The hidden winners of the semiconductor boom through artificial intelligence

By Tobias Rommel, Portfolio Manager DWS Invest Artificial Intelligence, DWS. 

ChatGPT already had around 100 million users after two months and almost without exception, every company we currently visit is talking about artificial intelligence (AI), its own large-language models, and their implementation in the operational business. You don't have to have a crystal ball to see that this will lead to a massive increase in demand for computing power and thus for AI semiconductors.

These chips are already effectively sold out, and in China, the black market is already charging double the price. This year, global sales of AI semiconductors are expected to reach around 43 billion dollars; within the next four years, this sum should triple. At the same time, the market is dominated by a US top dog, which should account for around 70 percent of sales in 2023, at around 30 billion dollars, and which does not sell these chips for expensive money, but also achieves a gross margin of 70 to 80 percent with them. But why is this company's position in so-called "accelerated computing" so strong? Because its history as a designer of graphics semiconductors plays into its hands. Because these chips can - in simple terms - calculate all the pixels of a monitor in parallel, it is precisely this architecture that is particularly helpful for AI applications, since these neural networks also require parallel processes.

However, in view of the high prices and tidy gross margins, the large cloud providers are understandably currently giving a lot of thought to how they can replace the semiconductors of the top dog with their own application-specific chips. Although the companies are at different stages of development, there is probably no question that in a few years these "application-specific integrated circuits" will account for a considerable share of the market. This is supported firstly by the fact that the cloud providers are also heavyweights in the technology sector. And the larger the market becomes, the easier it should be for them to invest the necessary millions in research and development. This is where the first hidden winners come into play. For companies that do not have more than 30 years of experience in the development and manufacture of semiconductors need the support of specialists who can take over the physical design of the chips and work closely with the contract manufacturers. The providers of the software used to design chips could also benefit from this development. And not only the cloud providers could be among their customers in the future, but also the providers of electric cars, of which a prominent representative recently decided to develop his own semiconductors.

Another development is remarkable with respect to AI chips: for a long time, semiconductors became smaller and more powerful at the same time about every 18 months. However, not only has this dynamic, which is laid down in "Moore's Law", already slowed down considerably in recent years, but the AI semiconductors of the US top dog, in particular, are even getting bigger again and are growing into the third dimension in the process. The company's latest product, for example, has 80 billion transistors, which is about five times the amount of these components mounted on an iPhone chip. On top of that, manufacturing is becoming more complex as a result, with the number of manufacturing steps in the production of AI semiconductors now exceeding 1,000. And these two trends are creating something of a boom for equipment manufacturers, who are equipping chipmakers with the necessary new machines.

Does this also mean the end of the proverbial "pig cycle" in the semiconductor sector? Unfortunately not, as evidenced by the drop in prices for unintelligent memory chips and also for semiconductors for the automotive industry, the warehouses are now full again after the lean period of the past few years. However, the demand for smart chips, for example for AI applications or for the "Internet of Things", should be well supported in the longer term. Now one might argue that supply will also increase, because in order to reduce dependence on Taiwan, for example, the governments of the Western world are engaged in a veritable subsidy race to attract semiconductor factories, which should ultimately also fuel production. But wait, those companies from Taiwan, South Korea, or the US that are now starting to build new capacities in the US and Europe are also the operators of the existing manufacturing plants in Taiwan. And they will probably manage the output of their highest-quality chips in such a way that margins should not come under much pressure.