While the world obsesses over the availability of Nvidia’s Blackwell GPUs, a far more boring—and far more critical—shortage is quietly threatening to choke the AI revolution before it truly begins. It’s not about silicon chips; it’s about steel and copper. Specifically, the Large Power Transformers (LPTs) that step down high-voltage electricity for use in data centers.
In late 2025, the wait time for a new LPT hit a staggering 4 years.
If you order a transformer today, you might get it in 2029. For an AI industry moving at the speed of light, this physical bottleneck is a disaster scenario. Here is why the grid is breaking, and why money alone can’t fix it fast enough.
The Physics of the Bottleneck: Why “Just Build More” Fails
To understand why we can’t just “print” more transformers like we do software, we have to understand the sheer physical magnitude of a Large Power Transformer (LPT). These are not the small grey cylinders on the utility pole outside your house. LPTs are building-sized behemoths, often weighing over 400 tons—heavy enough to require specialized rail cars (Schnabel cars) for transport. They sit at the critical intersections of the grid, stepping voltage down from transmission levels (345kV or 765kV) to distribution levels.
Designing them is less like manufacturing and more like bespoke architecture. Each unit is custom-wound for the specific voltage, impedance, and load profile of its destination substation.
The Magic Metal: Grain-Oriented Electrical Steel (GOES)
The heart of an LPT is its core, which must be made of a very specific material: Grain-Oriented Electrical Steel (GOES).
Standard steel is isotropic—its magnetic properties are the same in all directions. But for a transformer, you need to minimize energy loss (hysteresis) as the magnetic field flips 60 times a second. GOES is engineered so that its crystalline structure is aligned in a single direction, allowing magnetic flux to flow with near-zero resistance. This is achieved through a complex process of cold rolling and annealing at precise temperatures.
The physics of efficiency are brutal. The core loss formula illustrates why material quality is non-negotiable:
Where is the hysteresis coefficient and is the eddy current coefficient. GOES minimizes both. Without high-grade GOES, a transformer would overheat and fail under the massive loads demanded by modern grid interconnects. If you use persistent inferior steel, the “hum” of the transformer becomes a roar, heat builds up, and the unit eventually explodes.
The Manufacturing Cliff & The Cleveland-Cliffs Monopoly
Here is the problem: Making GOES is incredibly difficult. It requires a specialized annealing process at extremely high temperatures and precise rolling techniques.
In the United States, there is exactly one producer of GOES: Cleveland-Cliffs.
Yes, the entire U.S. electrical grid’s future rests on the output of a single company’s mills in Butler, Pennsylvania and Zanesville, Ohio. While they are investing in upgrades, they cannot single-handedly supply the surge in demand.
Globally, the situation isn’t much better. China dominates the global supply of electrical steel, but trade tariffs (Section 232) and national security concerns make relying on Chinese steel for critical U.S. infrastructure a non-starter. This leaves U.S. utilities fighting for allocations from limited friendly suppliers in Japan (Nippon Steel) and South Korea (POSCO), both of whom are already maxed out supplying their own domestic manufacturing partners.
Building a new steel mill capable of producing electrical steel isn’t a 6-month project. It requires billions in upfront capital and a regulatory permitting process that can stretch for half a decade. Even if we broke ground today, that steel wouldn’t hit the market until 2028 or 2029.
The AI Power Spike: A Demand Shock Like No Other
While supply is constrained, demand is going vertical. To understand the scale, we have to look at “power density.”
Traditional enterprise data centers operated at a power density of perhaps 5-10kW per rack. An AI training cluster equipped with Nvidia Blackwell GB200s operates at 50-100kW per rack—or higher.
A single gigawatt-scale data center—standard for the hyperscalers like Microsoft and Meta in 2026—consumes as much power as a city the size of San Francisco or a nuclear reactor. As we covered in our analysis of the AI energy crisis, when you drop a “nuclear reactor’s worth of load” onto a rural substation in Northern Virginia or Ohio, you don’t just need a new wire. You need an entirely new high-voltage substation, requiring multiple 500kV LPTs.
[!IMPORTANT] The Math of Shortage The U.S. grid needs to grow by roughly 2-3x its historical rate to accommodate AI. But transformer manufacturing capacity is only growing at ~5% per year. The math simply does not balance.
