IFI CLAIMS Technology Spotlight: Semiconductors

When the Chips Are Up

Among the many ways the global pandemic wrenched the supply chain, the unavailability of chips—needed for phones, computers, cars, appliances and many other systems—was a shock we won’t soon forget. To better protect the U.S. supply chain, President Biden signed the bipartisan CHIPS Act in 2022, granting nearly $53 billion to incentivize big chipmakers to build more manufacturing capacity for advanced semiconductors within the U.S., the birthplace, after all, of the transistor. Currently, the U.S. produces some 10 percent of the global supply. But the most cutting-edge chips, the class that has fueled the stock rocket called Nvidia, are produced in Asia, a concern for U.S. businesses and national security when supply chain snafus and geopolitical tensions arise. Chipmakers around the globe are battling for dominance in this vital industry that is projected to surpass $1 trillion, according to a recent analysis by the Wall Street Journal. And foreign governments are opening their coffers to shore up their homegrown manufacturers in order to build and maintain a national edge in the industry.

In the U.S., Intel has been granted $8.5 billion of the CHIPS Act funds to support development of its new facilities in Ohio. The company has also announced its own $100 billion investment in growing its U.S. manufacturing capacity. Taiwan Semiconductor Manufacturing Company (widely known as TSMC), Samsung, and Micron have also received sizable government grants to build out their U.S.-based fabrication. These manufacturing projects will take years to complete.

Allocating funds to the production of chips is a long-term investment that companies hope will pay big future dividends. The short-term stock moves, however, can be much more volatile, and market watchers have witnessed such among chip companies over the summer as Wall Street vacillated about the unknown timeline of future profits from these investments materializing into earnings weighed against the current and substantial capital expenditures to develop AI.

In view of all the attention from governments, businesses, and markets on the surging semiconductor business, IFI CLAIMS, the patent industry’s most trusted data provider, decided to take a look at the industry from the perspective of its patents because investment in the R&D needed to design and make advanced chips can have a time horizon comparable to the building out of brick and mortar facilities.

Patents hold important clues that can help investors decide whether or not a company deserves a place in the investment portfolio. Here is how the chip companies stack up.

The Number of TSMC’s Patents Is Fab

Manufacturing powerful computer chips for such companies as Nvidia and AMD is a highly technical and intricate process. Consider that transistors on a silicon wafer are smaller than a strand of human hair. Building chips then requires a magnitude of invention that few other industries call for, and the sheer number of the U.S. patents in this space over the past five years is staggering. Of the more than 300,000 published patent documents measured by the Cooperative Patent Classification (CPC) system covering the semiconductor devices space (CPC code H01L), TSMC holds about 10 percent of them.

Code H01L, it should be noted, is not the only one covering semiconductors. But because it holds a number of inventions that contain semiconductors, it is often viewed as a proxy for the industry. It’s also worth pointing out that this code is weighted heavily in the manufacturing and hardware segment of the chip industry. So it’s not surprising that the largest chipmaker in the world, which fabricates some 90% of the most highly developed chips, is on top.

Like chips themselves, the semiconductor industry is highly complex, with companies both fiercely competitive and partnering with each other at key steps along the way. Nvidia and AMD, for instance, design chips but don’t manufacture them. They have to use foundries like TSMC to make the physical product. Companies like Apple or Tesla aren’t chip firms but they often design specialized chips for their own products and then contract with other companies to make them. At the same time, they buy chips from the pure semiconductor players to use in their wares as well. And then companies like Intel or Samsung are integrated; they both design and manufacture chips. But that’s not to say they don’t have to play well with their competition. Every manufacturer has different capabilities, and believe it or not, Intel needs TSMC to make some of its more advanced chips. Suffice it to say that the industry is as interconnected as a computer network.

So why isn’t stock market darling Nvidia, the most magnificent of the Magnificent Seven, up near the top of this list? For that matter, where is Qualcomm? And AMD?

Semiconductor inventions appear under a number of classifications, and not every chip company is active in the proxy code of H01L. Such is the case with Nvidia (and others), which specializes in different patent classifications. Under the H01L umbrella, Nvidia has fewer than 100 patents over the past five years in the U.S. One example of Nvidia’s semiconductor patents in this classification: a multi-chip package for testing inter-chip communications (US-10317459-B2). Nvidia is active in areas such as deep learning, image analysis and enhancement, video and pattern recognition, and control units. All of these innovations are less about the hardware of semiconductors that we see in the current chart and more about the software side of the technology that goes into the design of the company’s graphical processing units (GPUs) and associated software (CUDA). Both are key components for developers to advance their AI, which is why Nvidia’s offerings are so indispensable right now. (More on Nvidia’s patent classifications below).

Bottom line though: with seemingly every company racing to adopt ChatGPT-like generative AI models, the sophistication involved with manufacturing the necessary chips should give rise to more patent protection in the future.

Semiconductor Technology Focus Groups

Patent applications are a leading indicator for the technologies that companies are pursuing at the moment, and different companies tend to develop their own specialties. In early 2023, IBM, which had topped IFI’s annual patent rankings list for nearly three decades, announced a different strategy of more selective patenting in just a handful of focus areas. In other words, having the greatest number of patents was no longer key to Big Blue; patenting in particular areas became the company’s new line of IP attack. One of the focus areas to make the selective patenting cut? Semiconductors. Within that space, IBM’s top innovations are around information retrieval, interface arrangements, machine learning, error detection and program controls. TSMC, on the other hand, concentrates on the machinery around chip manufacturing.

