Yasho Singh
No. 1 · November 2026

Silicon, Power, Steel

Why the AI bottleneck keeps sinking deeper

By Yasho Singh
Illustration of a miniature AI data center seen through the center of a hand-wound copper transformer coil

The summer I was sixteen, I worked inside a transformer factory in a tier-3 city in India.

It was the most dangerous place I’ve ever stood. Men in flip-flops and undershirts casually guided multi-ton sheets of steel swinging from cranes while torches burned at 5,000°F a few feet away. In one corner, men wound copper around giant cylinders by hand. A guy near the back spray-painted a finished tank the size of a small house with no respirator, just a cotton scarf pulled over his face in the false hope that it would keep out the fumes. Over it all came a deafening roar: metal being beaten, machinery whizzing, cranes screaming overhead, a forklift threading through the middle of it all.

For over a decade, I buried that memory. As my career moved through Georgia Tech, autonomous vehicle supply chain at Nuro, and eventually VC investing at the frontier of tech, I felt a strange shame about where I’d started. I wanted to align myself with the clean, frictionless future, not a 19th-century industrial workshop.

I was wrong about what I’d seen. No one in that room thought we were building something that would become the substrate of artificial intelligence, and of the robots it will one day run. But the machines we were making, giant coils of steel and copper that nobody outside the industry ever thinks about, have become one of the hardest constraints on AI.

In 2022, I was at Nuro, scrambling for chips like everyone else in tech, convinced silicon was the bottleneck. Four years later, the bottleneck has sunk twice: first from chips to power, and then from power to the steel beneath it. What’s throttling AI today isn’t only the GPU. It’s the same kind of machine I watched being built in the flames and chaos of that factory floor when I was 16. Each time the bottleneck moves, it lands somewhere money can’t fix as quickly. This essay is about why.

Act I: Silicon

At Nuro in 2022, the workday started with a hunt.

My job was to track every PCB component and sensor that went into our autonomous vehicles, from integrated circuits that cost a few cents to GPUs and lidar sensors that cost tens of thousands of dollars. Every morning, we got on calls to track down individual microcontrollers: chips that cost fifty cents and were holding up assemblies worth a hundred thousand dollars.

Across the industry, lead times for basic components stretched from around 12 weeks to a year or more. Buyers panicked and did the rational thing, which made everything worse: they placed duplicate orders across multiple distributors and suppliers, hoping one would come through. Order books filled with phantom demand, and no one could tell how many chips the world actually needed.

The headlines were about cutting-edge chips: TSMC, Apple, the most advanced processes in the world. But the most destructive shortages were in old, cheap, unglamorous parts made on legacy processes, like the chips inside power supplies and car controllers. Foundries had little reason to expand capacity for products with thin margins. So the chips nobody thought about were the ones that stopped everything.

Eventually, the crisis eased. As post-pandemic double-ordering cleared and consumer tech demand cooled, artificial order backlogs evaporated. When generative AI took off with the ChatGPT launch in late 2022, the conversation shifted to a new chip, the GPU.

Then it happened again. By 2023, AI labs and startups were waiting the better part of a year for NVIDIA’s H100. Again, the bottleneck wasn’t the chip itself but packaging, the obscure step where a chip is bonded to its memory, done almost entirely by TSMC. But this time, the money moved fast. TSMC doubled its packaging capacity in 2024, doubled it again in 2025, and cut the time to build a new packaging plant from several years to under two. By 2024, H100 wait times had dropped to a couple of months, and some companies that had hoarded GPUs were quietly reselling them.

For a while, it looked like silicon would be AI’s bottleneck forever. It wasn’t. Chips still pinch at times, memory especially, but they’re no longer the deepest constraint.

The lesson from both the 2022 and 2023 chip crises stuck with me: the bottleneck rarely sits where everyone is looking. It hides in the boring, low-margin layer the industry stopped investing in. And today, that layer is buried much deeper than any chip.

Act II: Power

If you want to understand the new bottleneck, start with a mismatch in clocks.

A hyperscale data center can go from empty field to finished building in under three years. Connecting it to the grid can take five to seven. Utilities plan in decades. AI companies plan in quarters. For nearly two decades, that didn’t matter, because electricity demand in the US barely grew. Then AI arrived and asked for more power, faster, than the grid was ever designed to deliver.

The result is a line. To plug a large new facility into the grid, a developer files an interconnection request and waits for the utility to study whether the grid can handle it. Those queues have exploded. In Texas alone, the large-load queue at ERCOT, the state’s main grid operator, grew from 63 gigawatts at the end of 2024 to over 470 gigawatts, most of it data centers.

