Making Tokens, Pt. 2: Pulling the Crystal

Part two of the Making Tokens series. In Part 1 we turned quartzite into 9N polysilicon. Now we have to turn that polysilicon into wafers, which means we have to coax a single perfect crystal of silicon out of a melt. Every wafer in every chip in every datacenter started its life as a single crystal that took two days to grow.

The constraint: every atom in the same place

A transistor only works if the electrons flow through silicon whose atomic lattice is undisturbed. Polycrystalline silicon (the kind that comes out of a Siemens reactor) is full of grain boundaries: places where one crystallographic orientation meets another. Carrier mobility tanks across those boundaries. So before you can build a chip, you have to convert poly-Si into a monocrystalline ingot, where every atom sits in the same regular lattice from one end to the other.

There are a couple of ways to do this. The dominant one, by an enormous margin, is the Czochralski process (CZ), invented in 1916 by the Polish metallurgist Jan Czochralski more or less by accident. He dipped a pen into what he thought was an inkwell and pulled out a wire of crystallized molten tin. The semiconductor industry has been pulling crystals out of melts the same way ever since.

The mechanics

Inside a CZ puller you have:

• A quartz crucible holding ~200-300 kg of polysilicon chunks
• A graphite susceptor wrapped around the crucible, heated to 1414 °C (silicon's melting point) by inductive coils
• An argon atmosphere to keep the melt clean
• A small seed crystal of monocrystalline silicon, with a precisely defined orientation, mounted on a wire that can be raised, lowered, and rotated independently of the crucible

The procedure is unsentimental and slow. The crucible heats up. The poly melts into a clear, near-mercury-colored pool. The seed is lowered until it just kisses the surface. The melt freezes onto the seed, atom by atom, in the same crystallographic orientation. Then you start pulling. Slowly.

A typical pull rate is 1-2 mm per minute. You rotate the seed in one direction and the crucible in the other to keep the thermal field axisymmetric. The crystal grows downward into the melt and is hauled upward into the cooler region of the puller, where it solidifies. Over the course of about 24-48 hours, you grow an ingot somewhere between 1.5 and 2.5 meters long and 300 mm in diameter, weighing 150-200 kg.

The atmosphere inside the puller is dead silent. There's no clatter of machinery. The crystal grows in slow motion. Operators check it through small windows.

Why this is hard

The CZ process looks deceptively simple, but it is governed by some of the more annoying coupled physics in industrial chemistry.

Heat transfer. You're balancing the heat input from the susceptor against the heat removed by conduction up the growing ingot and by radiation from the melt surface. Get the balance wrong and you get either constitutional supercooling (the melt freezes faster than the seed can guide the crystal structure) or a runaway shape change. Modeling the thermal field is a serious CFD exercise.

Mass transport. As the crystal grows, impurities segregate. Some elements (like boron) prefer to stay in the melt; others (like oxygen, dissolved from the quartz crucible) get pulled into the crystal at a known rate. This is what gives you the doping profile along the length of the ingot. The first wafers sliced from the top of an ingot are often a slightly different doping concentration than the wafers from the tail.

Vibration and convection. The melt has internal Marangoni convection (driven by surface tension gradients) and buoyancy-driven flow. These need to be controlled, often with a static magnetic field, to keep the growth front flat. "Magnetic CZ" (MCZ) is the standard for the highest-quality wafers used in leading-edge logic.

Crucible erosion. The quartz crucible dissolves slowly into the melt at 1414 °C. The resulting oxygen incorporation in the crystal is actually useful at controlled concentrations (it helps getter metallic impurities), but uncontrolled, it ruins the wafer.

The throughput economics are dictated by all of this. A modern CZ puller costs millions of dollars, runs for years, and produces somewhere around 100-200 ingots per year depending on size and recipe.

From ingot to wafer

What comes out of the puller is one giant rod of single-crystal silicon. The next stage is purely mechanical and pretty brutal.

Cropping: the ends of the ingot are sawn off and recycled. The seed end has the seed cone, the tail end has the segregated impurities.

Grinding: the ingot is ground to a precise outer diameter (300 mm) and a flat or notch is ground along one side to mark crystallographic orientation.

Slicing: a diamond wire saw slices the ingot into individual wafers, each about 775 microns thick. The wire is itself thinner than a hair and is impregnated with industrial diamond. The slicing process produces a lot of kerf loss: roughly 30% of the silicon by mass ends up as silicon dust that gets washed away in the slurry. This is one of the largest sources of waste in the whole upstream chain.

Lapping and grinding: the wafer surfaces are flattened mechanically.

Etching: a quick chemical etch removes any subsurface damage from the mechanical steps.

Polishing: chemical-mechanical polishing (CMP) brings one side of the wafer to an atomic-scale smoothness. The mirror finish you see in promo photos is the polished front-side. The back is rougher.

Cleaning: RCA cleans, megasonic cleans, repeat. By the time the wafer is boxed, the front surface has fewer than a handful of particles per cm² above the cleaning threshold.

What ships to the fab is a bare polished 300 mm wafer, weighing about 125 grams, in a hermetically sealed FOUP (Front Opening Unified Pod) containing 25 of them, and worth roughly $60-100 per wafer for the standard grade.

A single 2-meter ingot yields about 1500 wafers

Let me put a number on the input/output math:

• ~200 kg of polysilicon goes into the puller as raw material
• ~180 kg comes out as a usable ingot (after cropping)
• ~125 kg ends up as actual wafer material (after kerf loss)
• That works out to roughly 1500 polished 300 mm wafers per ingot

A leading-edge logic die for an AI accelerator (like an H100) is about 814 mm². A 300 mm wafer has about 70,650 mm² of usable area. So in theory you could fit 80 H100 dies per wafer, but in practice yield losses (defects in the silicon, defects in the lithography, edge exclusion) bring that down to maybe 50-60 working dies per wafer, and that's after months of fab processing that hasn't happened yet.

Per ingot: roughly 80,000 H100-class dies out of 200 kg of starting polysilicon. Or: every 2.5 grams of 9N silicon eventually becomes one AI chip.

The four companies

The wafer industry is even more concentrated than polysilicon. Four companies make almost every monocrystalline wafer used in chip production:

• Shin-Etsu Chemical (Japan)
• SUMCO (Japan)
• Siltronic (Germany)
• GlobalWafers (Taiwan)

Together they hold something like 90% of global wafer capacity. The barriers to entry are real: the CZ pullers cost millions, the process recipes are decades of accumulated know-how, and the quality bar gets higher every chip generation as feature sizes shrink. A new entrant would need years and billions to compete on the leading edge.

This is the quietest oligopoly in tech. Nobody writes about it. The chip companies (TSMC, Samsung, Intel) get all the attention. But every chip TSMC makes, every chip Samsung makes, started its life as a wafer from one of these four firms.

What's next

By the end of this stage, we have a clean, polished, single-crystal silicon wafer worth about $75. In the next step, photolithography, we will subject that $75 wafer to roughly 1500 individual process steps over four months and turn it into an AI accelerator die worth $15,000-25,000. That's a 200-300x value-add in three months of process time. The capex required to do it is in the tens of billions per facility.

Photolithography is where this stops being chemistry and starts being something else.