A Tour of the Semiconductor Industry

This week’s newsletter features a first tour of the semiconductor industry. The semiconductor industry faces some of the world’s hardest engineering problems, but interestingly hasn’t been a popular startup focus or academic research area for some time now. This week, we will provide a high level (and necessarily incomplete) tour of the industry, highlighting some interesting new trends along with challenges. This post will be the first in a series of posts highlighting different aspects of the semiconductor industry.


Semiconductor manufacturing is really important. Semiconductor companies like Nvidia increasingly don’t manufacture their own chips and instead outsource to dedicated chip manufacturers like TSMC (in Taiwan). These manufacturers (mostly located outside the US) are assuming outsize geopolitical and business importance. Investing in semiconductor manufacturing research and startups is necessary for the US to retain technology security in the 2020s and beyond.

Informal History

At a high level, the semiconductor industry manufactures chips. The chip powering the phone, tablet, or computer on which you’re reading this newsletter was produced by the semiconductor industry. Chips are complex electronic circuits (made of millions, billions, or even trillions of transistors) that provide a substrate upon which modern operating systems and programs can execute. It’s not unreasonable to consider the semiconductor industry as the bedrock foundation for all of computing.

Let’s start off with some history. Most of the American technology industry has focused on software for the last few decades. Moore’s law (the observation that the number of transistors on a chip tends to double roughly every two years) provided a steady drumbeat of speed improvements on chips that allowed companies to almost take progress for granted. In the past, it often wasn’t worth over optimizing code since the next generation of chips would arrive shortly and be faster anyways. Unfortunately though, over the last decade we’ve started to see persistent slowdowns in Moore’s law as it becomes more and more complicated to build chips

Note that the x-axis in the above diagram (source) is labeled with physical sizes (5nm, 7nm, etc.) This size is called the “process node” for the chip. Historically, this would refer to something like the physical size of a transistor, but is now more of a naming convention rather than anything concrete (discussion). Ignoring the different colors, the heights of the bars show the ballooning design costs to design chips at each new process node.

Why are costs rising here? The short answer is physics. As every part of the chip gets smaller, the physical effects that have to be managed become considerably more severe. Heat dissipation, noise, and quantum effects become harder and harder to control. Designing a cutting edge process node is now an act of high science, with increasingly exotic technology required to make everything work. Moore’s law has been slowing for over a decade but has arguably finally reached a breaking point. This slowdown means that building a “foundry” or “fab” (basically a chip manufacturing plant) has become more and more expensive. It now costs tens of billions of dollars to build a world class foundry.

This trend towards increasing complexity of manufacturing has spurred a movement towards fabless semiconductor companies. Many semiconductor companies have decided to outsource manufacturing to dedicated foundry companies (which focus solely on manufacturing chips for customers rather than designing their own chips) rather than building and maintaining their own foundries. This saves money and lets fabless companies focus on design rather than manufacturing. Nvidia is one of the most prominent fabless companies and relies on both TSMC (or Taiwan Semiconductor Manufacturing Corporation) and Samsung Foundries for its underlying manufacturing. The world’s dominant chip foundries (TSMC and Samsung) are located outside the US which poses unique challenges for American semiconductor companies.

Intel vs TSMC

To provide some more color to the evolution of the semiconductor industry over the last few decades, let’s do a deeper dive comparison of Intel and TSMC. Historically, Intel was the dominant player in the semiconductor space. Let’s take a look at Intel’s stock price since the 1980s.

Although the price doesn’t match technical history exactly, it is true that Intel’s relative technical position has weakened over the last 20 years. Intel has had persistent trouble breaking into the cell phone chip market, shrinking Intel’s market share in the overall semiconductor space. Another challenge has been Nvidia’s continuing dominance of the GPU (or graphics processing unit) space (with AMD coming up as a strong second). Increasingly, GPUs dominate gaming and deep learning workloads. Compounding its troubles, most recently Intel announced that its 7nm process node will be significantly delayed.

Intel’s troubles are notable since Intel is the last major semiconductor company that operates its own foundries in addition to designing its own chips. However, the increasing complexity of both design and manufacturing means that Intel has been struggling on both ends. It is likely that Intel will have to retool its strategy and shift its focus in order to excel.

A parallel trend we haven’t mentioned yet is the entrance of big tech into the semiconductor space. Google’s TPUs (tensor processing units) were introduced in 2016 as custom chips for deep learning workloads. Other players have followed suit. For example, Apple’s new M1 chip was designed wholly in-house and allows Apple to avoid using Intel or AMD chips to a greater degree.

How are these new players able to create competitive chips? The short answer is because of TSMC. TSMC is one of the most inspiring modern stories in business. Morris Chang founded TSMC in 1987 at the age of 56 after a 25 year career at Texas Instruments and over the last 30+ years has built it into the dominant player in the world semiconductor industry. TSMC started by providing excess manufacturing capacity to other semiconductor companies and has step by step improved its technical excellence. Here’s a plot of TSMC’s stock price over the last 20 years.