Aside from the proxy H01L semiconductor patent class, we thought it worthwhile to single out Nvidia, the superstar on Wall Street (current stock gyrations notwithstanding), and its areas of patenting focus. The company emphasizes technologies in the G06 class, which covers image data generation and simulators connected with the computing calculations of expected outcomes in an actual system. Computer systems based on biological models would be an example of a technology in this class. One of this class’s key technologies is deep learning. In particular, a form of it called convolutional neural networks, which helps machines see and interpret images as humans do. It is a proficiency that is foundational to generative AI and one of the reasons Nvidia’s GPUs are in high demand by AI developers.

When it comes to inventing measured by patents, the top three spots in AI chips, according to our calculations, go to IBM, Samsung, and Intel, which have been applying for the most in the past five years. Nvidia comes in ninth place.

One note for this section: with patents, well-established classification systems are used to refine searches within specific technologies. The patent coding system usually trails leading edge technologies, and currently, there is no single patent class that covers AI-related GPUs, though one will likely emerge down the road. So we fashioned our own search using technologies used in AI and GPUs.

Who Has the Highest Value Patents?

Even in the area in which Nvidia excels, the company doesn’t have the most patents. However, it has what we would consider the highest value patent, determined by forward citations.

First a primer: To receive a patent, an invention must be novel compared to what has come before, called prior art, in the patent industry lingo. When an invention is filed, the inventor references prior art in order to prove the uniqueness of the new idea. Those references are known as citations. With patents, citations are measured both backward and forward. A backward citation brings up previously granted patents that are foundational to the new invention. The more backward citations a patent mentions, the less original the invention is deemed to be. The improvement tends to be accretive and isn’t expected to disrupt an industry or make huge commercial strides. Forward citations, in contrast, measure how often subsequent patents refer to an earlier patent and indicate an opposite effect; the more forward citations a patent receives, the more potentially disruptive and valuable the technology is. If a startup company has a forward citation graph that looks like a hockey stick, it has invented something that is a big deal, and investors should take note.

Nvidia’s patent on safety systems for autonomous vehicles (US-20190258251-A1), granted in 2023, has the most forward citations (527) in the AI chip grouping. What does this runaway number tell us? Nvidia has technology that has caught the interest of other patent applicants. This one patent is necessary to many other patents—a big plus for the chip designer. Nvidia also holds the second most cited patent, granted in 2022, with this one on training and testing autonomous machines (US-20190303759-A1). Third is Intel with its patent for a way to speed up machine learning (US-20190205737-A1). It’s interesting to note that Nvidia and Intel each hold four spots in the top ten most cited patents.

Looking at the same forward citation information by company, Nvidia is again out far ahead with a significant number of highly relevant patents to other inventions. Intel is no slouch on that front either. Trailing far behind the two are Microsoft, Samsung and IBM.

Another way to slice the numbers is to divide out companies by the number of patents receiving at least 10 forward citations. In that sense, Intel comes out on top with 23 patents, while Nvidia is second with 11, about half of where Intel stands.

All of these metrics are certainly confidence building to any Nvidia investor. But they’re particularly compelling when it comes to Intel, the once mighty chipmaker that missed key opportunities over time and has lost its footing. On one day in early August, the stock plummeted more than 25%, its worst trading day performance in 40 years, after a terrible earnings release, followed by the announcement that it would cut 15% of its workforce and suspend its dividend. Plenty of Intel stockholders have sold and run for cover. But for the investor who thinks the company still has a turnaround story ahead and has the patience to hold the stock for years, these patent data points can bolster a carefully considered investment thesis that Intel will eventually deliver returns (That is, if it's not taken over by Qualcomm, as recently reported, or some other corporate suitor).

NPUs: The Newish Chips on the Block

Simply sticking with what’s tried and true in business can be a perilous stance if management becomes complacent. CPUs aren’t going anywhere. They’re foundational. But key chipmakers fumbled on the significance of GPUs, and when Generative AI took the world by storm, they were caught flatfooted and are now scrambling to catch up to Nvidia.

A new chip is starting to gain traction. Neural processing units or NPUs are chips designed for powering AI neural networks and act as helpers to CPUs and GPUs for AI-driven tasks. They can run immense workloads with less power. Intel’s Core Ultra integrates an NPU alongside CPUs and GPUs. So does Qualcomm’s Snapdragon.

The technology has been around for some years, and the numbers are still small in comparison to other processors. But IFI CLAIMS has found increasing references to NPUs in the text of semiconductor patents. When we conducted a specialized search within the area of AI and chips, the numbers were far greater. Samsung is the leader in this burgeoning technology with 120 patent applications in the past five years. A South Korean company named DeepX is second, followed by Arm and Qualcomm, which, early this year, was granted this patent (US-11861484-B2) for an NPU with a direct memory access core.

Are we seeing the beginnings of a new chip race taking shape? Possibly. Considering how much complex AI workload that NPUs can shoulder, these chips could decrease dependence on Nvidia, which might help realign the competitive landscape down the road. Whatever the outcome in the long term for NPUs, the conscientious investor should be paying attention to these patents.