Micron reported a 72.6% gross margin and an order book sold out through 2026 to the last gigabit the same week 17 plaintiffs walked into the Northern District of California to argue that those numbers are the fruit of a crime. The class action, Garciaguirre v. Samsung Electronics, was filed June 25th and assigned to Judge Noel Wise. It invokes Section 1 of the Sherman Act against three companies holding roughly 90% of the global DRAM market, and it alleges that the coordinated pivot to high-bandwidth memory was "cover" to curtail production of older DDR3 and DDR4 modules—conduct the plaintiffs say drove prices up roughly 700% over four years. Samsung and SK Hynix have pleaded guilty to criminal DRAM price fixing before; SK Hynix paid a $185 million fine in April 2005. So the question writes itself: is this scarcity a conspiracy, or just a very good year?
That is the trap. And the lawsuit cannot escape it, because the answer the court is being asked to choose between—conspiracy or free market—is the wrong question. Last week, the coverage here treated memory as a capacity-timeline problem: when do the fabs come online, and does AI demand outrun them. This week the antitrust suit forces a harder framing, because it surfaces the mechanism the timeline question hides. The demand-versus-cartel debate is a false binary, and the memory oligopoly wins either way.
Start with the pricing data, because it settles one thing decisively. In Q1 2026, DRAM contract prices rose 90–95% quarter-over-quarter. The DRAM market grew 80% quarter-over-quarter and 260% year-over-year. At first glance those numbers read as a demand explosion. But unit shipment growth did not move at anything like that pace. The revenue came from price, not volume—the exact signature of a supply-constrained cycle, not a demand-led one. When a market triples in value while shipping roughly the same number of chips, you are not watching buyers bid up scarce goods made as fast as possible. You are watching output held back.
When a market triples in value while shipping roughly the same number of chips, you are not watching buyers bid up scarce goods. You are watching output held back.
Here is where the conspiracy question dissolves into irrelevance. Micron has publicly disclosed a 3-to-1 conversion ratio: every unit of HBM production ramp directly displaces three units of conventional DDR5 output. HBM is the stacked DRAM that feeds AI accelerators, and it is a wafer hog—Bernstein's numbers show conventional DRAM yields more than twice the revenue and roughly three times the gross profit per wafer. By 2026, HBM is expected to consume about 25% of total DRAM wafer production. So when Samsung, SK Hynix, and Micron move wafers toward HBM, they are not choosing between "serving AI demand" and "disciplining commodity supply." Those are the same physical act, seen from two angles. The wafer that becomes an HBM stack for an Nvidia accelerator is, mechanically and simultaneously, three DDR5 modules that never reach Troy's Computers or a Florida repair shop.
This is what makes the plaintiffs' theory both intuitively right and legally doomed. They are correct that HBM diversion choked DDR3 and DDR4. They are correct that prices exploded as a result. Where they fail is the leap from "the three majors all did the same thing" to "the three majors agreed to do it." Because they didn't have to agree. The economics agree for them.
The courthouse has already run this experiment. In 2018, Hagens Berman filed a class action making nearly identical parallel-cuts claims. The district court dismissed it in 2020. The Ninth Circuit upheld the dismissal in 2022, ruling that the trio's conduct was "more likely explained by lawful, unchoreographed free-market behavior" than by an illegal agreement. The plaintiffs had assembled eight "plus factors," and they still failed to clear Section 1's threshold, which demands evidence of an actual agreement rather than the conscious parallelism that is simply native to a three-supplier market. The new complaint leans on the HBM pivot as the additional evidence the earlier suit lacked. But the HBM pivot is the worst possible exhibit, because it is the most economically rational move any of the three could make. Each company faces a wafer with two destinations: one yields a 72.6% gross margin and a 2026 order book sold out to the last gigabit, the other yields a commodity squeezed by Chinese legacy producers. No rational executive needs a phone call to pick HBM. The conduct is now indistinguishable from rational margin maximization—which is precisely what makes it durable and dangerous. A smoke-filled room is fragile; it can be subpoenaed, flipped, fined $185 million. An economic gradient cannot be deposed.
That is the structural point the suit will never reach: the floor under DRAM prices is no longer enforced by agreement. It is enforced by arithmetic. The feedback loop runs—high HBM margins reward capacity diversion, diversion starves conventional DRAM, scarcity sustains pricing—and it does not require perpetual AI demand growth to hold. It only requires that HBM stay more attractive per wafer than commodity DRAM. The post-2022 behavior change is what locked this in: Samsung, SK Hynix, and Micron stopped chasing volume and learned to limit capex and prioritize technology migration over share. SK Hynix's 2026 capex exceeds 30 trillion KRW, with plans to double wafer capacity within five years—but much of that is aimed at HBM, not the general memory pool. The capital expanding the industry is the same capital tightening the commodity end of it.
The consequences are already visible downstream. DRAM has gone from roughly half the bill of materials on a server in 2023 to somewhere between 60% and 90% today, averaging around 75%. Lenovo told the International Supercomputing Conference that "it will never be like it was last year," and laid out a five-step "RAMageddon Survival Guide." Apple, a buyer with enormous leverage, has reportedly sought U.S. permission to source DRAM from CXMT, a Pentagon-blacklisted Chinese maker. When Apple is reduced to petitioning the government to buy from a blacklisted firm, the rent capture has reached the top of the food chain. Micron, meanwhile, signed roughly $100 billion in multi-year supply agreements—take-or-pay terms with pricing floors that convert the moment into durable, multi-year earnings visibility. This is not a company bracing for a crash. It is a company that has been told the floor will hold.
There is exactly one force that can break a discipline enforced by economics rather than by men, and it is not in this courtroom. Not regulators, not lawsuits, not new fabs—new fabs cost tens of billions and take years, which is precisely why no rival can step in to undercut. The only genuine threat is Chinese legacy-node oversupply. CXMT and YMTC are scaling subsidized capacity on older, high-volume nodes, decoupled from Western export controls, and if they flood the legacy DRAM and NAND market they could drag spot prices down and break the incumbents' premium on the commodity end. That is the single variable that matters, because it attacks the only thing holding the loop together: the relative attractiveness of HBM. Everything else—the treble-damages claim, the eight plus factors, the 700% headline—is theater.
The plaintiffs will lose, and they will lose for the most uncomfortable reason available. Not because the squeeze isn't real. Because it is so rational that the law cannot tell it apart from a free market doing exactly what a three-firm market does.