The “Phantom Queue” Effect regarding Orders
Complicating matters is the “Phantom Queue.” Just as toilet paper vanished in 2020 due to panic buying, developers are now panic-ordering transformers.
A data center developer might place deposits for transformers for five different potential sites, knowing only one will get built. This artificially inflates the backlog, making lead times look even worse (now quoted at 210+ weeks). Manufacturers, burned by cancellations in the past, are now demanding non-refundable deposits of up to 50% just to hold a slot in the “production queue.” This capital inefficiency drains money that could be used for actual construction, further slowing deployment.
A History of Underinvestment: The Regulatory Trap
This crisis didn’t appear overnight. It is the result of decades of efficiency-seeking. Utilities, regulated to keep costs low for consumers, ran “lean” inventories.
The “Regulatory Compact” means utilities only get a return on “used and useful” assets. A spare transformer sitting in a field earning nothing was seen as “waste” by regulators. So, utilities stopped keeping spares.
Now, the fleet is geriatric. The average age of a transformer in the U.S. is 40 years. Their design life is roughly 40 years. We were already due for a massive replacement cycle before ChatGPT triggered the AI arms race. Now, we have a perfect storm:
- Replacement Demand: Old units failing from age and extreme weather events.
- AI Demand: New massive loads coming online.
- Renewable Demand: Solar and wind farms need more transformers per megawatt than coal or gas plants because they are distributed. A 1GW coal plant needs one big step-up station. 1GW of solar might need hundreds of pad-mount transformers and dozens of collector substations.
The Lead Time Reality
In 2020, lead times for LPTs were 60-80 weeks. Today, they are 210+ weeks.
This creates a brutal reality for data center builders. You can buy the land, permit the building, and secure the GPUs, but if you can’t get the transformer to connect to the 345kV line, your $10 billion facility is just a very expensive warehouse.
Forward-Looking Analysis: Coping Mechanisms & Innovations
So, how does the industry survive the “Transformer Gap” of 2026-2028?
1. Standardization vs. Customization
Historically, every utility ordered custom-spec transformers. One wanted the bushings on the left, another on the right. This “snowflake” approach kills manufacturing efficiency. The industry is rapidly moving toward standard voltage classes and sizes, allowing manufacturers to churn out “generic” units faster. If a utility refuses the standard unit, they go to the back of the line.
2. The Rise of “Small Modular” Substations
If you can’t get one gigantic 100MVA transformer, maybe you can get ten 10MVA units. We are seeing a shift toward distributed substation designs that use smaller, more available medium-voltage transformers, networked together to handle the load. This is less efficient but available now.
3. Strategic Stockpiles and The DPA
The Department of Energy is finally moving to create a strategic reserve of critical grid components, treating transformers with the same national security importance as oil or medical supplies. There is talk of using the Defense Production Act (DPA) to force domestic steel production for GOES, but again, physics dictates the timeline—you cannot decree a steel mill into existence overnight.
4. Vertical Integration: The “Amazon Steel” Era?
Don’t be surprised if Amazon, Google, or Microsoft starts buying steel mills or transformer factories. Just as they designed their own silicon (TPUs, Maia, Trainium) to bypass chip bottlenecks, the hyperscalers may be forced to backstop the supply chain for the physical infrastructure that powers them. We are approaching an era where “Tech” companies are actually “Energy & Infrastructure” companies in disguise.
5. Solid State Transformers (SST)?
The “Holy Grail” is the Solid State Transformer—using high-voltage power electronics to step down voltage, effectively replacing the giant copper coils with silicon carbide (SiC) chips. While promising, SSTs are still expensive and unproven at the 345kV scale. They remain a 2030 solution for a 2025 problem.
Conclusion
The bottleneck for AI isn’t code or compute—it’s copper and steel. The winner of the AI race in 2026 won’t necessarily be the company with the best model, but the company that pre-ordered their transformers in 2023. Until the physical grid catches up to the digital dream, the “Intelligence Age” will remain constrained by the humble, 100-year-old technology of the transformer.
🦋 Discussion on Bluesky
Discuss on Bluesky