Semiconductors. The oil of the 21st century. In 2022 more than 1 trillion microchips were manufactured, roughly 130 chips for every person on earth. Semiconductors have taken the front seat as supply chain issues brought industries to a standstill in 2021, geopolitical tensions have increased, and AI has hit an inflection point. Although the US makes up a large part of semiconductor demand, China became the largest user of semiconductors in 2021, buying chips to the tune of $190+ Billion. Although the US still dominates in the design and equipment segments, its manufacturing capacity market share has decreased from 37% in 1990 to about 12% today1 , trailing both Japan and Taiwan. This was a result of skyrocketing manufacturing costs along with global industry competitiveness from specialized players. Taiwan and South Korea are now the top manufacturers of leading-edge chips, with Intel, who was the “King of Chips” in the 90’s and early 2000’s, a distant third.
The current flux in the semiconductor industry could provide fertile grounds for attractive investments. This industry report is meant to provide an overview of the semiconductor industry, the players in the space, and trends.
What is a semiconductor? Semiconductors are made from materials such as Silicon, Germanium, Gallium Arsenide, Silicon Carbide, and others. What has made them the cornerstone of the modern world is that the electrical conductance in these materials can be precisely controlled by the addition of “impurities”, which is called “doping”. This feature gave birth to the first “Killer App” of the Semiconductor era, the invention of the Transistor (1948), where a narrow semiconductor region sandwiched between two other regions can either modulate the larger current going through the other two regions, acting like an amplifier, or it can completely switch on/off the current, acting like a digital switch. The second “Killer App” was the Integrated Circuit, invented by the Fairchild Semiconductor Company around 1960, which gave birth to the modern method of manufacturing computer chips, with Transistors, Resistors, Capacitors, and Interconnects all part of the same Silicon chip, all built on a flat wafer – it is called the Planar Process – no more testing each one separately. Two of the Fairchild founders, Robert Noyce and Gordon Moore, then left Fairchild and founded Intel, which gave birth to the third “Killer App” of the Semiconductor Age, the Microprocessor, and Moore’s Law, which was the Big Bang, so to speak, and 50 years later here we are with Laptops, Cellphones, Supercomputers, the Internet, EV’s and AI and so many other revolutions in technology based on semiconductors.
Semiconductors today are overwhelmingly composed of super high purity silicon (by “super high purity” I mean 99.99999999999% or 11 Nines of purity). These silicon ingots are created and then sliced into 12 inch wafers, each half a millimeter thick. The surface of these wafers can be protected by laying down a layer of Silicon Dioxide, SiO2, aka Quartz in one of its many forms. This “planar process”, using Silicon Dioxide passivation is the not-so-secret sauce of modern semiconductor manufacturing. Thus Silicon became the King of Semiconductors.
To get an idea of the mind-boggling progress made in this industry, in the 1950’s, the typical transistor was 1 centimeter long! This shrank to millimeters by the late 1950’s, and then came the integrated circuit. But it took a while for the Integrated Circuit to dominate. For example, the famous IBM 360, which came out in 1964, used “modules” of discrete Transistors, Capacitors, Resistors, still no Integrated Circuits. But then came the chip containing the first microprocessor (1971), the Intel 4004 “die”, before encasing it in its “Dual inline Package”, measured 3 mm by 4 mm and contained 2,300 Transistors. Today the Nvidia H100 chip contains over 80 Billion transistors, the sides of its die measure 28 millimeters, and like all modern chips, it is built on 12” Silicon wafers. The wonders of Moore’s Law!
The activities that make up producing a chip are (1) Design (2) Manufacturing (3) Testing and Packaging. Designing a chip is solely on the software side and is done by fabless players such as Nvidia. Manufacturing is how chips are created from silicon ingots. The steps for manufacturing microchips from a silicon wafer involve 1) photolithography 2) etching 3) doping 4) deposition 5) polishing. This entire process takes up to 15 weeks from start to finish. After the chips (or dies) are cut, they need to be tested and packaged for final sale. Even though semiconductor manufacturing is atomically precise, the yield is about 60% for leading-edge chips with trailing edge chips having a higher yield of around 80%. Roughly a couple hundred to about 15k chips are produced per wafer.