When I first saw those numbers, I realized I’d seen this movie before.

In 2022, panicked buyers placed duplicate chip orders across every distributor they could find, and order books filled with demand that didn’t exist. Today, data center developers do the same thing with power. They file requests at multiple utilities, for multiple sites, for the same project, because filing is nearly free and nobody knows which utility can deliver first. Wood Mackenzie estimates that of the more than 1,000 gigawatts of data center power requests on file, utilities are likely to commit to serving only about 28%. Some of the rest is real demand that won’t be served; much of it is duplicate filing. Texas is now trying to break the cycle by making speculative requests expensive. Under ERCOT’s new “Batch Zero” process, applicants must post $50,000 per megawatt in financial security, and only about 205 gigawatts, less than half the queue, was eligible for the first round.

Phantom demand hides a real shortage rather than disproving one. Even after you strip out the speculation, the grid can’t keep up. And like in 2022, the industry’s response tells you how serious it is.

Developers stopped waiting. They started building their own power plants.

“Behind the meter” generation, power produced on-site rather than drawn from the grid, has gone from a niche backup strategy to a mainstream plan. Data center campuses are installing natural gas turbines, reciprocating engines, and fuel cells. One of the companies supplying them is Cummins, where I interned years ago working on engines. Back then, nobody in tech paid Cummins much attention; it made diesel engines for trucks and generators. Now it’s supplying gas-fired generation for a microgrid that will power an AI campus in West Texas.

But even the workaround has a queue. The three companies that dominate large gas turbines, GE Vernova, Siemens Energy, and Mitsubishi, are effectively sold out for years. GE Vernova is now selling delivery slots for 2031. Mitsubishi has said deliveries for new orders are stretching to 2031 as well. Kentucky Utilities paid GE Vernova $25M just to hold a single reservation slot in line, before a single turbine is built. Developers have become so desperate that some are powering data centers with mobile generators on semitrucks and turbines originally designed for jet aircraft.

The chip-packaging crunch eased in about two years because chips are high-margin and capital moves fast. Power is different. The grid is regulated, shared by everyone, and built to last fifty years. You can’t double it in a year the way TSMC doubled its packaging capacity.

And underneath both the grid and the workaround sits the same machine.

Act III: Steel

All the power that reaches a GPU passes through a transformer, usually several. A transformer steps voltage up so power can travel across the grid, then steps it down so a building can use it. A data center connected to the grid needs them. A data center running its own gas turbines needs them too. There is no workaround for the workaround.

HOVER OR TAP A PATH TRANSFORMERS EVERY PATH PASSES HERE PATH 1 · GRID POWER Power plant Transmission grid GAS, NUCLEAR, WIND HIGH VOLTAGE STEP UP STEP DOWN PATH 2 · ON-SITE POWER On-site generation TURBINES, ENGINES STEP DOWN AI data center GPU RACKS POWERED
FIG. 1 — Two ways to power a data center, one shared dependency. Grid power is stepped up for long-distance transmission and back down for use; on-site generation still passes through its own transformers before it reaches the racks.

And transformers are in shorter supply than almost anything else on the grid. Before 2021, a large power transformer took about a year to arrive. Today, it takes two to four. Prices are up nearly 80% since 2019. And four of every five large power transformers the US uses are imported.

The obvious question is the one tech always asks: why not just build more factories? Because a transformer is three shortages assembled into one machine: a special steel, a mountain of copper, and a set of human skills that take years to learn. When I say “steel,” I mean all of it, the layer of the economy that is dug, smelted, rolled, and wound.

The first is the steel itself. A transformer’s core is made from grain-oriented electrical steel, rolled and heat-treated so its magnetic properties line up in one direction. The United States has exactly one producer: Cleveland-Cliffs, at a legacy mill in Butler, Pennsylvania, which can’t roll the widths or efficiency grades many modern large transformers need. A few years ago, the government worried the company might stop making it altogether, because the business barely made money. Now Cleveland-Cliffs is spending $200 million, $75 million of it from the Department of Energy, to raise the mill’s output by up to 25% by early 2028. It’s a real improvement, and nowhere near enough. One mill, upgraded over several years, can’t close a nationwide gap.

The second is copper: miles of insulated wire at the heart of every large power transformer, plus the busbars, cables, and grid connections around it. Data centers alone use more than three times as much copper as they did in 2020, and the grid needed to power them uses even more. If steel is slow to fix, copper is slower. A new copper mine takes, on average, around 20 years to go from discovery to production; in the United States, closer to 30. The miners aren’t short on money, earning more than $100 billion a year in profit, but they spend relatively little on new mines, because a mine is a bet that demand will still be there two decades from now. The copper in a transformer built today comes from mines someone decided to build a generation ago.

