The AI trend has brought microchip manufacturing back to the forefront of computer technology
Technological challenges are more important than political ones, argues Shailesh Chitnis.
Disclaimer 1: This is a free translation of Shailesh Chitnis's column for The Economist. The translation was prepared by the editors of Technocracy. To stay up to date with new material, subscribe to “Voice of Technocracy” — we regularly talk about news about AI, LLM and RAG, and also share useful mustreads and current events.
To discuss the pilot or ask a question about LLM, please Here.
Disclaimer 2: the text mentions the company Meta, which is recognized as an extremist organization in the Russian Federation
A century ago, 391 San Antonio Road in Mountain View, California, was an apricot-packing warehouse. Today, it’s just one of many low-rise office buildings on busy streets that house Silicon Valley startups and would-be billionaires. But in front of the building stand three large, strange sculptures — two-legged and three-legged shapes that resemble water towers. They’re giant versions of diodes and transistors, the components of an electronic circuit. In 1956, Shockley Semiconductor Laboratory opened on the site, a startup dedicated to the idea of making such components entirely out of silicon. And so began the history of Silicon Valley.
The company founded by William Shockley, one of the inventors of the transistor, failed to achieve commercial success. But its silicon work did not. In 1957, eight of Shockley’s employees, whom he called “traitors,” left to found Fairchild Semiconductor less than a mile from the lab. Among them were Gordon Moore and Robert Noyce, future co-founders of microchip giant Intel, and Eugene Kleiner, co-founder of Kleiner Perkins, a pioneering venture capital firm. Most of Silicon Valley’s famous tech companies can trace their roots, directly or indirectly, to these early Fairchild employees.
Before the advent of semiconductor components, computers were room-sized machines that used fragile, unstable vacuum tubes. Semiconductors, solid materials in which the flow of electrical current can be controlled, offered components that were more reliable, versatile, and compact. When such components began to be made primarily of silicon, it became possible to fit many of them onto a single piece of the material. Tiny transistors, diodes, and other components on silicon “chips” could be combined into “integrated circuits” designed to store or process data.
In 1965, Moore, while still at Fairchild, noticed that the number of transistors that could be placed on an integrated circuit at a given cost was doubling every year (he later adjusted this to two years). His observation, known as Moore's Law, turned out to be extremely significant. Chips produced in 1971 had 200 transistors per square millimeter. In 2023, the MI300 processor, created by the American company AMD, contained 150 million transistors in the same area. The smaller the transistors became, the faster they could turn on and off. MI300 components are thousands of times faster than their predecessors from 50 years ago.
Every major advance in computing, from personal computers and the internet to smartphones and artificial intelligence (AI), can be traced back to transistors becoming smaller, faster, and cheaper. The development of transistor technology has been the engine of technological progress.
For a time, the technological importance of silicon chips was reflected in the importance of the companies that made them. In the 1970s, IBM, which made both the chips and the computers and software that went with them, was an unrivaled giant. In the 1980s, Microsoft proved that a company that sold only software could be even more successful. But Intel, which made the chips that ran Microsoft’s software, was also a formidable force. Before the dot-com crash of 2000, Intel had the sixth-largest market capitalization in the world.
After the crisis, Web 2.0 services offered by companies like Google and Meta came to the fore, and the semiconductors on which their platforms were built began to be perceived as commodities. To explain this growth dynamic in the big tech sector, venture capitalist Marc Andreessen declared in 2011 that “software has taken over the world, not silicon.”
But the AI boom has changed all that. Developing AI requires enormous amounts of computing power. Until 2010, the amount of computing power needed to train leading AI systems increased according to Moore’s Law, doubling every 20 months. But now it’s doubling every six months. That means demand for more powerful chips is growing. Nvidia, the American company that specializes in chips ideal for running the large language models (LLMs) that dominate AI, is now the third-most valuable company in the world.
As AI has made chip manufacturing a trend again, companies looking to succeed in the field are starting to design their own chips. This is not just about training the models, but also about using them later (called “inference”). Using LLMs is computationally intensive, and the process must be done billions of times a day. Because specialized circuits can perform these tasks more efficiently than general-purpose chips, some LLM companies are choosing to design their own chips. Apple, Amazon, Microsoft, and Meta have all invested in building their own AI chips. Google has more processors in its data centers than any other company except Nvidia and Intel. Seven of the world’s ten most valuable companies are now chip makers.
The technological complexity of a chip depends mainly on the size of its elements; the cutting-edge technology is currently measured in parameters below 7 nm. It is at this level that key processes for AI occur. However, more than 90% of the semiconductor industry’s production capacity works with elements of 7 nm and larger. These chips are less technologically complex, but more common – they are used in TVs, refrigerators, cars and machine tools.
In 2021, at the height of the COVID-19 pandemic, a severe shortage of such chips disrupted production across industries including electronics and automobiles. The industry, in an effort to be efficient, had globalized: the chips were designed in America; the equipment to make them was in Europe and Japan; the factories that use the equipment was in Taiwan and South Korea; and the chips were packaged and assembled into devices in China and Malaysia. When the pandemic disrupted these supply chains, governments took notice.
In August 2022, the U.S. government proposed a $50 billion package of subsidies and tax breaks to bring chip manufacturing back to the U.S. Other regions have followed suit, with the European Union, Japan, and South Korea pledging nearly $94 billion in subsidies. The situation has been complicated by U.S. efforts to limit China’s access to advanced chips and the technology that makes them through export bans. China has responded to those bans by restricting exports of two materials vital to chip production.
But the biggest challenges for chipmakers aren’t industrial policy or national rivalries, but technology. For five decades, shrinking transistors has improved chip performance without increasing power consumption. Now, as transistors per unit area increase and AI models become larger, power consumption is skyrocketing. To keep up with exponential performance gains, new ideas are needed. Some are incremental, like tighter integration of hardware and software. Others are more radical, like rethinking silicon use or abandoning digital processing in favor of other methods. In future posts, we’ll explore how such advances can keep the exponential engine going.
