The Heart of Intelligence ― GPUs and the Chokehold of Silicon

In the previous episode, we saw that the machine called AI grows by feeding on a new kind of oil—data. But oil, on its own, is not power. It needs a furnace to burn it in: a blazing furnace of computation that turns vast quantities of data into intelligence.

That furnace goes by the name of semiconductor. And burning at the hottest center of the furnace is a black plate only a few centimeters across—the AI chip.

The struggle for supremacy over intelligence converged, in the end, not on ideas, nor on talent, nor on data, but on this small crystal of sand. The heart of AI was beating on silicon.

From Accident to Strategic Resource ― The GPU as a Heart

The story begins in an unexpected place. The GPU (graphics processing unit) was originally a component for rendering the 3D imagery of games smoothly. It computes, in parallel, the countless pixels that make up a single screen. That quality—doing a great many simple calculations all at once—would later change the fate of AI.

The deep learning revolution of 2012, which we saw in Episode 2, was itself triggered when researchers lined up gaming GPUs and repurposed them for training. Training a neural network is, in essence, the endless repetition of enormous matrix calculations. GPUs, running thousands of small arithmetic units at the same time, were orders of magnitude faster than conventional CPUs that handle operations one after another.

In this way, a computing device once dismissed as a toy transformed into the heart that generates intelligence. The larger an AI model grows, the more GPUs must keep spinning and roaring for ever longer stretches of time. The chipmaker NVIDIA was pushed from being a maker of gaming chips into being one of the most critical suppliers of the AI age.

The One Who Turns Sand into Crystal ― Taiwan’s TSMC

But here a question arises. Who makes that heart?

Design and manufacturing demand separate skills and separate capital. A plant (a fab) that etches the most advanced chips requires investment on the order of trillions of yen to build, and it draws its circuits with atomic-level precision. The companies in the world capable of this can be counted on both hands.

Standing at the summit is Taiwan Semiconductor Manufacturing Company—TSMC. Founded in 1987 by Morris Chang, the company brought a revolution to the semiconductor industry. Until then, it had been taken for granted that a company would manufacture the chips it had designed in its own factories. Chang severed that link. TSMC would not design at all; it would only take on and manufacture the chips that other companies designed—the world’s first “foundry” (a pure-play contract manufacturer) business model.

Precisely because it held no competing designs of its own, TSMC gathered the trust of its customers in full. The circuit diagrams drawn by minds the world over—NVIDIA foremost among them—flowed into the factories of Taiwan. According to reporting and industry materials, TSMC is said to have come to manufacture the great majority of the world’s most advanced chips single-handedly. The heart of AI was beating on a small island called Taiwan.

The world of manufacturing equipment shows the same pattern. The extreme ultraviolet (EUV) lithography machines indispensable to the most advanced chips can, in effect, be made by only one company: the Netherlands’ ASML. The light sources that etch the circuits, the chemicals that polish the wafers—each is concentrated among a small handful of companies somewhere in the world. The very process of inscribing intelligence onto sand was concentrated in the hands of a very limited few. Design in the United States, manufacturing in Taiwan, equipment in the Netherlands and Japan—the heart of AI rested on a long and delicate chain of international division of labor that no single nation could ever forge alone. This very “concentration” and “division of labor” would later become an enormous strategic card.

Tightening the Grip on the Throat ― The Export Controls of 2022

The heart of intelligence, the factory that makes it, and the equipment that runs the factory. AI’s lifeline was thus strung together along a long chain. And the chain, when traced, leads largely back to American technology and its allies. If someone seizes a single point along the chain, they can stop the heart of a rival far away—nations were beginning to grasp this fact.

In 2022, the United States seized that single point. That year, sales to China of NVIDIA’s most advanced AI GPUs, the A100 and H100, were restricted. Then, in October 2022, the U.S. Department of Commerce rolled out measures broadly regulating exports to China of advanced-computing semiconductors and the equipment used to manufacture them. The aim, as it was explained, was clear: to prevent China from diverting cutting-edge computing power to military and surveillance uses.

In response to the controls, NVIDIA was reported to have introduced alternative chips for the Chinese market, such as the A800 and H800, whose performance had been partly throttled. But the United States is said to have caught such derivatives in its regulatory net through subsequent measures, and the cat-and-mouse game of restriction and evasion continued. What is depicted here is no longer a mere matter of trade. It is a struggle to sever the very roots of a rival’s AI development at the highest point upstream—the supply of computing power itself.

There is a sense of déjà vu in this story. In the twentieth century, it was Middle Eastern oil that held the fate of nations. At the dawn of the twenty-first, the semiconductor—the brain of every electronic device—took up that mantle. And now the cutting edge of that semiconductor holds the throat of a new kind of power: intelligence. Resources, components, intelligence—the wellspring of power keeps changing its form, yet it has always thrust upon humanity the same question: who holds it?

How far the controls will prove effective, and how far China can catch up on its own, are matters on which assessments differ, and no firm conclusion can be drawn at present. What is certain is a shift in the structure itself: AI supremacy came to be contested over an invisible supply network of computing power—a chain known as the supply chain.

Who will hold the heart burning on sand? Once that question had been set down, AI was no longer a dream of the laboratory but a national strategy in itself.

And in late 2022, the intelligence born of that scorching heart finally slipped from the hands of researchers and engineers. An era in which anyone could hold an AI in the palm of their hand—as something to talk to—was quietly about to begin.