Why America Has to Make Its Own Sunlight

I had a crane certification before I had a business card.

That was the start of it. Most of my first decade went to mega-projects — the UK, Africa, Australia, and home in the States — and work at that scale teaches you a particular kind of alertness. Being wrong on a site is loud. It happens in about a second, everybody sees it, and you fix it before lunch or you don’t work there anymore.

Factories are the opposite, and it took me longer than I’d like to admit to adjust. In a factory, being wrong is quiet. A solder profile drifts two degrees on a Tuesday and nothing happens. Nothing keeps happening for six years. Then a string of modules on a roof in Arizona starts giving back half a percent a year more than it should, and by the time anybody traces it to that Tuesday, you’ve built four hundred thousand of them.

I’ve put somewhere around 900 megawatts of other people’s modules on the ground since then. Projects taught me what a loud mistake costs. The factory taught me to be afraid of the quiet ones. Both are why I’m writing this.

What matters is the machine that converts it. A photovoltaic module is a laminated sandwich of glass, polymer, silver paste, and refined silicon that has to survive thirty years of hail, humidity, salt, and thermal cycling while giving up a fraction of a percent of its output a year. Cells get soldered into strings. Strings get laid up between glass and encapsulant. The stack goes into a laminator, comes out cured, gets framed, gets a junction box, and gets flashed under a solar simulator to prove what it’ll do on a roof in year one and, by inference, in year twenty-five.

None of that comes from the sky. It’s metallurgy, process control, adhesion chemistry, and yield.

America has largely stopped doing it. We didn’t stop installing solar — we install more every year, and I’ve been on the buying end of that for most of my working life. We stopped making it. The distinction sounds academic right up until you ask what happens when the people who do make it decide, for reasons of their own, to make less of it for us.

Five stages: polysilicon, ingot, wafer, cell, module. The IEA puts China’s share of every one of them above eighty percent, and its share of ingots and wafers at roughly ninety-five percent or better.

Ninety-five percent of anything, anywhere, is a single point of failure. This one has a flag on it.

The wafer is the part that should keep people up at night, and almost nobody thinks about it, because nobody puts a wafer on a rooftop. A wafer plant is capital-heavy, energy-hungry, and slow to stand up — years, not quarters. Module assembly is the opposite: the easiest stage to bring home and the loudest to announce. We’ve gotten reasonably good at the loud part.

I want to be careful here, because there’s a comfortable version of this story and an accurate one. China didn’t steal this industry. China won it — with sustained capital at a scale Western investors wouldn’t tolerate, vertical integration from quartz to module, utilization rates our plants never held, and an industrial policy that treated a manufacturing base as an asset instead of a cost center. I bought from those companies for fifteen years. The product was good, it was cheap, and it showed up when they said it would. Anyone who tells you otherwise is selling you a grievance, and a grievance won’t produce a single wafer.

The weakest version of the reshoring case is nostalgic: American hands should build American things. I run a manufacturing company and I don’t find that persuasive, because it doesn’t survive contact with a spreadsheet.

The strong version is about coercion. Concentrate a supply chain inside one jurisdiction and you’ve handed that jurisdiction’s government a lever. Whether they ever pull it is their call, not ours. That’s the whole problem.

We’ve watched the lever get used. In 2010, rare earth shipments to Japan slowed during a territorial dispute. In 2023, export controls landed on gallium and germanium, then graphite, then antimony. Each time, the affected industries found out the same thing about the input they’d been buying for years.

Solar crossed the line from consumer good to infrastructure sometime in the last few years, which is what makes this urgent instead of interesting. Solar and storage are the overwhelming majority of the generating capacity sitting in American interconnection queues. They’re the marginal new electron on this grid. And electricity isn’t just what runs the lights anymore — a federal-laboratory study projects that data centers alone could be pulling something on the order of seven to twelve percent of U.S. electricity by 2028.

