The Hard Bottleneck at the Edge of AI Compute: America’s Silicon Steel Monopoly and Its Investment Logic

In financial markets, we are forever hunting for the kind of “choke point” that everyone calls a stranglehold.

Author: Koutian Wu; GitHub: ktwu01

Today I was reading a research report on the US power grid supply chain, and one detail in it stopped me cold.

Frankly, over the past year everyone has talked about AI, about GPU capacity, about where Jensen Huang went and how big the deal was. All eyes were glued to TSMC’s CoWoS packaging line, treating it as the lifeline that determines the fate of large models.

But if you pull your view outward from Silicon Valley’s data centers to the substations out in the middle of nowhere, you find a much harder, much more traditional bottleneck.

Here’s the thing.

This is the classic bottleneck investment logic. Consider it: AI data centers’ power draw is growing exponentially, and grid modernization is moving in parallel. Those two forces converge on an enormous demand for power infrastructure.

And the transformer is the heart of the grid. Without it, even if you built a nuclear fusion plant, the electricity would never reach your H100 cluster.

But this is not something you can make whenever you want. The core material in a transformer is called grain-oriented electrical steel, GOES for short. Its grains are specially processed to align in a single direction, giving it superb magnetic properties, and it is used specifically for transformer cores.

What makes your scalp crawl: in the entire United States, only one company, Cleveland-Cliffs, can produce this steel.

Can you believe it???

Even though the US still has old-line steel mills like U.S. Steel, what they produce is mainly non-oriented electrical steel, which goes into electric vehicle motors and cannot be used for large-scale grid transformers.

This artificial restriction on supply, against a backdrop of exploding demand, has directly produced a massive supply chain disaster. Lead times are now absurdly long; sometimes you wait years for a transformer. The US government has even listed this as a national security vulnerability. If the slightest thing goes wrong at Cleveland-Cliffs’ Ohio plant, America’s ability to build transformers basically grinds to a halt.

At this point some people may wonder: if the monopoly is this ironclad, can’t I just blindly buy Cleveland-Cliffs stock and be done with it?

That’s what I thought at first.

But when I mulled it over, I couldn’t fully convince myself. You see, steel is ultimately a heavy-asset, low-margin cyclical industry. Even if you monopolize one niche like silicon steel, the company’s overall share price will still be dragged around by ordinary construction steel and automotive steel prices. Cleveland-Cliffs has been posting losses in its last couple of quarterly results, which says a lot.

The ones who really skim all the profit off the top are the downstream giants that take that silicon steel and build transformers.

Following on from the above: rather than buying cyclical stocks of raw materials, consider the “water sellers” who sell transformers and grid equipment.

Eaton, for example (ticker ETN).

This company has now become almost the most solid foundation in AI infrastructure. Look at their Q3 results: data center-related orders surged nearly seventy percent year over year. That is not traditional industrial stock growth; it is pure tech-stock euphoria.

What moved me most was the order backlog in their hands. Their core segment sales grew fifteen percent, but the backlog grew twenty percent. That is like cushioning yourself with a multi-billion-dollar safety net. Even if the economy suddenly slows tomorrow, the work they have stacked up could keep them busy for years.

And they are genuinely attacking the bottleneck. Eaton just poured over $300 million into a new plant in South Carolina specifically to build large transformers, expected to come online in 2027. They also spent $100 million in Texas expanding capacity. And they even acquired the liquid cooling company Boyd, because everyone knows chips like Blackwell keep getting hotter and traditional air cooling simply cannot keep up.

One link after another.

Still, it is a bit jarring to say, but Eaton now trades at a P/E of over thirty times. For an industrial company, that is a fairly expensive price. The market is pricing them like a tech stock, so the error tolerance is very low. If growth slows even a little, the share price could see a pullback.

My own unpolished suggestion: if you are also bullish on this track, do not go all-in at once. Start a watch position first, and add gradually after a pullback to below a P/E of twenty-five. Or, to keep it simple, just buy an infrastructure ETF like GRID or PAVE, sweeping up GE Vernova and Quanta Services, the ones who build and repair the grid, all in one go.

For a moment I was speechless. We grind ourselves into the ground in frontier tech, trying to use large models to change the world. And yet the physical-world Transformer gets strangled by the capacity of a handful of traditional steel mills.

That feeling is too surreal.

It is like spending millions on a top-tier supercar, only to discover the entire town has one gas station, and the owner is in a bad mood today and decides to ration fuel.

This is the undercurrent coloring of this era, folks.

Just when you think we are sprinting toward AGI, the friction of the physical world always yanks you back from an unexpected place. Whether it is Nvidia’s chips or America’s domestic silicon steel capacity, it all reminds us that the revelry of the digital world ultimately gets paid for by the iron and fire of the physical world.