How to make millions on the collapse of Nvidia?

My name is Alexander Stolypin and I am a professional investor in technology companies. Let's start with the fact that right now the demand for artificial intelligence chips has become colossal. Almost every company is expanding or implementing AI technologies that require computing power, and huge ones at that. Giants like Google, Amazon, Netflix, Tesla, Meta* are seriously investing billions of dollars in chips alone to keep up in the ongoing innovation race.

The appearance and boom of ChatGPT became a catalyst. The world saw the real potential of neural networks, called them Artificial Intelligence (which is not quite correct yet), and began to actively implement them: Tesla and Waymo are training their self-driving cars, Google and Meta* are fighting for dominance in fundamental language models, and so on. Everyone is doing what they did before, but now they are giving tasks to machines for (almost) complete consideration.

We will not delve deeply into the technical features of chip operation: they are simply needed to train neural networks or maintain their operation. Some companies offer solutions for the first task, some for the second, some make universal options, but the essence is the same – they are produced in huge quantities.

And since there is such a large amount of money there, let's clarify the situation as a whole.

What is happening: two families in Taiwan are ruling the world and geopolitics

I could loudly say that the main chip manufacturers have a cartel agreement, but the situation is a bit more complicated than it seems.

So, the majority of all the world's chips are produced by Nvidia and AMD. Meta is trying to cope with its own efforts and developments more than anyone else, while Google or Amazon produce their own chips for specific solutions. However, more than 90% of chips from the entire global turnover are purchased from Nvidia and AMD. At the same time, in their pair, more than 80% of the total volume is produced by Nvidia.

But there is one catch: both company leaders are from Taiwan, they are relatives and just good friends. We are talking about Jensen Huang and Lisa Su: their respective grandmother and grandfather turned out to be siblings.

On the one hand, there is nothing unusual: Taiwan is a small island where the probability of being someone's relative is far from zero.

On the other hand, in the current environment of investors and business in general, kinship is avoided, especially when it comes to the same field. The likelihood of monopoly, fraud, and family interests can outweigh and contradict the interests of investors, as well as attract the attention of the FAS and all related structures.

Given that almost all the world's chips are made by two families, it becomes a bit unsettling. Of course, both Nvidia and AMD are based in the USA and are subject to state laws, but there is a nuance. Both companies are only engaged in chip design and software creation/maintenance.

The hardware is directly made by the Taiwanese company TSMC, which is recognized as the leader in semiconductor manufacturing and development: their solutions are the most efficient and productive. They are preparing to release working 2-nm processors (for which a queue of well-known companies has already formed), and it is from them that both Nvidia and AMD buy – in fact, both manufacturers depend on TSMC products, as they build their technologies based on their chips. Similarly, Apple, thanks to TSMC's developments, is rapidly updating its own chip lineup, having released three new generations (M2-M4) in less than two years.

The CEO of TSMC is also connected with Nvidia and AMD, and the production facilities and headquarters are also located in Taiwan. It turns out that the AI hardware industry is in the hands of literally a few families of a small state, which can be called a monopoly de facto.

This is why China and the USA (to put it mildly) are showing interest in Taiwan. AI chips are becoming something like oil for the entire economy. And whoever owns the fuel essentially controls the world. Since 2022, the territorial conflict between Taiwan and China has seriously intensified: China began active military exercises in the Taiwan Strait, and the USA increased its military presence on the island.

In August, there was a famous visit by the Speaker of the US House of Representatives – Nancy Pelosi. This caused a sharp reaction from Beijing, which took such actions as support for Taiwan's independence. Immediately after the visit, China conducted large-scale military exercises around the island, which included missile launches and the blockade of sea and air routes.

In response, the US began economic pressure on China: first of all, they withdrew more than a thousand of their microelectronics specialists from China, literally forcing them to return home. Then Biden signed a document tightening the export of semiconductors to China, and a number of Chinese specialists were restricted access to semiconductor technologies.

To date, China is actively demonstrating its military power, literally hinting at a possible invasion of Taiwan, and the US is supplying air defense systems and providing them with public support.

At the moment, China's technology with all its capabilities allows the production of 28-nm processors – lagging behind Taiwan by at least 15 years, and to obtain (or better yet directly own) the advantages of Taiwan. And therefore, in fact, to influence all the advanced technologies of the world.

Hence, there will be an influence on the world economy as a whole – for a country like China, this will become a factor of dominance even over the US.

But Taiwan is not at war, processors are being produced, and chips are being assembled more and more, outpacing Moore's law by a whole six months. So let's get back to Nvidia, AMD, and TSMC.

The current state of affairs with companies cannot last forever: the law of the invisible hand never allows a giant, a syndicate, or anything similar to exist alone. Sooner or later, new players come with a completely different approach to the product. And we will talk about this a little below.

