1,000 is no longer a questioned astronomical figure. Goldman Sachs has given a new 12-month target price of $1,100 (previously $875 in March), as institutions have increased their EPS (Earnings Per Share) forecasts for the fiscal years 2025 to 2027 by an average of 8%, indicating a sustained strong demand for AI servers and supply improvements. The boundless prospects of data center revenue are undoubtedly Nvidia's ballast.

Advertisement

Despite Nvidia's high moat, market opportunities are not limited to it. AI is still in its early stages of transformation, so more investment managers prefer to "buy a little of everything."

Nvidia's standard AI GPU may capture 80% of the market share, and customized AI chips are also benefiting some companies, such as Tesla, Amazon, Microsoft, Google, Alchip, Broadcom, and AMD.

The Asian market also presents many opportunities. Y.T. Boon, a semiconductor industry expert and head of the Asia-Pacific Equity Thematic Research Department at Robeco, mentioned to me in Shanghai that AI investment is growing rapidly, and AI chips will grow at an annual rate of 50% over the next five years.

Asian countries such as Taiwan, Japan, South Korea, and Southeast Asia (Thailand, Malaysia, etc.) have gathered a series of companies along the industry chain, which are expected to continue benefiting from the AI wave.

01

Is Nvidia Heading Towards the Thousand-Dollar Mark?In April, the release of U.S. inflation and retail data exceeded expectations, leading to the evaporation of expectations for a rate cut in June. At that time, U.S. tech giants plummeted, with the "Tech Seven" losing a combined market value of $950 billion, setting a historical record. Nvidia suffered the most severe loss in market value, with its stock price plummeting by 13.6% during the week of April 15-19, wiping out nearly $300 billion in market value.

However, the giants quickly recovered after falling, and now they are approaching $900 again, with a nearly 90% increase this year. As early as March of this year, several U.S. stock fund managers mentioned to me that $1,000 sounded exaggerated before, but it is now within sight.

At that time, Nvidia's dazzling financial report slapped all short positions in the face. The financial report for March showed that Nvidia's sales in the fourth quarter of last year reached $22.1 billion, a year-on-year increase of 265%, of which data center revenue related to AI grew by 409% to $18.4 billion, and net profit reached $12.29 billion, a year-on-year increase of 769%.

Although Nvidia's data center revenue in fiscal year 2024 grew more than three times year-on-year, Goldman Sachs expects it to achieve more than double growth again in fiscal year 2025.

"We expect large cloud service providers and consumer internet companies to continue to grow in generative AI infrastructure spending, while enterprise customers in various industry verticals, as well as an increasing number of sovereign states, will also increase the development and application of AI. In the short term, we see a series of new product launches, including H200 (inference performance is twice that of H100), Spectrum-X (Ethernet-based AI network solution), and B100 (next-generation data center GPU platform), as well as supply improvements, all of which have strengthened an already strong demand backdrop."

However, some institutions that have laid out Nvidia are cautiously optimistic. A QDII investment manager mentioned to me that from a medium to long-term perspective, Nvidia's data center GPU business faces two challenges:

On the one hand, the competitive landscape of general-purpose GPUs is deteriorating, with AMD and Intel both entering the market. Although Nvidia is still very likely to be the absolute leader with a market share of 70%-80% in the long run, compared to the current 90% market share, the situation is deteriorating, and it may not be able to maintain such high prices and profit margins.

On the other hand, each CSP (cloud service platform, such as Amazon's AWS, Microsoft's Azure, etc.) has a huge motivation and ability to develop their own dedicated AI acceleration chips. Although it is still difficult to provide general computing services to the outside world in the foreseeable future, as a fact that is happening, it is replacing a part of the existing Nvidia chip share for internal use.

In his view, Nvidia's reasonable trading range is between 30-40 times the price-to-earnings ratio, with a market value cap of around $2 trillion (recently broken through), and a better entry point is around $1.5 trillion.

However, it is normal for AI to inflate bubbles in its early development, and it is not easy to buy at a reasonable price. Recently, there are extremely optimistic institutions that predict Nvidia's market value will reach $4.5 trillion by the end of 2025.Let's take a look at NVIDIA's business segments, which mainly include the following areas:

Gaming Platform, Data Center, Artificial Intelligence, Automotive, and Professional Visualization. The Data Center is undoubtedly the leader.

The reason why many Wall Street institutions have recently reiterated their buy ratings for NVIDIA and raised their target prices is due to the continued strong demand for AI servers and supply improvements, which have led institutions to upgrade their earnings forecasts.

Especially considering that NVIDIA is currently trading at a price-to-earnings ratio of 35 times, which is only equivalent to 36% of the premium covered by Goldman Sachs, while its median premium over the past three years has been 160%. In other words, institutions believe that NVIDIA's stock price seems high, but in reality, it is not expensive.

