The Anatomy of a Compute Reserve

If compute is a strategic reserve, it helps to understand what the reserve actually consists of, because the answer reveals where the pressure points lie. A useful way to think about it is as a stack of four dependent layers: the chips themselves, the machines that make the chips, the facilities that house the chips, and the energy that powers everything. A weakness at any layer constrains the whole.

Start at the top with the chips. The processors that matter for frontier AI are specialized accelerators, most prominently graphics processing units, or GPUs, adapted from their origins in video rendering to perform the parallel arithmetic that machine learning demands. A single company, Nvidia, designs the majority of the world's most sought-after AI accelerators. It does not manufacture them itself. That distinction matters enormously, because it means the design of the most important chips and the manufacture of them sit in different hands and different countries.

Drop to the next layer: fabrication. Turning a chip design into a physical processor requires a semiconductor foundry, and the most advanced foundries on Earth are astonishingly concentrated. The Taiwan Semiconductor Manufacturing Company, TSMC, produces the overwhelming majority of the world's leading-edge chips. A handful of others, in South Korea and increasingly the United States, can approach the frontier, but the concentration in Taiwan is so extreme that a single island has become the most strategically significant piece of industrial real estate on the planet.

Drop again: the tools that make the fabrication possible. To etch circuits mere nanometers wide, foundries rely on extreme ultraviolet lithography machines, and precisely one company in the world makes them: ASML of the Netherlands. Each machine costs on the order of hundreds of millions of dollars, contains hundreds of thousands of components, and cannot be replicated quickly by anyone, at any price, without decades of accumulated expertise. This is the deepest choke point in the entire stack. Whoever influences ASML influences who can build advanced chips at all.

Now the physical layer most people overlook: the data center. Chips do not train models sitting on a shelf. They must be assembled into vast clusters inside purpose-built facilities with specialized cooling, networking, and power delivery. Building these at frontier scale requires land, construction, water for cooling, and above all electricity, in quantities that rival small cities. A single large AI training facility can demand hundreds of megawatts of continuous power, which is why the map of future AI capacity increasingly overlaps with the map of available electricity.

That brings us to the base layer: energy. You can own the chips and the buildings, but without abundant, reliable, affordable power you cannot run them. Energy is emerging as a binding constraint on AI expansion, which is why national compute strategies now routinely involve power generation, grid upgrades, and even direct arrangements with nuclear and natural gas suppliers. The reserve is only as deep as the electricity behind it.

Stack these layers and a striking picture emerges. The design of chips is concentrated in the United States. Their manufacture is concentrated in Taiwan and South Korea. The tools to manufacture them are concentrated in the Netherlands and Japan. The assembly into working clusters and the energy to run them are contested everywhere. No single nation controls the whole stack. Every major power has leverage over some layers and dependency on others.

My read is that this interdependence is the master fact of AI geopolitics. It is what makes export controls possible in the first place, because a choke point is only useful if your rival must pass through it. And it is what makes those controls dangerous, because pressure at one layer ripples through all the others and invites retaliation at whatever layer the other side happens to control. When the United States restricts chip designs, it is pulling a lever it holds; when a rival restricts the rare earth minerals that go into the machines, it is pulling a lever it holds in return.

Understanding the anatomy of the reserve is therefore not a technical detour. It is the foundation for reading every headline that follows. Before you can grasp why export controls reshape the global compute map, you have to see the map as it truly is: not a single resource in a single place, but a fragile chain of dependencies stretched across allies and rivals alike, where a squeeze at any link is felt along the entire length.