Can the People Building the AI Boom Share in it?

I’m based in the US, so it is almost debilitating to keep up with every headline… let alone start dipping internationally. So when someone made a throwaway comment this week mentioning the stock implications of the Samsung strike, I almost let it go.

On the surface, it reads like one of the ever-growing number of labor disputes we’ve learned to file away — factory workers, a pay dispute, all a world away in South Korea. Easy to scroll past. But sitting underneath is a story that actually captured my dwindling attention span. It’s the story about the actual physical infrastructure of AI — where it comes from, who builds it, and what it costs those people. It’s also a story about a shift happening in how technical talent moves, who gets paid, and what happens when the people doing the most physically demanding work in tech finally decide to demand a different arrangement.

For that reason, this week I decided to pull you down the rabbit hole with me and go deep into what I discovered. Because beyond being interesting, I think anyone building on AI today should understand what’s underneath the hardware they’re waiting on — and why this particular strike, in this particular moment, is more connected to your workflow than it might look.

What’s happening with the Samsung Strike and why?

Let’s start with the dispute itself, because the details matter much more than the headline.

Samsung’s first and largest labor union (Samsung Electronics Union), representing more than 90,000 employees, spent the last month in an escalating conflict with their management over one thing: a proportional share of the profits their work generates.

To understand why that feels beyond reasonable, you need to follow Samsung’s recent journey as a company. Like most tech / device companies, Samsung rode the COVID consumer boom hard. Everyone is stuck at home, needing to connect to the world through their devices. Those devices need memory. (Samsung $$$) Then, post-COVID, everyone had already bought their laptop. Demand collapsed, memory prices cratered, and Samsung posted operating losses for much of the past two years.

In that time, workers received zero bonuses. Their pay structure runs through something called the EVA formula — described by the union as a black box, with two known features: it deducts capital costs and losses before calculating the bonus pool, which means in a down year, the floor is zero. And the ceiling, in any year, is capped at 50% of base salary.

Then AI arrived and flipped everything. Data centers started consuming memory at rates the industry, once run by human consumer demand, could never have predicted. Samsung’s profits went from operating losses to record returns in a matter of quarters. They quickly went from a surplus to a shortage of product, and preorders scheduled for months to a year in advance were in high demand.

However, this rebound was not seen by Samsung employees. They saw zero in the bad year. And nothing again in the record year… because the black box formula, with its deductions and cap, was stacked against them. This obviously sucks, but it reads like any company story today. Nobody likes to pay.

It’s obviously not fair, but what can any employee do to change that? Unfortunately for Samsung, that’s not an argument their employees have to have in the abstract. The answer is to just go down the street.

SK Hynix — Samsung’s Korean rival, same industry, same job, same country — restructured its compensation last year. 10% of annual operating profit is allocated directly to a worker bonus pool. No deductions. No cap. Written into their contracts for ten years. The formula is public, letting all employees do the math themselves. Based on 2026 projections, that works out to an average bonus of about $460,000 per worker this year, up from $95,000 in years prior.

This has obviously caused an exodus at Samsung. The engineer doing the same work, in the same country, for the rival company: $460,000. If they stayed at Samsung: nothing.

This comparison is what brought tensions to a head between Samsung and the union. And it raises a real question: in a corporate landscape that fights this hard against paying people, what is so important about what they’re building that workers — and rival corporations — are starting to fight over them?

Why are these workers a bottleneck?

To understand really why a labor dispute in Korea connects to your estimated 40-week GPU wait-time, you need to understand what actually comes off that factory floor.

A modern AI GPU — the H100, the Blackwell, whatever comes next — is not one chip. It’s an assembly. At its center sits the GPU die: the compute engine, the brain, designed by Nvidia or AMD and predominantly manufactured by TSMC in Taiwan. Their main competitor is Samsung Foundry, which has been struggling to keep pace — though that’s a different part of Samsung than the workers we’re talking about today.

The AI chip supply chain has three layers. Designers like Nvidia create the blueprints. Manufacturers like TSMC build the compute die. Memory suppliers — led by Samsung — build the HBM that makes the whole thing run. Samsung sits at the intersection of two of those three layers, which is why this labor dispute matters.

The brain handles the math. But math requires moving data, constantly and fast. So, surrounding that die is a stack of High Bandwidth Memory (HBM). Think of it as the engine’s fuel system. Without it, a Blackwell GPU is a very expensive piece of nothing. HBM also requires significantly more wafer capacity to produce than conventional memory, so the more Samsung shifts production toward HBM4 to meet AI demand, the tighter overall memory supply gets. More demand, more production, less memory for everything else.

A modern AI GPU is an assembly, not a single chip. The HBM stacks on either side are produced by Samsung’s workers. Without them, the compute die in the center does nothing.