Follow that to the end. Compute runs on electricity. Increasingly the marginal electricity comes from photovoltaics. And the photovoltaics come from a supply chain we don’t control. We’re building the most consequential technology of the century on a power platform whose manufacturing base sits with a strategic competitor. You don’t have to believe that competitor will ever act against us. You only have to notice whose option it is.

I don’t have to imagine what any of this does to an island. Three hundred people work for us in Puerto Rico. The storm everybody remembers was 2017, and people talk about it in the past tense, but the aftermath never really ended here — reliable power and reliable water are still open questions for a lot of households, nine years on. That isn’t history I’m invoking for effect. It’s the operating environment.

And when a grid goes down, what holds up the rebuild isn’t engineering talent and it isn’t money. It’s whether the hardware is sitting on the dock or eight weeks out on a boat, from a port that can close for a policy reason or a weather reason or because it’s simply backed up.

Now the objection, which deserves a real answer instead of a slogan. Chinese modules land at roughly a dime a watt. American-made modules cost two to three times that. Anyone who tells you patriotism closes that gap has never carried a P&L, and I’ve carried a few.

I bought Chinese modules for years, and with the same information I’d do it again. For most of the last fifteen years, buying anything else would have been malpractice against my own numbers. I don’t get to pretend otherwise now that I’m on the other side of the transaction.

Direct labor is on the order of a tenth of module cost. The rest is materials, capital equipment, energy, logistics, freight — and, dominating all of it, yield and utilization. A line at ninety-eight percent yield and ninety percent uptime and a line at ninety-two percent yield and sixty percent uptime are different businesses making the same product.

The first time an engineer walked me through what two points of yield does to cost per watt, I asked him to run it again on the whiteboard. I’d spent a decade thinking about this as a wage problem. It mostly isn’t one. The cost gap is a scale-and-yield gap wearing a wage costume, and that changes what kind of factory you should build — a tariff protects the wage line, which is the tenth that barely matters.

The last American solar manufacturing wave gets remembered as a story about subsidies. It was really a story about cost structure.

Those companies built plants whose economics needed one of two things to hold: an exotic technology had to win on the merits, or the price of a conventional Chinese module had to sit still. Neither happened. Prices fell roughly eighty percent in four years. They’d bet their competitor would stand still, and nobody stands still.

I watched a few of those companies from the other side of the table, being pitched a module that cost more today and promised to cost less later. The pitch was always technically interesting and it never penciled. The lesson isn’t “don’t build in America.” It’s narrower and more useful: don’t build a labor-substituting plant to fight a labor-cost war that’s already over. Build the plant that wins on the axis where the fight actually gets decided.

If labor is a tenth of the cost and yield is the whole game, an American line should be organized around the variables that respond to engineering instead of wages: yield, throughput, uptime, scrap.

On a floor, that’s four things.

See the defect where it’s still cheap. A microcrack caught at the stringer costs one cell. The same crack caught at flash test — after lamination, framing, and a junction box — costs a finished module plus every minute of line time that went into it. Electroluminescence imaging and machine vision belong upstream, at the station where the damage happens, not at the end where it’s merely confirmed. I’d rather throw away a cell than explain a module.

Close the loop on the processes that drift. Solder profiles drift. Encapsulant cure drifts. Tab alignment drifts. Every one of those shows up in data long before it shows up in a failed module. Correcting a drifting process automatically is a completely different discipline from inspecting finished goods and throwing some of them away, and most module plants in the world still do the second one.

Give every module a birth certificate. Serialized, permanent traceability: which materials, which machines, which recipe, which shift. When a field failure surfaces in year six, a company that can name every module built from that lot has a contained warranty event. A company that can’t has something that ends companies.

Make the loop fast. A defect signal that arrives in seconds changes what an operator does next. One that arrives at the end of a shift only changes what gets scrapped.

None of this is exotic. It’s the discipline semiconductor and aerospace manufacturing have used for decades, pointed at a product that’s historically been built like a commodity. What’s new isn’t the algorithms — it’s deciding a photovoltaic module is worth that much rigor.