The reason for raising the target price is also due to the recent positive outlook on capital expenditures related to generative AI from NVIDIA's major customers (mainly members of the "Tech Seven"), as NVIDIA's GPUs are in short supply:

1. Alphabet stated that it has made good progress in generative AI-related services and said that due to investments in technology infrastructure, capital expenditures in the remaining quarters of 2024 may be higher than in the first quarter (about $12 billion);

2. Microsoft emphasized that AI contributed 7 percentage points to Azure's growth in March, higher than 6 percentage points in December and 3 percentage points in September. Currently, the demand for Azure's AI exceeds its existing capacity. In terms of capital expenditures, Microsoft guided that there will be a substantial sequential increase in the June quarter and stated that capital expenditures in the fiscal year 2025 will increase year-over-year, as the company aims to meet the growing demand for cloud and artificial intelligence products;

3. Meta raised its 2024 capital expenditure guidance to $35-40 billion (previously $30-37 billion) and shared expectations that capital expenditures in 2025 will grow year-over-year, mainly due to investments to support its AI research and development work;

4. Amazon expects a significant year-over-year increase in capital expenditures in 2024, mainly due to high infrastructure capital expenditures to support the growth of AWS, including generative AI.

Considering the above positive comments from the AI ecosystem and the launch of various new products, institutions predict that NVIDIA's data center business will achieve sequential growth of 10%, 17%, and 5% in the second, third, and fourth quarters of this year, respectively (previously 10%, 15%, and 0%), reflecting the continued strong demand.Goldman Sachs believes that constructive capital expenditure comments from hyper-scale cloud service providers, early signs of AI commercialization, and the launch of various new products by NVIDIA have bolstered Wall Street's confidence in the short term.

TSMC's management indicated in a recent earnings call that despite plans to more than triple the annual growth of CoWoS capacity, the CoWoS capacity remains tight. Almost all of the CoWoS capacity is occupied by NVIDIA.

The key point is, how many GPUs do these well-known giants have on hand, and how many do they need? Last year, there was a rough estimate that the market demand would reach about 432,000 H100s. If we talk about a selling price of around $35,000 per card, it equates to $15 billion worth of GPUs, not including Chinese companies like ByteDance, Baidu, and Tencent, which are eager for the H800.

Let's take a look at NVIDIA's production capacity. Looking back at February to April 2023, the data center revenue was $4.28 billion. If we follow the earlier estimate, it's already enough for NVIDIA to last until 2024.

The supply bottleneck for H100 lies in the fact that only TSMC is producing H100. TSMC has a total of four production nodes that provide capacity for 5-nanometer chips, namely N5/N5P/N4/N4P, and H100 is produced on a private node of N5 or N5P, an enhanced 5-nanometer node, where NVIDIA needs to share the node's capacity with Apple, Qualcomm, and AMD. The A100, on the other hand, is made on TSMC's N7 production line. TSMC's wafer fabs need to regulate the capacity allocation for each customer 12 months in advance, and H100 takes about half a year from production to shipment.

According to industry insiders, the wafer fab is not the production bottleneck for TSMC; the CoWoS 3D stacking packaging is the gateway to TSMC's production capacity.

02

How Deep is the Moat?

All along, intensifying competition has been the main point for those who question the sustainability of NVIDIA's stock price, but NVIDIA's moat is hard to break.

In terms of its own merits, the reason why more companies are choosing NVIDIA's H100 over the A100 is that H100's 16-bit inference speed is roughly 3 to 5 times faster, and the training speed is about 2.3 times faster, thus the overall efficiency is tripled, but the cost is only 1.5 to 2 times. When combined with the overall system cost, the performance per dollar that H100 can generate is higher, possibly 4 to 5 times that of the A100.For many startups, the speed of model initialization training or improvement is crucial, and the H100 can scale better with a larger number of GPUs, offering faster training times. So why don't these companies consider using AMD's GPUs? There are several reasons, such as the potential for AMD's GPUs to require more time to operate, which could affect the speed of product development and, consequently, the speed of market launch.

It is evident that NVIDIA's CUDA is a very "dangerous" moat. The full name of CUDA is Compute Unified Device Architecture. The CUDA ecosystem allows developers to more efficiently utilize NVIDIA's GPU parallel computing capabilities to accelerate computational tasks.

This ecosystem includes the CUDA Toolkit and applications. The CUDA Toolkit is a software development kit provided by NVIDIA for GPU programming development. It offers a complete development environment, including a compiler, debugger, performance analysis tools, GPU-accelerated libraries, and runtime APIs, among other tools, enabling developers to write CUDA applications using programming languages such as C++, Python, and Fortran, and to leverage the parallel computing capabilities of GPUs.

The reason AMD has never been able to keep up with NVIDIA is also due to the fact that TSMC's CoWos production capacity, which is now a chip foundry, is almost entirely absorbed by NVIDIA. Even if the MI250 could be a viable alternative, it is not very available at the moment.

Some also ask why AMD does not develop an ecosystem like CUDA. There is speculation that this is because AMD has an existing market with x86 processors, and if they were to seriously engage in neural networks, they would likely prefer to do so on the CPU, which is their core business. However, NVIDIA is different; it only has GPUs and must fully utilize their capabilities.