TSMC is now the dominant semiconductor foundry in the world with a market cap (at time of writing) of ~$600B. For comparison, Intel’s market cap is ~$240B. (If you’re interested to learn more, here’s an amazing translated history of the company.) TSMC has been dramatically scaling up its R&D investments recently.

There are a number of other foundries out there. Samsung Foundries is probably TSMC’s closest competitor. TSMC is headquartered in Taiwan and Samsung Foundries in South Korea. It’s worth noting this since it’s an important sign that the center of the semiconductor manufacturing business has shifted to Asia from the US. As will be discussed later, this situation creates some dangers for American companies.

It’s worth noting though that there are still major foundries in the US. Intel does the majority of its manufacturing in the US. GlobalFoundries is another major foundry with large manufacturing operations in the US and elsewhere. (Interestingly though, GlobalFoundries is owned by Abu Dhabi’s Mudabala). However, it’s worth noting that GlobalFoundries decided to stop their 7nm development and focus on specialty chips, so GlobalFoundries cannot compete with TSMC on the cutting edge. One other player worth highlighting is Skywater foundries, based in Minnesota. Here’s a recent profile. Skywater primarily focuses on manufacturing at the 90 nm node and serves a number of specialty customers such as DWave (the quantum computing company) and genomics companies (building microfluidic chips for DNA sequencers). Skywater also works with DARPA and academic labs on carbon nanotube transistors.

Geopolitical Considerations

One of the biggest themes of the current decade (which we will return to repeatedly in this newsletter), is rising competition and tensions with China. The Chinese government has acted in an increasingly empowered fashion, cracking down on Hong Kong, ratcheting up tensions with India, and ramping up military activity in the South China sea. One of the long time tenets of Chinese foreign policy has been the conquest of Taiwan. The rising profile of TSMC has even led some to examine whether seizing control of TSMC could be a factor in accelerating military action against Taiwan.

One of the bitter realities of the 2020s is that America’s foreign profile has been so weakened that it may well not be possible for the US to prevent a conquest of Taiwan. Given this reality, one of the most important things for American semiconductor companies to do is to make sure they have a backup plan in case access to TSMC is suddenly cut off by invasion, blockade or coup. One option is to have some degree of redundant manufacturing with Samsung or GlobalFoundries. These foundries may not be at the cutting edge of TSMC, but the alternative is to face the likelihood of having an entire product line fall off a cliff. Apple, for example, is looking into shifting some M1 manufacturing to Samsung.

The United States defense and semiconductor establishment is aware of the risks from dependence on TSMC (article). However, it isn’t easy to fix the issue. In future weeks, I hope to discuss some of the policy changes that have been floated recently to encourage growth of the American semiconductor industry.

Musings on Opportunities

This has been a long post, so kudos if you made it down this far! If you’re an entrepreneur or scientist or engineer, what can you do with your new knowledge of the semiconductor industry?

For academics, I’d encourage taking a closer look at semiconductor manufacturing research. This field is now red hot, powered (as we have just discussed) by geopolitical tensions and the demise of Moore’s law. If your lab can invent a new technology to manufacture chips more efficiently, I suspect there will be many American government agencies interested in funding your work. On a note of foundational science, chip manufacturing features some of the hardest problems in materials design, physics, and computational modeling.

One long standing opinion of mine is that AI research will inevitably have to turn to semiconductor manufacturing. We’ve known for some time that the computational needs of AI workloads are ever increasing (analysis). There are only so many chips in modern clouds. We’re already nearing a point where even large cloud providers are struggling to handle heavy AI training workloads. Google, Apple, and Nvidia (among many others) have been investing in new AI chip designs, but they haven’t yet started investing in chip manufacturing. I suspect this will change over the coming decade as these companies become more aware of the potential speed-up and performance gains from improved manufacturing techniques.

For entrepreneurs, I’d caution that the semiconductor industry is extremely hard to break into. The capital requirements are large, and hardware startups (aside from the lucrative AI chip startup space) haven’t gotten much airtime. Semiconductor manufacturing in particular has extremely high capital expenditure (capex) requirements that make it very hard for new startups to establish themselves. That said, I have a suspicion that there are strategies that could enable cheaper manufacturing that will develop over the coming decade as more research focuses on modern semiconductor manufacturing. I anticipate this could be a rich space for new startups and daring entrepreneurs in the years to come.

As a parting thought for the week, I’ll note that this has been a very cursory tour. More to come soon!

Highlights for the week

Feedback and Comments

Please feel email me directly (bharath@deepforestsci.com) with your feedback and comments! In particular, if you’re currently working in the semiconductor industry, please get in touch! I’d love your input for future iterations in our semiconductor series.


Deep Into the Forest is a newsletter by Deep Forest Sciences, Inc. We’re a deep tech R&D company specializing in the use of AI for deep tech development. We do technical consulting and joint development partnerships with deep tech firms. Get in touch with us at partnerships@deepforestsci.com! We’re always welcome to new ideas!


Author: Bharath Ramsundar, Ph.D.
Editor: Sandya Subramanian