The AI Pipeline, Part 1: Dirt, Sand and the Software That Draws Chips

The short answer: AI hardware begins as high-purity quartz and silicon (China alone mines about 5.4 million tonnes of silicon a year), refined into wafers and drawn into chip designs by three software companies (Synopsys, Cadence, and Siemens EDA, holding roughly 70 percent of the EDA market) on two proprietary architectures: x86, controlled by Intel and AMD, and ARM, owned by Arm Holdings and running about 99 percent of the world's smartphones.

You type a prompt. Half a second later, an answer appears. What happened in between is one of the most complex supply chains humans have ever built, and it passes through a surprising number of chokepoints where a single company, sometimes a single factory, controls the entire world's access. Sand from a handful of quarries. Quartz refined in a few furnaces. Chip design software made by three companies. One Dutch machine that prints the chips. One Taiwanese company that fabricates most of them. Three memory makers. Three clouds. If any single link in that chain fails, AI development does not slow down, it stops.

I have wanted to write this for a while, because most AI commentary stays at the model layer, and the models are the least interesting part of the power structure. So this is a series. Five parts, one layer at a time, from the ground up to the data center, with the geopolitics that runs through every single link. Part 1 starts literally at the bottom: the minerals, and the software that turns minerals into blueprints.

Everything starts with dirt#

Before silicon, there is sand. Not any sand: high-purity quartz, the kind found in commercially useful quantities in only a few places on Earth. It is the most extracted solid material on the planet. The world pulls roughly 50 billion tonnes of sand and gravel out of the ground every year for construction alone, a rate far above what rivers and beaches naturally replenish, and the supply of the high-purity stuff needed for electronics is shrinking. Your smartphone contains a meaningful quantity of it, your laptop more.

Silicon itself is mined at industrial scale. China alone mines about 5.4 million tonnes of it a year. But mined silicon is not a wafer. It gets purified to nine-plus nines of purity (99.9999999%, one foreign atom per billion), pulled into single-crystal ingots, sliced into discs polished to near-atomic flatness. Every wafer is a finished industrial product before a single transistor is printed on it.

And the pipeline's mineral dependencies do not stop at silicon. Rare earths, the 17 elements that make magnets, motors, missiles, phones, and hard drives possible, have a geography all their own. China mines about 70% of the world's rare earths and refines close to 90%. That refining share is the real chokepoint, and it was built deliberately. Deng Xiaoping reportedly said back in 1987 that the Middle East has oil and China has rare earths. Three decades of patient industrial policy later, that is simply the map.

Then there is gallium, the metal that goes into the GaAs and GaN compounds used in radio chips and power electronics. China accounts for about 98% of primary gallium supply. Not 70%. Ninety-eight.

Why does this matter for AI? Because when this part of the chain got weaponized in 2023-2025 (export curbs, price shocks in terbium and dysprosium, gallium exports falling to zero), the effects did not stay in the mining sector. They showed up inside the AI supply chain, as you will see in part 3. The minerals are the first layer, and the first lever.

The design layer: three companies draw every chip#

Once you have wafers, someone has to design the chip. A modern AI GPU has over a hundred billion transistors. Nobody draws that by hand. It is designed in software called EDA (electronic design automation), and that software market is effectively a three-company affair: Synopsys, Cadence, and Siemens EDA (the old Mentor Graphics). Together they hold roughly 70% of the global market, and every serious chip on earth, NVIDIA's, AMD's, Apple's, Qualcomm's, Huawei's, was drawn in their tools.

Think about what that means for the chip war. The United States does not need to block a chip physically. It can restrict the software used to design it. Which is exactly what happened when EDA tool access became one of the levers in the US-China fight.

The instruction sets the chips speak are equally concentrated. x86, the architecture in your laptop and most servers, is controlled by exactly two companies: Intel and AMD, a duopoly so stable that Intel holds over 80% of PC processors and roughly 99% of the server market. ARM, the architecture inside essentially every smartphone (about 99% share), is owned by one company, Arm Holdings, which collects licensing fees on nearly every phone on earth while being small enough that people forget it exists.

Then there is the design houses themselves. NVIDIA, AMD, Apple, and Qualcomm are fabless: they design chips but fabricate nothing. This is not laziness, it is economics. A leading-edge fab costs tens of billions of dollars, so design companies rent manufacturing capacity instead. The natural follow-up question is: who owns the machines and the fabs that everything gets rented from? That is part 2, and it is the layer where the concentration gets almost comical.

The full series#

• Part 1: Dirt, Sand and the Software That Draws Chips (you are here)
• Part 2: One Machine and One Island
• Part 3: The Buyers and the Export Control Era
• Part 4: The Geography of Compute and the India Lens
• Part 5: Nobody Owns the Stack, Everyone Owns a Switch