Making Tokens, Pt. 1: Sand into Silicon

This is the first piece in a five-part series tracing the actual supply chain behind a single AI token. You can see all five steps on the home page. This one is about the first step: turning rocks into the purest material humans know how to produce at scale.

The starting material isn't sand. It's quartz.

People use the word "sand" loosely when they talk about silicon. The actual feedstock is high-purity quartz, not the beach sand you're picturing. Beach sand is contaminated with shell fragments, organic matter, iron oxides, and a parade of other minerals. You can't make a chip out of it.

The real input is quartzite mined from a small number of deposits where the SiO₂ content is high enough (typically >98%) and the trace metals are low enough to be useful. The best deposits in the world are in surprisingly few places: Spruce Pine in North Carolina, the Iberian Peninsula, parts of Brazil, and a handful of sites in China, Russia, and Norway. The Spruce Pine deposit in particular is so unusual in its purity that a meaningful share of the world's high-purity quartz for the semiconductor industry comes from one small region in western North Carolina.

The price of this raw material is around $200/ton for high-purity quartz by the time it leaves the mine, compared to a few dollars per ton for ordinary construction sand. That premium is going to get a lot bigger by the end of this piece.

Stage 1a: Metallurgical-grade silicon

The first chemical transformation is a carbothermic reduction, run in a submerged-arc electric furnace at around 1900 °C:

SiO₂ + 2C → Si + 2CO

You feed quartzite and a carbon source (typically a mix of metallurgical coal, charcoal, and woodchips) into the top of the furnace. Massive electrodes carry hundreds of amps through the bed. The carbon reduces the silica, the molten silicon collects at the bottom, and the carbon leaves as carbon monoxide off the top. What pours out at the bottom is called metallurgical-grade silicon (MGS), which is roughly 98-99% pure.

The energy intensity of this step alone is substantial: roughly 10-15 MWh per ton of MGS. The CO that leaves the furnace is enough that there's a whole secondary discussion about how to handle the CO₂ footprint of silicon, which is one reason the industry is increasingly siting these furnaces near hydropower (Iceland, Norway, Quebec, the Pacific Northwest).

MGS is fine for things like making silicones, aluminum-silicon alloys, and steel. It is many, many orders of magnitude too dirty to make chips out of. You wouldn't get a working transistor out of MGS if you tried for a thousand years.

Stage 1b: From MGS to trichlorosilane

To get to electronic-grade purity (we'll come back to what "electronic-grade" actually means in a moment), the silicon has to be put through a chemical purification loop. The trick is that silicon itself is hard to purify directly, because it has a very high boiling point and tends to drag impurities along with it through any straightforward refining. Instead, the industry converts the silicon into a compound that is easy to purify, then converts it back.

The standard route is via trichlorosilane (SiHCl₃, "TCS"):

Si (MGS) + 3 HCl → SiHCl₃ + H₂

This reaction is run in a fluidized bed at around 300 °C. TCS is a liquid at room temperature with a boiling point of just 32 °C, which means it can be distilled. And distilled. And distilled. Modern TCS plants run multi-stage fractional distillation columns that achieve breathtaking purity, removing boron, phosphorus, iron, aluminum, and the rest of the periodic table down to single-digit parts per billion.

The output of this stage is ultra-pure trichlorosilane, which is then ready for the next conversion.

Stage 1c: The Siemens reactor

Here's where the magic finally happens. Inside a Siemens reactor (named after the German company that developed the process in the late 1950s), you have a bell jar with thin polysilicon "seed rods" hanging vertically and heated electrically to about 1100 °C. You introduce a flow of TCS and hydrogen into the chamber. The reverse reaction takes place on the hot rod surfaces:

SiHCl₃ + H₂ → Si + 3 HCl

Silicon deposits epitaxially on the rods, growing them outward over the course of days to weeks. The HCl byproduct is recycled back to the first conversion step, which is one reason TCS chemistry has dominated polysilicon production for so long: the chlorine cycle is closed.

At the end of a Siemens run, you crack open the reactor and pull out polycrystalline silicon rods that are typically 150-200 mm in diameter and several meters long. These rods are smashed into "polysilicon chunks" the size of fist-sized rocks, which is how electronic-grade poly is shipped to wafer makers.

What does "9N" actually mean

The semiconductor industry talks about purity in "nines." 6N means 99.9999% pure (one part per million of impurity). Each additional "N" is a factor of ten cleaner. Electronic-grade polysilicon is typically 9N to 11N, meaning 99.9999999% to 99.999999999% pure: one part per billion to one part per hundred billion of impurity.

To put a number on what this actually means: at 11N purity, a kilogram of polysilicon contains less than 10 micrograms of total non-silicon atoms. That is less impurity, by mass, than there are humans on Earth in a kilogram of the periodic table.

This is the purest commodity humans produce at industrial scale. Pharmaceutical chemistry has its very-high-purity feedstocks, and isotope separation produces specific things at high purity, but in terms of raw tonnage of an element at this level of cleanliness, polysilicon is the king of the periodic table.

The energy intensity to get here is real. Total energy across the chain from quartzite to 9N polysilicon is roughly 100-110 kWh per kg of poly. The Siemens deposition step alone accounts for a large fraction of that. The industry has spent decades trying to displace it with cheaper alternatives (fluidized-bed reactors that produce granular polysilicon at lower energy intensity, the Union Carbide silane route, monosilane processes), and each new method has won market share for some applications, but the Siemens process remains the dominant route for the very highest purities required by leading-edge chips.

The market is tiny and oligopolistic

Global polysilicon production runs around 600,000 tonnes per year. By the standards of bulk chemicals, this is a small market: a single ethylene cracker produces more output than half the polysilicon industry combined.

The capacity is dominated by a small number of producers:

Wacker Chemie (Germany / Tennessee, US)
GCL-Poly Energy (China)
Hemlock Semiconductor (Michigan, US, jointly owned with Corning)
OCI (Korea / Malaysia)
Tongwei, Daqo, Xinte (China)

The other crucial fact: about 90% of all polysilicon produced in the world goes into solar panels, not chips. The semiconductor industry only buys roughly 10% of global poly output, and it pays a premium for the very-highest-purity grades that solar doesn't need. When you read about polysilicon prices crashing or spiking, that's almost always a story about the solar market. The semi-grade market is a small, stable, expensive corner of it.

The value-add by mass

Here's the punchline of this stage. The pricing roughly looks like this:

Quartzite at the mine: ~$0.20 / kg
Metallurgical-grade silicon: ~$2 / kg
Trichlorosilane (purified): ~$5-10 / kg
9N polysilicon (electronic-grade): ~$25-40 / kg

That is a roughly 150-200x value-add by mass, accomplished entirely by purifying the same atoms you started with. The element doesn't change. You're not adding gold or rare earths. You're only removing impurities. And it is hard enough to remove them that the resulting price multiple is comparable to what you'd see in pharma intermediate chemistry.

This is the first thing a token costs. The next stage, pulling a single crystal out of this polysilicon, is where the chemistry gets boring and the physics gets weird.

Sources and pointers

USGS Mineral Commodity Summaries, "Silicon" (annual): production and reserve data
Bernreuter Research polysilicon market reports: the standing reference on production capacity and pricing by producer
Vince Beiser, The World in a Grain: long-form journalism on industrial sand, including a chapter on Spruce Pine
SemiAnalysis posts on TSMC's silicon supply chain: good for the downstream economics