Semiconductors are expensive and very light thus makes shipping parts to other countries more conducive. The industry has become a complex global interweaving of partners. Semiconductor parts on average cross-country borders about 70 times before becoming an end product. This hyper-specialization benefits in an ideal world but has significant drawbacks in a non-ideal one. The years 2020 and 2021 exposed the fragility of these supply chains. In March 2021, Renesas Electronics, a company based in Japan that manufactures roughly 30% of automotive micro-controllers, had a fire at their plant and it took 100 days to reach normal baseline production. This caused massive ripple effects in the automotive industry, bringing plants to a standstill.
By end use application, currently about 30% of semiconductors are used for mobile phones, 25% are used for data centers and IT infrastructure, 20% are used for computers, and the rest are used for automobiles and consumer electronics. The semiconductor industry revenue is forecasted to reach about $1T by 2030 as EV adoption and computational and data storage needs increase. Semiconductor costs per vehicle are expected to go $500 dollars today to over $4k dollars by 2030, an 8x increase.
Although the name “chips” is thrown around ubiquitously, it is important to distinguish the types of chips as the players and industry dynamics vary significantly. There are primarily three types of chips made: these are Logic, Memory, and Analog. About 42% of chips produced are logic, 26% are memory, 14% are analog, and the other 18% represent non-integrated circuit semiconductors. Chips are grouped as leading edge or trailing edge. This roughly means that the width of the gate of the transistor is below 6 nanometers versus above 6 nanometers. CPUs, GPUs, or ASICs (application-specific integrated circuits) all refer to logic chips.
A table of the players and respective market shares is shown below2.
The players in the semiconductor ecosystem are OEMs, integrated device manufacturers (IDMs), Foundries, Fabless, and EDAs.
The OEMs provide manufacturing equipment that goes into manufacturing facilities. Equipment ranges from etching to photolithography to testing. Most tools in a leading-edge fab can deposit, polish, or etch materials within the precision of a handful of atoms. ASML is a notable player as it essentially has a monopoly on photolithography (the large orange bar in the graph below in lithography is ASML2). Photolithography, or the patterning of chips with focused flashes of Extreme Ultraviolet Light (EUV) that pass through a photomask, requires the most expensive and sophisticated machines. To highlight how complex the lasers being used are, replicating it requires identifying and assembling 457,329 parts3.
The critical software used to design chips is called Electronic Design Automation or EDA. Semiconductor design used to be done in-house by semiconductor companies but with the arrival of the foundry business model, EDAs and foundry effectively bifurcated the design stage from manufacturing. EDAs design the circuitry but have to coordinate their solutions with IP providers such as NVIDIA and ARM as well as PDKs (process design kits) from foundries. PDKs are design constraints so that the chip is compatible with manufacturing processes. The largest EDA players are Synopsys and Cadence however there is a push to open-source EDA software. This will be hard to do with IP providers continuing to want a hefty payment. Currently, the cost to design a 7nm chip is $300mm and the cost to design a 5nm is $550mm and takes about 12-20 months.
Integrated Device Manufacturers, or IDMs, as the name implies, are vertically integrated from design to the final testing of the chip. Intel and Samsung are IDMs, however to compete with TSMC, they recently have been opening up their foundries to third party designers or fabless companies. IDMs have seen significant market share erosion as the industry has become modularized with the rise of fabless, EDAs, and foundry companies.
Foundries are pureplay contract manufacturers for the semiconductor industry. TSMC is the largest of these with more than 50% market share. TSMC dominates the leading edge, being the first fab to come out with 7nm and 5nm chips. Customers of TSMC are Apple and Nvidia. Other large foundries are Global Foundries, UMC, and SMIC. These business models revolve around volume as only then can a manufacturer sink billions of dollars to develop a leading-edge foundry. TSMC alone will spend around $30 billion in capex in 2023.
