A History Lesson On Semiconductors

The Alchemy of Silicon

Why is silicon miraculous? Unlike a conductor (gold or copper), which lets electrons flow freely, or an insulator (glass or rubber), which blocks electrons entirely, silicon sits in a very peculiar middle ground. Its electrons only need a small nudge of energy to start conducting. That property of semiconductor behavior is actually extremely useful.

The trick lies in doping: deliberately contaminating ultra-pure silicon with traces of other atoms. Add in phosphorus and you create N-type silicon, full of free electrons. Add in boron and you create P-typer silicon, full of “holes” (the absence of electrons that functionally behave like positive charges). If you press these two materials together, you have a P-N junction. This allows current to flow in one direction. This is an example of the simplest active semiconductor device, a diode. Add in a third layer and you have a transistor, a device where a tiny control voltage at the gate can switch a much larger current on or off. The genius of the integrated circuit, pioneered by Jack Kilby (TI) and Robert Noyce (Fairchild) in 1958, was to etch thousands [then millions, then billions] of these transistors onto a single, very small slice of silicon.

Gordon Moore, co-founder of Intel, noticed that the number of transistors that could fit on a chip was doubling every ~2 years. Moore’s Law became an observation/prophecy of this exponential growth in the chip industry that has since sustained – today’s chips contain over 100 billion transistors, which are measured in angstroms (0.1 nanometers).

Early Innings in America

The semiconductor industry was a creature of the American national security program. The first customers for integrated circuits were the US Air Force and NASA. The Minuteman II ballistic missile and the Apollo guidance computer were mass purchasers of chips. This entanglement of chips with military power has since remained. The Pentagon, through DARPA (the Defense Advanced Research Projects Agency), luckily invested early in this notion of compute as a strategic multiplier. The thought was that a nation that commanded more advanced chips could build smarter weapons (and win with less volume but more accuracy), decode more encrypted communications, and run sophisticated logistics.

Silicon Valley grew from this military foundation. Fairchild Semiconductor, founded by the “traitorous eight” who left Shockley’s lab in 1957, was the seed from which most of modern tech came through. Fairchild alumni founded Intel, AMD, and dozens of other companies. By the 1970s, the US dominated global semiconductor production commercially. Intel’s 4004 microprocessor put an entire CPU on a single chip and set the template for the personal computer paradigm that followed.

Japan

The first great geopolitical battle over semiconductors was with Japan – a nominal US ally. Through the late 1970s and 1980s, Japanese firms like Toshiba, NEC, Hitachi, and Fujitsu systematically entrenched themselves in the global memory chip market (with the blessing of America). They did so through government backing (the Ministry of International Trade and Industry), very patient loans from Japanese banks, and a manufacturing discipline with dramatically lower failure rates. By the 1980s, Japanese companies commanded 50%+ of the global semiconductor market.

As a result of this losing American DRAM (DRAM chips provide for temporary storage) manufacturer market share, American semiconductor leaders banded together to form the Semiconductory Industry Association and in 1985 filed a Section 301 complaint alleging that (i) Japan’s government had rigged the domestic market against US chip imports and (ii) Japanese firms were dumping chips below cost in global markets. A year later, the US-Japan Semiconductor Agreement forced Japan to open its market and stop the dumping in exchange for the US pausing its tariffs.

Still, the US learned a much deeper lesson of not allowing any single foreign nation, even ally, to dominate semiconductor manufacturing. So, America cultivated alternatives in Taiwan and South Korea. Over the following decade, Samsung and TSMC would overtake Japan to become the world’s most dominant manufacturers. Chris Miller in Chip War explains at length at the way commanding the semiconductor supply chain enables leverage over the entire global economy.

Foundries and Fabless

In 1985, Taiwan’s Economics Minister (K.T. Li) asked Morris Change (a veteran in TI who was passed up for the CEO position) to build Taiwan a substantial semiconductor industry. Taiwan Semiconductor Manufacturing Company was founded thereafter in 1987.

Chang’s hypothesis was that designing and manufacturing chips under the same IBM/Intel roof was not necessary; he wanted TSMC to be a pure manufacturer (foundry). It would accept other companies’ chip designs and fabricate them at scale. The implication was immense –if a company no longer needed to build its own billion dollar fab, it could focus on design in entirety. Qualcomm, Nvidia, eventually AMD, and Apple’s chip division are examples of companies that now compete on chip design and contract with foundries for manufacturing. TSMC competes on manufacturing quality. It now manufactures more than 90% of the world’s most advanced chips.

Here, it’s really fascinating to think about the globalization of the chip supply chain – both in terms of its efficiency and fragility. American firms designed the chip. A Dutch company (ASML) made the only lithography machines capable of printing them. A Taiwanese company (TSMC) printed them. Korean companies (Samsung, SK Hynix) supplied the memory. Japanese companies supplied the chemicals and materials. Chinese factories assembled the finished electronics.

