Everyone's talking about the US chip ban. But the real restrictions were imposed by China.

I'll tell you about a component.

You've probably never heard of it.

It's in your phone's touchscreen. It's in satellite communications. It's in fiber optic cables.

It's in AI data centers, in the components that allow thousands of chips to communicate with each other.

Without that component, chips can work individually, but not together.

Approximately 70% of its production is in China.

This component is called indium.

Indium isn't mined on its own.

It's obtained as a byproduct when processing zinc ore. So, if you want to increase indium production, it's not enough to just look for more indium; you need to refine more zinc.

That's why it can't be accelerated by simply paying for it.

First, you need to grow another industry, and a large part of that industry is in China.

So what exactly does this component do in artificial intelligence?

Everyone is measuring the competition by chips. How many chips does each company have? How many parameters was used to train each model?

However, a data center isn't just a stack of chips.

When training a large model, thousands of chips are working simultaneously, constantly sending data to each other.

So those thousands of chips have to act like a single machine.

Within the same rack, chips communicate via copper wires. But as the cluster grows, data between racks is transmitted by light.

The lasers that produce that light are based on indium phosphide.

Adding chips doesn't change the outcome if the connection slows down, because the system operates at the speed of the slowest link.

China restricted exports of this material in February 2025.

In the following 14 months, China's indium exports globally fell by two-thirds, and to the US by 77 percent. The price of indium phosphide flakes rose from approximately $1,400 to $5,000.

The result can be seen in one company.

Lumentum, the American company that produces these connectors, quadrupled its production. Despite this, it has no stock left to sell until 2028.

Not because demand is high. Because it couldn't find raw materials.

To solve this bottleneck, Nvidia committed $2 billion each to Lumentum and its competitor Coherent. Lumentum is building a new facility in the US, and production will begin in mid-2028.

So, the world's most valuable chip company wrote checks to two suppliers to solve a material shortage, and the solution will begin at the earliest in two years.

Why isn't this being discussed?

Because chips are quantifiable. They have a brand, a price, and a market value. The number of chips a company buys becomes news.

The interconnect layer isn't visible. Nobody discloses how many optical modules they buy.

So the public is measuring the competition with the wrong unit.

When China imposed restrictions on rare earth elements, this was discussed, and counter-policies were developed.

The same move was made with indium, and it was hardly discussed.

It's the same weapon used without much fanfare.

Now let's return to the initial picture.

The US chip bans prevented China from accessing advanced AI chips, and this has largely worked.

But the two restrictions aren't preventing the same thing.

One stopped the chip itself. The other stopped the chips from working together.

Making a chip and making thousands of chips work together as a single machine are two separate tasks. In the first, China lags behind; in the second, it holds the gate.

Moreover, this isn't the only bottleneck.

When building a data center, it needs to be connected to electricity before the chip. Transformers make that connection.

China has approximately 60% of the global transformer production capacity. The US only meets about 20% of its own demand.

Orders for large power transformers can take up to 4 years. A data center, on the other hand, is a project planned to be built in 18 months.