The third is people. Inside Hitachi Energy’s factory in South Boston, Virginia, workers called winders coil that copper by hand around giant cylinders to form the heart of each transformer. One unit takes four to six weeks. The skill takes three to five years to master and can’t be fully automated. Hitachi’s new $457 million plant next door will need about 800 more workers, in a county of roughly 34,000 people.

When I read that, I thought of the men in that corner of the factory in India, winding copper by hand. I had assumed that work belonged to the past. But in Virginia in 2026, the heart of every large transformer is still wound by human hands. That work isn’t the past. It’s the critical path of the future. And it’s exactly the kind of work no robot can yet do.

It’s also, increasingly, where America is buying from. The United States is now the largest destination for India’s transformer exports. This fall, a manufacturer in Prayagraj, India, shipped its first extra-high-voltage transformer to the US. The world forgot this industry was strategic, and the places that never stopped building it are now filling the gap.

New capacity is coming. Hitachi is spending more than $1.5 billion on US transformer plants, but its biggest new factory, in Mississippi, won’t start production until 2029. Compare that with TSMC, which cut the time to build a new chip packaging plant to under two years.

That’s the difference between silicon and steel. When chips ran short, money flooded in and factories appeared within two years. Here, money isn’t enough. You can’t wire capital into metallurgy, rush a copper mine that takes 20 years to open, or download a winder’s five years of training. AI now waits on the layer of the economy that moves at the speed of mills, mines, and human hands.

The Pattern

Line up the three crises side by side, and a pattern appears. Each time the bottleneck moved, it sank to something slower to fix.

Chip packaging plant Transformer factory Trained transformer winder Gas turbine, ordered today Grid connection New copper mine UNDER 2 YRS ~3 YRS 3–5 YRS ~5 YRS (2031) 5–7 YRS ~20 YRS GLOBALLY ~30 US 051015202530 YRS
FIG. 2 — Years to add capacity at each layer. Lighter segments show the upper end of the range. Sources below.
  • A new chip packaging plant: under two years.
  • A new transformer factory: about three years from groundbreaking to first production.
  • A trained transformer winder: three to five years, one person at a time.
  • A gas turbine ordered today: delivery around 2031, about five years out.
  • A grid connection: five to seven years.
  • A new copper mine: about 20 years globally, closer to 30 in the US.

Each line measures the time to add new capacity: build a plant, train a worker, take delivery, connect, or open a mine. Expanding what already exists is faster, but only at the margin.

These layers don’t take turns. Packaging, memory, interconnection, turbines, and transformers can all bind at once, depending on the project and the place. What sinks is the deepest constraint: AI keeps getting tied to industries whose supply can’t respond on a short horizon.

Deeper doesn’t only mean slower. It means what everything above depends on: no grid connection or gas turbine delivers power without transformers, and no large power transformer gets built without electrical steel, copper windings, and the people who wind them.

AI is the fastest-moving industry in the world, and it now runs at the speed of its slowest one.

The obvious explanation is margins. That’s what happened with legacy chips in 2022: foundries didn’t expand capacity for parts that barely made money, so the cheap parts were the ones that ran out. Transformer steel followed the same script. For years, it barely made money, and America let its capacity wither to a single mill.

But margins don’t explain copper. Copper miners are making more than $100 billion a year in profit, and they’re still not building many new mines.

The deeper explanation is time. Time does two things at once. Physically, it makes supply inflexible: a mine or a factory can’t be rushed. Financially, it makes the bet uncertain: the longer the wait, the less anyone knows whether the demand will still be there. The first makes shortages last; the second discourages anyone from fixing them. Every layer down asks investors to be right for longer. A packaging plant pays back within a few years of a demand surge, so when chips ran short, money moved instantly. A transformer factory is a bet on demand a decade out. A copper mine is a bet on the world two decades from now. Capital is happy to chase a shortage; it’s much less happy to underwrite one that has to last twenty years to pay off. So the deeper the layer, the slower money arrives, and the longer the shortage lasts.

That’s why the bottleneck keeps sinking. Every time capital fixes one layer, demand hits the next one down, where the fix takes longer and the bet is harder to make. AI is the fastest-moving industry in the world, and it now runs at the speed of its slowest one.

What It Means

If the real bottleneck is time, then a few things follow.

Time is the moat.

