On May 27, Valve raised the price of the Steam Deck OLED to as much as $949 — a handheld that cost $399 when it launched in 2022 — citing "rising memory and storage costs." The same day, Samsung's largest union approved a pay deal handing its chip workers an average bonus of about $340,000, explicitly framed as their share of the AI boom's profits. Two facts that shouldn't sit on the same calendar day. The buyer of memory is being priced out of a gaming console. The maker of memory is collecting the biggest payout in the history of the industry. Same component. Opposite ends.
The reversal hidden inside those two stories is bigger than either. For thirty years, memory was the part of a computer engineered to disappear — abundant, interchangeable, and cheaper every cycle. Making it was the worst job in semiconductors. Today it is the most valuable thing the industry produces, and the people who make it are being paid like they discovered oil. The Wall Street Journal's headline, four days later, used exactly that word: AI has made memory chips more valuable than oil.
The Commodity
Memory was designed to be a commodity, and for a generation it behaved like one. DRAM and NAND were the chips you didn't think about — the gigabytes on the spec sheet, the cheapest part of the bill of materials, the thing that got better and cheaper on a schedule so reliable it underwrote the entire economics of personal computing. A laptop got more memory every year for the same price because that was simply what memory did. It fell.
The falling came in a specific shape: boom and bust, the most violent cycle in all of semiconductors. When demand softened, prices didn't drift down — they collapsed. In the fourth quarter of 2022, DRAM's average price fell 34% in a single quarter, among the worst drops since 2006. By early 2023, the sector was facing "one of its worst-ever routs." Bloomberg's framing captured the recurring delusion of the business: "This time was supposed to be different." It never was.
And the people who made memory paid for the cycle with their jobs. In February 2023, SK Hynix reported a $1.4 billion quarterly operating loss — its biggest on record. Two months later, Samsung's operating profit fell more than 96% to a 14-year low, and the company announced it would cut memory production "to a meaningful level." Micron cut headcount. This was the pattern memory workers lived inside: when the cycle turned, the memory division was the first thing the company sacrificed. The component that was always getting cheaper was made by the workers who were always first in line for the cuts.
The Inversion
Then a customer arrived that did not care about the cycle.
AI infrastructure does not buy memory the way a PC maker buys memory. A data center's appetite is contractual, multi-year, and effectively unbounded. Every model needs more of the physical chips that hold data while a GPU works. The high-bandwidth variety that matters most is made on the same lines, by the same three companies, that make the chips in everything else. When that demand hit, the cycle did not soften. It inverted. By late 2025, nearly every major memory maker was running at or near full capacity, with 2026 production slots already "sold out." Samsung raised some prices 30% to 60% in two months. Contract prices for NAND would eventually run more than 600% above their September 2025 level; DRAM nearly 400%.
The clearest measure of what changed is allocation. TrendForce estimated that data centers would consume more than 70% of all high-end memory produced in 2026 — not 70% of some AI-specific chip, but 70% of the entire category that also feeds laptops, phones, and consoles. We traced earlier how that squeeze passed straight through to consumers: Surface laptops up $500, Galaxy tablets up $280, cheap smartphones priced out of existence in India and Africa. The Steam Deck's jump to $949 is the latest invoice from that same reallocation. But the price tag is only one face of the inversion. The other is the paycheck.
The Windfall
When a commodity that was supposed to stay cheap forever becomes the scarcest thing in the supply chain, the value has to land somewhere. For two years it landed in the price of devices. In 2026 it started landing in the wages of the people who make the chips.
The numbers arrived all at once. SK Hynix's union had already, in September 2025, secured 10% of the company's annual profit as a bonus pool — about $2.7 billion split across 33,625 staff. By May 2026, Samsung's workers were negotiating the same logic on a larger base: the deal approved on the 27th will distribute roughly $26.6 billion to about 78,000 chip employees, an average of $340,000 each, explicitly described as sharing the windfall from the AI boom. The same week, TSMC's CEO pledged a 30%-plus bump in profit-sharing payouts after staff complained their share was too small. Across three companies on three different parts of the supply chain, the workers had reached the same conclusion at the same time: the scarce thing is ours to make, and we want our cut of what scarcity is worth.
The most revealing number is internal. Inside Samsung, the bonus deal opened a roughly 100-to-1 gap between what memory-division staff receive and what the colleagues making smartphones, TVs, and home appliances receive. It is the inversion drawn inside a single company. The division building consumer devices — the one absorbing the higher memory costs and passing them to customers — is paid a fraction of what the memory division earns from creating the shortage. The payouts grew large enough that they have become a national debate in Korea over how the profits of the AI boom should be shared at all. The component nobody used to want to make is now the one whose workers everyone resents.
The workers who used to be first in line for the layoffs are now first in line for the bonus. The chip didn't change. The thing it became scarce against did.
Why This Time Is Different
The obvious objection is that memory has always done this. It is the most cyclical business in technology; every boom has a bust waiting behind it, and the $340,000 bonus could be the same kind of peak that preceded the $1.4 billion loss. The workers celebrating in 2026 could be the workers cut in 2028. The cycle is the one thing about memory that has never broken.
Except the cycle ran on a mechanism that is now gone. Memory prices collapsed in every past bust because demand was discretionary — a PC maker could wait, a phone buyer could skip a year, and when they did, the glut crushed the price. The 2026 demand is not discretionary. It is the $1.1 trillion cloud backlog, the multi-year capacity commitments, the buildout that has already been sold. And the makers know it. Memory companies are now leveraging their newfound power to lock customers into long-term agreements — the exact instrument the industry never had, because in a commodity business the buyer always held the leverage. A long-term contract is the structural admission that the cycle is over. You do not sign a ten-year deal for something whose price you expect to collapse.
The Reversal
For thirty years, the defining fact about memory was that it got cheaper, and the defining risk of making it was that you would be the first one laid off when it did. Both halves of that sentence have now flipped. The component engineered to disappear is the bottleneck everything else has to wait behind. The workers who used to absorb the downturn are the ones holding the windfall. On May 27, the same scarcity that pushed a $399 handheld to $949 handed the people who built the chips inside it an average of $340,000 — the squeeze and the bonus printed on the same day, the same shortage read from opposite ends of the line.