GPUs

Early AI research in the 1950s/60s was very ambitious but computationally starved. There were two AI winters, in fact, that followed periods of overhyped promise (1970s and the late 1980s). Funding fried up during these periods.

Ultimately, AI’s fortunes were changed by a new chip architecture. Nvidia developed the graphics processing unit (GPU) in the late 1990s, which was designed to perform the parallel floating-point calculations needed to render 3D game graphics. Founders Jensen Huang, Chris Malachowsky, and Curtis Priem had no stated interest in AI – they wanted to win the gaming market.

In 2012, a team at the University of Toronto led by Geoffrey Hinton used Nvidia GPUs to drain a deep neural network (AlexNet). It won the ImageNet visual recognition competition by a massive margin. The reason why this was so monumental was the GPUs’ parallel architecture that permitted matrix multiplication operations at orders of magnitude faster than conventional CPUs.

From this point onward, the history of AI and the history of semiconductor advancement were entangled. AI progress as it followed (image recognition, language models, generative AI, etc.) required increasingly more powerful chips to train and run. Nvidia’s A100 and H100 GPUs became very important commercial products – demand largely exceeded supply with waitlists of over two years.

LLMs like GPT-4 required clusters of tens of thousands of H100s running for months to train. The compute requirements for frontier AI models have been doubling every six to twelve months (faster than Moore’s Law!). The chip is now the engine of AI, not just the consumer internet. Therefore, in the eyes of major governments, the chip is the engine of future economic and military power.

China

Under Xi Jinping’s comprehensive national security plan, Beijing began treating chip self-sufficiency as an existential priority as early as 2014. Semiconductors were China’s largest import (even greater than oil) – and that dependence was something America had noticed.

China’s response was to spend. The National Integrated Circuit Industry Investment Fund (the “Big Fund”) put hundreds of billions of yuan into domestic chip design and manufacturing. SMIC was set up as China’s equivalent of TSMC. Chinese fabless designers like HiSilicon and Cambricon began making technical progress. Xi had a “Made in China 2025” goal that aimed for semiconductors to be domestically produced from zero to one.

America saw this with increasing alarm, and the trigger for confrontation came via Huawei’s 5G equipment. US intelligence agencies argued the telecommunications gear was designed for surveillance and somehow built on chips procured from American technology. In 2018, the Trump administration restricted Huawei’s access to US components. A year after, Huawei was put on the Commerce Department’s Entity List. In 2020, the rules were extended globally – any company that used American chip-making tech needed a US license to supply to Huawei, so TSMC had no choice but to halt the manufacturing of Huawei’s most advanced processors.

We now see the US employing a new strategy of wanting an absolute lead on the global scale. Rather than continuing with the approach of staying a generation or two ahead of rivals, America the nation has poised itself to maintain a total lead – in weapons, AI research, commercial computing, and any application that could compound into a form of military capability over time (this is a very expansive umbrella).

Chokepoints

As I have alluded, the entire global economy rests on a handful of irreplaceable bottlenecks.

First, ASML is a Dutch company headquartered in Eindhoven (right by TU Eindhoven). It is the sole manufacturer of extreme ultraviolet lithography machines. Without these, no one can print chips at advanced process nodes. The EUV machines use laser-generated plasma to produce light with a wavelength of 13.5 nm. This light is then reflected off a sequence of nearly perfect mirrors. Building these machines has required decades of work, hundreds of supplier relationships, and north of 450,000 components. No nation has come close to replicating them. That should really put into perspective the implications of the US’ pressure on the Netherlands to restrict EUV exports to China.

Second, TSMC, as we know, manufactures a vast majority of the world’s chips. TSMC has accumulated operational knowledge of tens of thousands of engineers, who have iteratively refined the manufacturing process over decades. This knowledge appears in the head of the employees, the internal culture, and the ecosystem of Taiwanese suppliers around the country. That’s why I feel that despite massive investment, TSMC’s fabs in Arizona and Japan have taken years longer than planned to reach the Taiwan plants’ average yield rates.

Third, American electronic design automation software is a less visible but still critical aspect. Synopsys and Cadence provide the EDA tools without which it is practically impossible to design advanced chips. US export controls have increasingly targeted this software layer as well.

We can understand the asymmetric quality of the chip industry. The US doesn’t need to be the best chip manufacturer to win; it just needs to control enough chokepoints to deny adversaries access. The export controls of 2022-2025 were effectively an attempt to leverage ASML’s monopoly and America’s dominance in cheap design to impose a ceiling on China. Of course, China has still been making headway in chip innovation on its own. There is an all-out effort to build domestic alternatives (China’s “50% Mandate,” SMIC’s usage of the older deep ultraviolet lithography equipment, and Chinese research on a novel carbon nanotube-based 2D transistor are examples). Whether that will help China reach the forefront of the global standings remains to be seen. While US reshoring is certainly less smooth than perhaps planned, I would bet on the US retaining its lead.

Note: will definitely write more on this soon. There are really exciting/more recent developments.