Those wafers are what the Samsung workers we’re discussing today build in specialized cleanrooms. (~40% of the world’s DRAM inventory is built by them, to be exact).

What makes cleanroom workers irreplaceable?

Beyond being likely great people, they know how to operate in a data cleanroom.

A semiconductor cleanroom is nothing like a typical factory floor. For context, regular room air has about 35 million particles per cubic meter — dust, skin cells, fibers. In an ISO Class 1 cleanroom, where Samsung’s HBM wafers are made, that number drops to just 10. Even a single stray hair can ruin millions of dollars’ worth of work. Before each shift, workers spend up to 30 minutes getting suited up: full-body coverings, respirators, triple gloves, sealed boots. Once inside, they’re in for twelve-hour shifts day and night in dry, filtered air under yellow lighting, with hardly any chance to eat, drink, or step out.

A researcher in a Samsung Electronics chip cleanroom (https://www.kedglobal.com/korean-chipmakers/newsView/ked202503060008)

And what they do inside those suits is far from mundane or simple assembly work. They operate chemical vapor deposition systems, plasma etching machines, and extreme ultraviolet lithography equipment that costs millions of dollars per unit. And while many may start with a two-year technical degree — the Korean equivalent of an Associate’s or Trade Specialization — the ability to do the job, manage chemicals, use machines, and interpret data takes years to master.

In many ways, clean room workers are modern-day trade specialists — a high-tech evolution of the blacksmith apprenticeships of yore. The difference between what a seasoned technician has learned on the job and a new hire is the human judgment that determines whether a $2 million wafer batch survives the night shift.

To top it all off, these technicians are currently working under more pressure than ever before. To keep up with AI demand, Samsung has accelerated its factory expansion in Pyeongtaek by months, with construction and equipment calibration occurring simultaneously on the same floors. Cleanroom staff are keeping everything sterile while construction crews work just next door. The yield requirements for HBM4 are extremely strict, and the factories are running at full capacity around the clock just to meet delivery targets.

Unsafe Conditions in Clean Rooms are the Reason Korea has Unions at All

For most of Samsung’s history, there was no union — and that was true across Korea’s chaebol system broadly. In South Korea, your company is closer to an identity than a job. That cultural weight discouraged independent organizing. Samsung went further and actively suppressed it. It took a 2019 court ruling to change that. But the reason clean room workers specifically wanted a union in the first place goes back further and is much more permanent than any pay dispute.

In 2003, a 19-year-old named Hwang Yu-mi took a job at Samsung’s Giheung chip plant, dipping silicon wafers into chemical baths. By 2005, she was sick. In 2007, she died of acute leukemia in the back of her father’s taxi on the way to the hospital. She was 23. Her father, Hwang Sang-ki — a taxi driver — refused to sign Samsung’s non-disclosure agreement. He spent the next eleven years documenting other cases. The group he helped found, SHARPS, eventually identified over 200 Samsung workers who had developed leukemia, lymphoma, brain tumors, and miscarriages — most in their 20s and 30s. Seventy-six had died. In 2018, Samsung’s co-president bowed at a press conference and issued a formal apology. Two years later, Jay Y. Lee, Samsung’s Chairman, renounced the no-union policy. Samsung Electronics Union — the company’s first recognized union — was formed.

This is the same union that is now negotiating for a share in the AI boom for Samsung workers. The direct descendant of what Hwang Sang-ki started in the back of that taxi.

… And Back to The Politics of Pay

At the beginning of this article, we introduced the challenge posed by the union and the standard set by SK Hynix, which made it impossible to ignore. But zoom out from Korea, and a pattern emerges across every company building this hardware. Compensation has become a strategic weapon.

TSMC has used aggressive base salary increases to fend off poaching from Chinese firms. Micron’s wage structure was built into its CHIPS Act deal from the start, a condition of $6.1 billion in government funding. China is deploying unlimited state capital and equity packages to recruit engineers it can’t build domestically fast enough. Every company in this ecosystem has made a calculated decision about what these workers are worth — and priced accordingly. Samsung was the last holdout.

A look at different companies approaches to competing for cleanroom workers

Not one of those changes happened because a company decided it was the right thing to do. Every dollar these workers have gained came from a fight — or from a rival making the fight inevitable.

So can the people physically building the AI boom have a piece of it? It seems that at least for now — yes. The Samsung team has struck a deal with the union in line with the SK Hynix methodology. Not because the industry decided it was fair, but because, in this day and age of GPU waitlists, the most irreplaceable jobs in tech are among the least glamorized and physically demanding. The companies that figured that out first are winning the talent war. The ones that didn’t had to be dragged there.

So somewhere in Pyeongtaek, on the third shift, someone is tending a wafer that will eventually end up in a data center that will eventually run a model that will eventually generate revenue for a company that will never know their name. And the good news is that this week, that person got a raise.