Fabless players solely design chips and then partner with foundries to manufacture those chips. Players in this space are AMD, Nvidia, Qualcomm, Broadband, and Mediatek. The rise of the foundry business model in the late 80s allowed the creation of these players.
The end of Moore’s Law
Moore’s Law — which states that the number of transistors per integrated circuit will double every 2 years — has held for almost 60 years. This law of competition has enabled a lower cost of computation, more power efficiency etc. However, transistors per chip are nearing a physical limit with the most advanced chips having a gate length of 3 nm. Creating these leading-edge chips are also becoming prohibitively expensive with only the largest companies being able to utilize them such as Apple. The cost to build a leading edge fab is almost $20B (up 20x from 2000)4 and doubling transistor density will see a corresponding doubling in capital investment.
As a result, other parts of the supply chain will see an inflection in demand. Advanced packaging is beginning to see this as the combination of multiple chips could see efficiency and computing gains. Advanced packaging helps gains in efficiency due to system scaling instead of transistor scaling. It is currently done by IDMs and other smaller players such as Micross, Skywater, and Qorvo. China has signaled aggressive expansion into this domain to offset its weakness in leading-edge chip manufacturing.
Customization of chips
The trend of chip customization will increase with the rise of IOT and automotive devices. Those chips need to have a variety of metrics tweaked such as power efficiency, latency, communication etc. Demand will come from niche applications in which hardware is a major differentiator or it enables different functionalities that generic chips can’t meet. The supply side will drive the time and costs lower from the design stage to manufacturing. As these are not leading-edge chips, there will be a push for more open-sourced design software and new business models. Skywater is a US-based foundry whose business model is predicated on these ‘the long tail’ of custom chips.
The semiconductor cold war will continue
The United States relies on TSMC for 92% of its leading edge chips. This extreme dependence on a foreign entity poses a national security risk to the US and the Chip Act aims to address this concern with $39B in manufacturing incentives meant to establish at least two leading fabs in the U.S., along with an additional $11B for R&D to support and sustain the industry. The notion is to seed the semiconductor manufacturing environment, subsequently seeing a multiplicative effect from public and private investments.
Similar actions have been taken in Europe (European Chips Act) and in China. However, without the continued support of the government, US foundries will likely suffer. Significant capex is needed every year to sustain technological prowess. TSCM’s capex plan alone for 2023 is $30 billion which almost matches the aggregate US funding amount.
There will be continued retaliation between countries such as what has recently been occurring between US and China. The US since 2020 has restricted US advanced semiconductor technology to Chinese companies such as Huawei and recently China banned Apple products from use by government employees. China has proven resilient however and recently Huawei introduced a 7nm, cutting edge phone. This has prompted in further US clampdowns such as barring Nvidia from supplying A100 chips to China.
Another aspect of domiciling semiconductor manufacturing aside from protecting supply chains is increased security. Chip security is becoming ever more important with the insertion of any compromised microelectronics possibly affecting millions of devices.
AI eats the world
The rise of AI has recently occurred because Deep Neural Network technology took off. This in turn was due to the increase in computational ability offered by the evolution of Graphic Processing Units (‘GPUs’). GPUs are superior to CPUs in tasks that involve parallelizable tasks and High Performance Computation. A GPU contains hundreds of times as many Cores as a CPU, with each Core being independently programmable to some degree, which brings in the fourth Semiconductor Killer App, the age of cheap parallel computation. A ChatGPT 4.0 query consumes 12x more energy than a Google search. Nvidia alone will sell about 500k H100 GPUs in 2023 with each H100 having about 60 Teraflops of processing capacity. Cloud computing is on track to triple in size by 2030 driven by a virtually unlimited demand for computation.
- Semiconductor Industry Association. 2022 SIA Factbook. Semiconductor Industry Association, May 2022. https://www.semiconductors.org
- The White House. Building Resilient Supply Chains and Revitalizing American Manufacturing.
- Miller, Chris. Chip War: The Fight for the World’s Most Critical Technology. Harper Business, 2022