Nvidia staggered: new players on the horizon

Taiwanese chip alliance

Despite Nvidia's dominant role in the chip sector, many do not realize that a significant part of the success of leading manufacturers is associated with Taiwan. It is not only TSMC — the largest contract chip manufacturer in the world, but also their strategic cooperation with the Dutch company ASML.

ASML is a key player without which neither Nvidia nor any other manufacturer will be able to make their chips. The company produces lithographic machines used to apply silicon to semiconductors, which is the basis of chips.

The problem is that all major chip manufacturers are actually dependent on ASML technologies. The company itself is located in the Netherlands, and this is the first counterbalance to the global monopoly.

In recent years, ASML has been actively discussed in the context of geopolitics and technological rivalry. For example, the United States restricts the sale of ASML machines to China, fearing increased competition. These restrictions hold back the expansion of Chinese manufacturers, but also increase the dependence of the United States and the EU on a single supplier.

ASML equipment costs hundreds of millions of dollars, and companies produce it piece by piece. The assembly and calibration process takes months, making it a rare and expensive resource, limiting the number of companies that can afford it.

Cerebras: a new challenge for Nvidia

Yes, Nvidia has a serious competitor on the horizon – Cerebras. They also make chips using TSMC in Taiwan.

But there is one feature: their chips are built on a unique architecture that allows processing large volumes of data simultaneously, making them significantly more efficient for deep learning AI.

They have developed one of the largest chips in the world with 4 trillion transistors. The most powerful Nvidia Blackwell B200 model has over 200 billion transistors, which is tens of times less. 900,000 cores, bandwidth of 21 petabytes – the gap from Nvidia's current performance is huge.

And it can be said that superior characteristics are often not an indicator of efficiency, but even now Cerebras' performance shows better results than competitors – only due to the size of the chips.

Unlike Nvidia, whose architecture requires the installation of tens of thousands of chips in large data centers (for example, Elon Musk's data center uses more than 100,000 Nvidia chips), Cerebras can manage with much fewer chips per data center.

And thanks to this, efficiency losses are significantly reduced – the more chips you install, the less efficient each chip becomes individually. It is necessary to hire teams of developers to optimize and minimize lost power programmatically.

But Cerebras went further and managed to achieve linear power progression. In other words, 2 Cerebras chips will be exactly 2 times more powerful, making them fully competitive.

The company was founded in 2016, and already in 2022 released the world's most efficient supercomputer for training AI algorithms – Andromeda, capable of delivering 1 exaflops on AI tasks.

1 exaflops is the ability of a computer to perform a billion billion floating-point operations per second. For example, it would take a person performing 1 operation per second 32 billion years to do what a 1 exaflops device does in one second.

Cerebras solutions are unique and more efficient than Nvidia's. Let's look at the benchmarks:

The image shows a comparison of various computing platforms using the neural network Llama 3.1 Instruct 8B from Meta, by output speed, measured in tokens per second.

The more tokens a platform can process per second, the faster it works.

Cerebras (first on the left) shows the highest performance among all the platforms presented, processing 1,846 tokens per second. This significantly exceeds the performance of the nearest competitor Groq, which processes 750 tokens per second.

All other platforms in the image run on Nvidia chips (except for Groq and Cerebras). Their performance is significantly lower. For example, Fireworks processes 251 tokens per second, while giants like Amazon and Azure (Microsoft) show speeds of around 9 and 47 tokens per second respectively.

The cost of using Cerebras is also one of its key advantages — according to the information in the image, its use costs three times less than, for example, Microsoft Azure, while showing performance tens of times greater.

And if a year ago the phrase Cerebras - a competitor to Nvidia, really sounded like a fantasy, today the situation has changed significantly.

And an interesting fact: Sam Altman, head of OpenAI, owns a stake in Cerebras, which means he is betting on the company's success – their shares are traded on the pre-IPO market.

The price of Cerebras shares has already increased 5 times on the over-the-counter market over the past year. This greatly pleases investors (including myself personally). When directly listed on the stock exchange, their shares are capable of growing many times over when the first real competitor to Nvidia becomes available for investment to billions of investors worldwide.

But syndicates and pools (which we also have in Russia and the CIS) allow you to enter a deal with several thousand dollars – so there are opportunities for the average investor as well.

At the moment, the Cerebras IPO has been put on hold by the US regulator, again due to geopolitics. The main investor of the firm – the Saudi Arabian fund – can sell chips directly to China, bypassing all current sanctions, which contradicts US national security.

This is a sure sign that the company is creating truly breakthrough technologies.

Today we learned a little about who is breathing down the neck of Nvidia and AMD. Soon we will talk about even more interesting topics.

Also, I run a blog about tech companies that bring innovation to the world, successfully implement themselves on the stock exchange, at pre-IPO and IPO stages, and tell where they can be bought.

With you was Alexander Stolypin.
See you in the future!