Source: Content compiled and translated from the WSJ.

We are in the microchip era, which signals the imminent arrival of an industrial revolution where artificial intelligence (AI) will be applied to nearly all human activities.

NVIDIA Corporation is a paragon of this era. With a market value of approximately $5 trillion, it stands as the world’s most valuable company. Jensen Huang, NVIDIA’s founder and CEO, delivered a compelling speech last week at the company’s AI conference held in Washington, D.C. During his keynote address, Huang elaborated on the advancements achieved by NVIDIA’s chips. He expressed gratitude to former President Trump for bringing chip manufacturing back to the U.S. from Asia through energy policies, which have boosted the domestic production of AI microchips in the country.

Most of NVIDIA’s latest chips feature plastic packaging and resemble ants or beetles in appearance, with "legs" made of copper wires. Each chip contains up to 20.8 billion transistor switches and is priced at around $30,000. A revolutionary breakthrough lies in the fact that these data center chips no longer operate independently like the central processing units (CPUs) in laptops. Instead, within data centers, tens of thousands or even millions of such chips are interconnected to form a "hyperscale" computer, whose collective computing power is known as artificial intelligence (AI). Colossus 2, located in Memphis, Tennessee, is one of the world’s top-tier data centers and serves as the core of Elon Musk’s xAI. As the backbone for Grok and autonomous vehicles, Colossus 2 integrates approximately one million NVIDIA chips into a single massive computer.

"Chips" have captivated people in our era to such an extent that even manufacturers of new devices refer to their potential successors as "giant chips" or "super chips." In reality, however, these new devices are the polar opposite of microchips—they lack independent processing units or memory and are not enclosed in plastic casings with wire "pins."

The U.S. government regards the chip industry as a crucial strategic sector. The 2022 CHIPS and Science Act authorized over $200 billion in funding to support domestic chip manufacturing in the U.S. From the Netherlands, home to ASML—the leading manufacturer of chip-making equipment—to Taiwan and its powerhouse semiconductor firm TSMC, microchips are shaping U.S. foreign policy. TSMC controls over 95% of the market for cutting-edge chips, which are used in smartphones and other advanced devices.

By restricting access to the Chinese chip market—a market that houses the majority of the world’s semiconductor engineers—U.S. industrial policies have hindered the growth of American manufacturers of wafer fabrication equipment (critical for chip production). Yet, these policies have not slowed China’s rise. Since the introduction of these protectionist policies around 2020, China’s output of semiconductor capital equipment has grown by 30% to 40% annually, compared to approximately 10% annual growth in the U.S. during the same period.

This shift mirrors the impact of the U.S. ban imposed on Chinese telecom giant Huawei starting in May 2019. The ban led to a $33 billion decline in sales from U.S. companies to Huawei between 2021 and 2024, even as Huawei’s global market share expanded.

Industrial policies and protectionism almost always favor existing industries that are at risk of obsolescence. In this regard, the CHIPS Act, along with its associated bans and tariffs, is no different from Louisiana’s subsidies for ethanol blending in gasoline or sugar beet cultivation, or subsidies for rare earth mining despite the development of new technologies at Rice University that can efficiently extract rare earths from electronic waste. All efforts to salvage U.S. microchip production are taking place amid clear signs that the microchip era is drawing to a close.

The intricate physical properties of key machinery—what some of us refer to as "EUV machines"—clearly demonstrate this. These properties determine and limit chip size and density. The latest version, manufactured by ASML, enables high numerical aperture extreme ultraviolet lithography. If you are not Chinese, you can purchase an EUV machine for approximately $380 million. To date, around 44 units have been sold. Each machine is shipped in roughly 250 crates and requires hundreds of specialized engineers to install over a period of about six months. Dario Gil, IBM’s Director of Research, has called it "the most complex machine in the world."

EUV lithography functions like a "camera." It projects light patterns onto a "film" or "photoresist" on the surface of a 12-inch silicon wafer through a quartz-chromium photomask that contains the chip design.

The fundamental operation of an EUV machine lies in the combination of physical laws and engineering constraints, which ultimately boils down to the EUV photomask. The photomask determines chip size, which in turn dictates the granularity of AI computing. Consequently, the EUV photomask determines how many graphics processing units (primarily from NVIDIA) need to be connected to perform a specific AI task. Beyond a certain threshold—approximately 800 square millimeters, or 1.25 square inches—the laws of light speed restrict larger designs.

The impact of photomask limitations is evident in the growing complexity of hyperscale data centers defined by NVIDIA. The result—smaller, denser chips and "chiplets," each with its own complex packaging—requires a more thorough reintegration of processes to achieve consistent results. Computing must first be distributed across multiple chips and then recompiled. This leads to higher communication overhead between chips, necessitating more complex packaging, additional cables, and fiber optic links.

The inevitable size limitations of photomasks are bringing an end to the chip era. What comes next? The wafer-scale integration model, which will completely bypass individual chips. Mr. Musk first pioneered this concept during his tenure at Tesla, when he launched the now-disbanded Dojo computer project; today, DensityAI has revived this technology.

Cerebras, a company based in Palo Alto, California, has applied this concept in its WSE-3 Wafer-Scale Engine. The WSE-3 contains approximately 4 trillion transistors—14 times more than NVIDIA’s Blackwell chip—and offers 7,000 times more memory bandwidth. Instead of placing memory in distant high-bandwidth memory networks such as chips and chiplets (the traditional approach), Cerebras embeds memory directly onto the wafer. The company has stacked 16 layers of its wafer-scale engine, condensing an entire data center into a small box with 64 trillion transistors.

Also working to build the wafer-scale future is David Lam, founder of Lam Research Corp.—the world’s third-largest manufacturer of wafer fabrication equipment. In 2010, Mr. Lam founded Multibeam, a company that has developed equipment capable of multi-column electron beam lithography. This technology allows manufacturers to overcome the size limitations of photomasks. Multibeam has already demonstrated the ability to etch 8-inch wafers.

The post-microchip era is on the horizon, where data centers will be integrated into boxes housing wafer-scale processors.

Disclaimer: This article is created by the original author. The views expressed in the article are those of the original author. Our reprinting is solely intended to convey a different perspective and does not imply our endorsement or support of this view. If you have any objections, please contact our backend team.