There is one object that the world’s largest companies are fighting over with billions of dollars, that states are trying to stockpile like a strategic reserve, and that sometimes ends up smuggled across borders. It’s neither a smartphone nor a vaccine: it’s an electronic chip, the GPU. To understand why it became so precious is to read a map of power — where every link in the chain is a point of leverage.

What does a GPU actually do?

A GPU (Graphics Processing Unit) was born for one task: computing images, in video games. But rendering a 3D scene means performing millions of identical little calculations at the same time. Where a conventional processor (CPU) has a few very powerful cores that handle tasks one after another, a GPU lines up thousands of simpler cores that work in parallel.

This architecture, designed for pixels, turned out to be perfect for artificial intelligence. Training an AI model essentially means stacking up enormous matrix multiplications — exactly the kind of massively parallel calculation a GPU devours. The same chip that used to run a game now runs ChatGPT.

The truest image: during the AI gold rush, GPUs are the shovels and picks. And the one selling the shovels never loses.

The craze: demand that exploded

From 2023 onward, the rise of generative AI turned an already strong demand into a frenzy. Every major cloud company — see the major cloud providers — wanted to build its own “compute factories.” Delivery times stretched to several months, and one player swept the board: Nvidia, which designs the most coveted GPUs and corners the overwhelming majority of the AI market, in part thanks to its in-house software (CUDA) that the entire industry has learned to use.

The result: a handful of chip models became a global bottleneck. Not because we’re short of ideas — because we don’t know how to make enough of them.

The bottleneck: it’s not just a question of factories

Here’s a strategic nuance many people miss. Nvidia doesn’t manufacture its chips: it designs them. Production is handed to a subcontractor, the foundry. And for cutting-edge chips, there’s essentially only one: TSMC, in Taiwan.

But the real chokehold isn’t even etching the chip. It’s two lesser-known steps:

• Advanced assembly (packaging): AI GPUs bond the compute chip to stacks of ultra-fast memory on a single substrate (a technology called CoWoS at TSMC). This assembly capacity has long been the bottleneck of production.
• HBM memory (High Bandwidth Memory): these memory stacks are made by only three players (SK Hynix, Samsung, Micron). No HBM, no GPU.

In other words, the shortage doesn’t come from a lack of sand or silicon. It comes from a handful of steps that very few factories in the world know how to perform.

The mandatory passage points

When you reason like a strategist, you look for the chokepoints: those places where everything has to pass through, and which you only need to block to bring everything to a halt. The GPU chain concentrates two of them, both formidable:

• TSMC, in Taiwan. The bulk of the world’s most advanced chips comes out of a single island, at the heart of a zone of tension between China and the United States. One disruption in Taiwan, and the entire global industry teeters.
• ASML, in the Netherlands. The EUV lithography machines, indispensable for etching the finest chips, are made by a single company in the world. No cutting-edge factory can exist without it.

Two companies, two countries: that’s what an immense share of the global digital economy hangs on.

Helium and neon: the invisible dependencies

This is where the story becomes dizzying. Even owning the factories and the machines, you can’t make a chip without certain rare gases — and their supply, too, is dangerously concentrated.

• Neon. The lasers that etch the chips need neon of very high purity. Yet a large share of semiconductor-grade neon historically came from Ukraine (a byproduct of steelmaking). The war that broke out in 2022 brutally disrupted this source and sent prices soaring — a chilling reminder that a regional conflict can paralyze factories on the other side of the planet.
• Helium. Indispensable to chip manufacturing (inert atmosphere, cooling, leak detection) and to the operation of data centers, helium is a non-renewable resource, extracted alongside natural gas. Its global production rests on a handful of countries (United States, Qatar, Algeria, Russia), and the sector has already gone through several severe shortages. A gas we associate with party balloons is, in reality, a critical link in the most advanced industry in the world.

The strategic lesson: vulnerability is never where you look for it. It nests in the most inconspicuous link — often a consumable we had forgotten to count.

The cold war of compute

All these dependencies have become weapons. Since 2022, the United States has restricted exports to China of the most powerful GPUs and of the machines that make them, dragging the Netherlands (ASML) and Japan along in its wake. The goal: to slow Beijing’s access to cutting-edge compute.

China struck back on its ground of strength: by limiting exports of metals it dominates (gallium, germanium, rare earths), essential to electronics. Each presses on the other’s chokepoint.

Ultimately, a conviction has taken hold in capitals: compute has become a strategic resource, just as oil was in the 20th century. Whoever controls the GPUs controls the speed at which a country — or a company — can develop its AI. Hence the race for local fabs (the massive investment plans in the United States, in Europe, in Japan) and the idea of “sovereign compute.”

Reading the map like a strategist

What should you take away, without being an engineer?

1. Scarcity is organized, not natural. Silicon is abundant; what’s missing is a few ultra-specialized industrial capacities. They can be built — but it takes years and fortunes.
2. Resilience plays out on the weak links. A chip depends on an island, on a Dutch machine, on a Ukrainian gas. The real strategic question isn’t “who has the best chip?” but “who holds the tap?”
3. Compute is a geopolitical asset. Watching the diversification of factories, export controls, and tensions over raw materials tells you more about the future of AI than the next spectacular demo.

In one sentence

The GPU rush isn’t a geek bubble: it’s the battle for the fuel of AI. And like any strategic resource, its value rests not just on the chip itself, but on the long, fragile chain that makes it possible to produce — right down to the gas we thought was reserved for birthday balloons.