Of course, whether AMD can be used on a large scale is a business decision, which also depends on factors such as the stability and cost-effectiveness of AMD's performance. It is estimated that the price of AMD's MI300 is approximately $13,000. Compared to NVIDIA's H100, which has similar computing power but half the memory, the MI300 has a price advantage of about 43%.

This 43% advantage is offset in software upgrades. It is worth noting that the comparison is between the newly launched MI300 and the H100, which was released in 2021, and this comparison may not be appropriate in itself. Therefore, all comparisons must consider the timeline and whether a long-term advantage can be formed.

In fact, just in March, after a two-year hiatus, Huang Renxun made a significant announcement at NVIDIA's AI conference GTC, unveiling the new generation of Blackwell architecture GPUs.

He said, "Hopper is great, but we need a more powerful GPU." Reports indicate that the highest specification of the Blackwell chip has a floating-point operation speed (FLOPS) that is about five times faster, with further optimized energy consumption, showing strong competitiveness compared to AMD's MI300X GPU, and consolidating NVIDIA's technological advantages in performance and energy efficiency.03

But Opportunities Go Beyond NVIDIA

Although NVIDIA has an unbreakable moat, AI concept stocks have seen a widespread increase over the past year, with growth across markets from the United States to Asia. This actually implies that there are many opportunities throughout the industry chain, both upstream and downstream. Moreover, despite NVIDIA's strengths, investment institutions cannot possibly invest entirely in one company.

Wen Yandao mentioned to this author that NVIDIA can be compared to "Uniqlo"; everyone will go to buy standard products from it, but many cloud service providers, such as Amazon, Google, Microsoft, and others, are actually developing their own large language models, and they all require customization. Although they are already major clients of NVIDIA, they also want customized AI chips, so they will look for other companies, such as AMD in the United States and Alchip in Taiwan, China. Alchip is helping Amazon and Tesla to customize AI chips, while Broadcom in the United States is customizing AI chips for Google. Customized products are like tailored suits, fitting better, and their operating costs are far lower than those of general-purpose NVIDIA, with power consumption being only 1/5.

Super Micro Computer (SMCI), known as NVIDIA's "beloved son," is also a dark horse this year. From last year to March 2024, its stock price has seen a maximum increase of up to 13 times, surpassing NVIDIA; its market value has grown from $5 billion to over $50 billion.

To illustrate, NVIDIA's GPUs are not simply connected by wires and ready to use; they are grouped in sets, plugged into Super Micro Computer's racks, and combined with other components to create servers that are sold to clients such as OpenAI, Google, Microsoft, and Meta for training large language models or for cloud computing.

Super Micro Computer also has a unique liquid cooling technology to reduce the power consumption of GPUs, allowing businesses to save costs significantly. Some analyses point out that, given the multitude and heat generation of GPUs required for generative AI, the next generation of data centers must use liquid cooling instead of wind cooling for temperature regulation.

Furthermore, the role of the Asian industry chain should not be underestimated, with many companies driven by NVIDIA. TSMC, which naturally needs no introduction, has taken on the production of all NVIDIA chips; Japan's semiconductor industry chain is also rising in the context of geopolitical dynamics, with more American companies establishing factories in Japan. The expanding demand for AI chips also benefits Japanese manufacturers.

For example, Tokyo Electron, a leader in semiconductor equipment cleaning devices, has over 60% of its sales in China, doubling the proportion from earlier years, due to its accelerated establishment of new factories in the country. Tokyo Electron's stock price has nearly doubled in a year; another Japanese company, Disco, known as the "Japanese sushi knife," holds nearly 100% of the market share in cutting and grinding machines in the industry chain and has also benefited from the AI chip wave, with its stock price increasing by over 160% in a year.

In the downstream AI application end, there may also be more products in the future. In addition to the various AI applications available now, Wen Yandao stated that companies such as Samsung, Huawei, and Xiaomi will soon introduce AI phones, and the penetration rate of AI computers will gradually increase. However, the terminals for AI may not be limited to just phones and computers; AI Pin is a great example of this.To put it most bluntly, AI Pin is a brand-new concept of an "AI assistant." The founding team, Humane, aims to break the existing consumer impression of all electronic products. The functions demonstrated by AI Pin are quite diverse, including chatting, writing, playing music, etc., all of which can be accomplished on this small device that were originally only possible through online ChatGPT. The most distinctive feature may lie in its detection and projection capabilities.

Thanks to the 13-megapixel camera equipped on AI Pin, it can be paired with different gestures to instruct AI Pin to perform various functions, even detecting the nutritional content of food, which are features that traditional wearable devices currently cannot achieve. Compared to the bulky and easily "socially awkward" VisionPro helmet, this Pin is much more compact and lightweight.

In fact, AI has always existed, but quantitative changes lead to qualitative changes. Today, generative AI is a revolutionary phenomenon. By 2030, 70% of companies worldwide will adopt AI. Automation and AI-enhanced workforces will increase productivity, and more personalized and AI-powered high-quality products will also drive an increase in consumer demand. The future is already here, but the changes in the next decade will be absolutely unimaginable.