We’re a manufacturer that uses AI. We’re not an AI company. I say that a lot, and it isn’t positioning, it’s a budget. An AI company spends its money on models. We spend ours on the line, and use models to keep it in tolerance at three in the morning in the second week of a ramp, which is exactly when a human being stops noticing that something has been drifting for four hours.

But a runway only helps if you take off before it ends. There’s a version of this build-out that produces tariff-shelter factories — plants that pencil only inside the incentive, that never drive yield because they never had to, and that go dark the month the credit steps down. That would be worse than not building at all, because it would discredit the whole project for another decade and hand the argument to the people who think we shouldn’t try.

So the test we apply to anything we build is one question: does this line pencil at world prices in 2032, with no credit? If the answer’s no, we’ve rented a supply chain and called it reshoring.

A module plant that imports its cells is an assembly operation with a domestic address. Ours imports cells today. It’s worth doing — it builds the workforce, the quality systems, the logistics muscle, the muscle memory of running a line at all — but nobody should tell you it takes a hand off the valve, and that includes me.

The real work is upstream and sideways: cells, wafers, ingots, polysilicon. Alongside it, all the parts of the bill of materials nobody puts in a press release — solar glass, encapsulant, backsheet, silver paste, aluminum frames.

And underneath all of it, the layer almost nobody counts: the tool base. The stringers, the laminators, the diffusion furnaces, the metrology. Who builds the next generation of that equipment? I genuinely don’t know. Every answer I’ve heard, including the ones I’ve given, ends with buying the tools from the same place we’re trying to stop buying the modules from. That capability took a generation to lose. I’ve got no clean plan for getting it back, nobody I’ve asked has one either, and I’ve come around to thinking it may be the part of this problem that matters most.

What I do know is that you don’t get any of it at once. You get it the way a manufacturer gets anything — start where you can actually run, earn the right to go upstream by proving you can hold quality downstream, compound.

Automation-first gets heard as “fewer Americans employed,” and it’s backwards.

A line built this way needs technicians, controls engineers, process engineers, quality engineers, and people who can read production data and act on it before the shift ends. Those are careers, and they have a property that matters strategically: they can’t be arbitraged away by cheaper hands somewhere else, because they’re not competing on hands.

They’re also portable. A workforce that can hold tolerances on a photovoltaic line can hold tolerances on whatever this country decides it needs to build in a hurry next.

I came up through the trades, so I’ll say the part that people in my seat usually skip. The industrial base isn’t buildings. I’ve stood in beautiful empty buildings. It’s the people who know how to make a building produce something, and that knowledge lives in individual heads, gets transferred by standing next to someone, and evaporates in about a decade if nobody’s hiring.

We build in Aguadilla, Puerto Rico — three hundred people and a gigawatt of module capacity — and the reasons aren’t sentimental. U.S. jurisdiction, U.S. law, U.S. workforce protections, domestic content on day one. Deepwater port access. A real precision-manufacturing heritage in pharmaceuticals and aerospace, which means the résumés that come across my desk already have clean-process discipline on them — people who’ve held tolerances for a living and don’t need to be sold on why it matters. A bilingual engineering workforce.

And a grid whose fragility makes the customer’s problem impossible to abstract away. If you’re building energy hardware, that’s about the most useful thing an environment can give you: you can’t spreadsheet your way past a problem you drive through every morning.

That’s a longer argument, and probably its own essay. What matters here is that the case for American solar manufacturing doesn’t have to be made in the abstract. It can be made in a building, with a line running in it.

A country that can’t build the machine doesn’t own its energy — it rents it, on terms renegotiated annually by people whose interests aren’t its own. That was tolerable when solar was a niche. It’s harder to defend now that solar is the marginal source of power for the grid that will run this country’s compute, its industry, and its recovery from the next storm.

So: build the plant that’s automated, instrumented, and honest about having to win on cost eventually, without help. That’s what we’re trying to do in Aguadilla. The line is running and three hundred people run it. Whether it’s still running at world prices in 2032, with no credit propping it up, is the only version of this argument that counts, and I don’t get to grade my own paper.

Ask me then.