In 2023, the winners were whoever got GPUs first. Today, they’re whoever made the long bets years ago: the utilities and hyperscalers that reserved turbines and transformers before anyone else saw the shortage, and the equipment makers that kept building through the lean years. GE Vernova says it is mostly sold out through 2030. A turbine slot, a signed grid connection, or a site with power already secured is now worth more than almost anything else an AI company can buy, because no amount of money can make it arrive sooner. A power contract may now be worth more than a GPU order.

The opportunity is shortening the bet.

If the real bottleneck is how long investors have to be right, the opportunity is to shorten the bet: turn long, uncertain industrial commitments into shorter, more predictable ones. There are three ways to do it, from easiest to hardest.

The first is to skip the new bet entirely by getting more out of capacity that already exists. The grid is built for its worst few hours of the year, which means it sits partly idle the rest of the time. A 2025 Duke University study found that if new data centers cut their power use for just 0.25% of the year, averaging about two hours at a time, the existing US grid could absorb about 76 GW of new load without a single new power plant. It won’t work everywhere, but it’s the rare fix that’s mostly software, aimed at a problem that otherwise takes five to seven years. The catch: it eases the grid connection, not the deepest layer. A flexible data center still needs its own transformers.

The second is to make the bet smaller. Modular on-site power, like the gas generators Cummins is supplying in West Texas, skips the five-to-seven-year wait for a grid connection, though it still needs its own transformers. Equipment designed to use less electrical steel and copper shrinks the bet at the deepest layer. Smaller units turn one decades-long commitment into many short ones that can stop if demand does.

The third is the hardest, and the only one that goes straight at the deepest layer: making the long bet certain enough that someone will place it. That is already happening. Buyers now pay to reserve equipment years ahead, like the $25 million Kentucky Utilities paid GE Vernova for a single turbine slot. A reservation moves part of the risk from the manufacturer to the buyer, and certain demand is what lets a factory expand. Long-term purchase commitments work the same way: they turn a decade-long gamble into a contract.

Procurement is where I’d start, with a caveat my own history taught me. In 2022, I watched companies win or lose on whether they’d secured fifty-cent chips in time, and I watched everyone’s rush to reserve parts make the shortage worse. Winning scarce capacity for one company doesn’t create any more of it. The bigger prize is helping manufacturers tell credible long-term demand from phantom demand, so they can build with confidence. Credible demand looks like what Texas is trying to force with Batch Zero: money posted up front, not free filings at every utility. Done right, procurement stops being a way to win the shortage and becomes a way to end it.

Robots are on both sides of the bottleneck.

People are the hardest layer to scale with money alone. An electrician’s apprenticeship takes four to five years. A transformer winder takes three to five years to master the craft. Software can’t compress that. Physical AI is one of the only technologies that could.

It’s already starting. On one data center site, four drilling robots from DEWALT and August Robotics completed a floor-drilling job that takes crews eight to nine weeks in about seven to nine days. And the work of the winders in Virginia, and of the men in that corner of the factory in India, is exactly what physical AI is trying to learn: precise, variable, and physical. Whoever teaches a robot to wind a large power transformer won’t just build a better robot. They’ll loosen a constraint on the entire AI economy. The two aren’t on the same clock, though: drilling robots are relieving shortages now, while a robot that can wind a large power transformer is still years away.

But robots are also competitors for the same slow inputs. Every joint in a humanoid runs on a motor made of copper windings and rare-earth magnets, about 1.3 kilograms of rare earths per robot by Morgan Stanley’s estimate, roughly double what an EV motor uses. In 2025, China’s export controls on those magnets were enough to slow Tesla’s Optimus. The faster physical AI scales, the harder it presses on the very layers it could help fix.

The scarcest skill is understanding both clocks.

Silicon Valley runs on one clock: ship in weeks, scale in quarters. The industries that now decide how fast AI can grow run on another: plan in years, build in decades. Most people have only ever worked on one. The companies that win the next decade will be built by people who can work on both, moving at software speed where they can, and planning on steel time where they must.

The summer I was sixteen, I thought I was standing in the past.

Men in flip-flops guided steel through the air. Torches burned a few feet away. In one corner, men wound copper by hand around giant cylinders while the whole floor roared around them. I was sure the future would leave all of it behind.

So I spent the next decade moving toward the other clock: Georgia Tech, Nuro, venture capital, the clean, frictionless side of the economy, where problems get solved in a sprint. And from there, I watched AI run straight back into that room. The GPUs wait on power. The power waits on transformers. The transformers wait on steel, copper, and people who wind it by hand.

I used to be ashamed of where I started. Now I think it’s the most useful place I’ve ever stood.

Sources

Act I: Silicon

Act II: Power

Act III: Steel

What It Means